Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Structure from Motion Photogrammetry in Forestry: a Review

Structure from Motion Photogrammetry in Forestry: a Review Purpose of Review The adoption of Structure from Motion photogrammetry (SfM) is transforming the acquisition of three-dimensional (3D) remote sensing (RS) data in forestry. SfM photogrammetry enables surveys with little cost and technical expertise. We present the theoretical principles and practical considerations of this technology and show opportunities that SfM photogrammetry offers for forest practitioners and researchers. Recent Findings Our examples of key research indicate the successful application of SfM photogrammetry in forestry, in an operational context and in research, delivering results that are comparable to LiDAR surveys. Reviewed studies have identified possibilities for the extraction of biophysical forest parameters from airborne and terrestrial SfM point clouds and derived 2D data in area-based approaches (ABA) and individual tree approaches. Additionally, increases in the spatial and spectral resolution of sensors available for SfM photogrammetry enable forest health assessment and mon- itoring. The presented research reveals that coherent 3D data and spectral information, as provided by the SfM workflow, promote opportunities to derive both structural and physiological attributes at the individual tree crown (ITC) as well as stand levels. Summary We highlight the potential of using unmanned aerial vehicles (UAVs) and consumer-grade cameras for terres- trial SfM-based surveys in forestry. Offering several spatial products from a single sensor, the SfM workflow enables foresters to collect their own fit-for-purpose RS data. With the broad availability of non-expert SfM software, we provide important practical considerations for the collection of quality input image data to enable successful photogrammetric surveys. . . . . . Keywords SfM Point cloud UAV Close-range photogrammetry (CRP) Forest inventory Forest health This article is part of the Topical Collection on Remote Sensing * Jakob Iglhaut CETEMAS, Centro Tecnológico y Forestal de la Madera, Área de j.iglhaut.919076@swansea.ac.uk Desarrollo Forestal Sostenible, Pumarabule, Carbayín Bajo s/n 33936, Siero, Spain Carlos Cabo Department of Mining Exploitation and Prospecting, Polytechnic carloscabo.uniovi@gmail.com School of Mieres, University of Oviedo, Campus de Mieres, C/ Gonzalo Gutiérrez Quirós s/n, 33600 Mieres, Spain Stefano Puliti stefano.puliti@nibio.no Department of National Forest Inventory, Division of Forestry and Livia Piermattei Forest Resources, Norwegian Institute of Bioeconomy Research livia.piermattei@geo.tuwien.ac.at (NIBIO), P.O. Box 115, 1431 Ås, Norway James O’Connor Department of Geodesy and Geoinformation, TU Wien, james.oconnor@kingston.ac.uk Gußhausstraße 27-29/E120, 1040 Vienna, Austria Jacqueline Rosette 6 Physical Geography Catholic University of Eichstätt-Ingolstadt, j.a.rosette@swansea.ac.uk Ostenstraße 26, 85072 Eichstätt, Germany 1 7 Department of Geography, Swansea University, Singleton Park, Department of Natural and Built Environment, Kingston University, Swansea SA2 8PP, UK Penrhyn Road, Kingston upon Thames, Surrey KT2 1EE, UK 156 Curr Forestry Rep (2019) 5:155–168 Introduction spectral RS data has found wide application and acceptance in physical geography [23]: Structure from Motion (SfM), paired The use of remotely sensed (RS) data in forestry is motivated by with multi-view stereo (MVS) algorithms (SfM-MVS, com- efforts to increase cost efficiency, precision and timeliness of monly abbreviated to just SfM). SfM is based on computer forest information [1]. Differently, to traditional field-based sam- vision and facilitates the photogrammetric reconstruction from pling, the availability of full-coverage RS data enables the pro- images alone. Contrary to traditional stereophotogrammetry, duction of maps of key forestry variables, which are useful for 3D information can be computed from overlapping images, forest management purposes. First examples of aerial imagery without the need for prerequisite information on camera loca- usage for forestry purposes date back to the beginning of the tion and orientation, camera calibration and/or surveyed refer- 1920s [2, 3]. Over the past century, there has been tremendous ence points in the scene. This allows the use of inexpensive growth in the number of RS data sources available for the as- imaging platforms, both for aerial or terrestrial applications. sessment and monitoring of forests. Three-dimensional (3D) RS SfM photogrammetry has been comprehensively reviewed data, which can describe tree or canopy height, have shown in the geosciences [24, 25, 26�� ], where it has been gaining great potential for forest inventory [4]. In the past 20 years, the prominence for topographical surveys. We complement these use of airborne laser scanning (ALS) has been widely used for findings with a summary of SfM photogrammetry use specific forest inventory purposes and has become the standard data to forestry. We present an overview of the theoretical princi- source for operational forest inventories in many countries ples of a SfM-MVS workflow and its applications in forestry around the world [5–7]. Nevertheless, the acquisition of ALS by reviewing a representative sample of key research in this data requires a degree of planning and investment, making these field. Challenges and technical considerations are discussed, data sources cost-effective only on a relatively large scale [8]. concluding with opportunities and practical implications for Up to the beginning of 2010, there were no cost-effective means operational use of SfM by forest practitioners. of acquiring high-resolution 3D RS data for smaller areas, such as single forest properties or forest stands. Furthermore, in those Structure from Motion: Theoretical Principles cases, where ALS-based forest management is implemented, surveys are carried out infrequently, e.g. at intervals of 10– Traditional stereophotogrammetry methods are based on an anal- 20 years [5]. Hence, for some forest stands, the information ogy of the binocular human vision. Depth can be perceived from may be too unreliable for decision-making. Timeliness is a key two points whose relative position is known. However, depth, requirement to enable the adoption of precision forestry prac- volumes or 3D features can also be perceived from a single tices. This is especially true when the forest structure is changing observing point if either the observer or the object is moving rapidly, as is the case in fast-growing regeneration forests, or [27, 28]. SfM is a photogrammetric technique that is based on when growth is hindered by biotic or abiotic disturbances. both these principles: (i) the binocular vision and (ii) the chang- Photogrammetric approaches to obtain 3D information on ing vision of an object that is moving or observed from a moving forest structure have become popular, offering substantial cost point [29]. SfM is used for estimating 3D models from se- savings in the case of aerial photogrammetry compared with quences of overlapping 2D images. It gained popularity in recent ALS [9, 10]. Photogrammetry is limited to the reconstruction years due to its ability to deal with sets of unordered and hetero- of surfaces visible in the image data, providing ground informa- geneous images without prior knowledge of the camera param- tion only where large vegetation gaps exist. However, photo- eters [30]. SfM differs from traditional photogrammetry mainly grammetricdatacanbecombinedwithpre-existinggrounddata, in three aspects: (i) features can be automatically identified and derived from light detection and ranging data (LiDAR) for ex- matched in images at differing scales, viewing angles and orien- ample. This data synergy has been thoroughly discussed by tations, which is of particular benefit when small unstable plat- Goodbody et al. [11], indicating the potential for cost-efficient forms are considered; (ii) the equations used in the algorithm can forest inventory updates. Similarly, Kangas et al. [6] suggest an be solved without information of camera positions or ground equal value of photogrammetric and ALS data in forest manage- control points, although both can be added and used and (iii) ment planning, given that a ALS ground information is available camera calibration can be automatically solved or refined during from previous campaigns. Additional to the proven complemen- the process. SfM can thus automatically deliver photogrammet- tary use of LiDAR and photogrammetric data [9, 11, 12], recent ric models without requiring rigorous homogeneity in overlap- attempts at deriving inventory relevant forest metrics from pho- ping images, camera poses and calibrations [31–33]. togrammetric data alone show potential for aerial [13� , 14]and ‘SfM’ photogrammetry is commonly used to define the terrestrial [15� , 16] acquisitions. Further standalone use of photo- entire reconstruction workflow, from image set to dense point grammetry was shown for forest health monitoring [17, 18� , 19] cloud; however, strictly speaking, SfM only refers to a specific species classification [20] and biodiversity assessments [21, 22]. step in the workflow that provides camera parameters and a In the last decade, a photogrammetric approach offering sparse point cloud (see Fig. 1). Although some studies use the flexible and cost-effective acquisition of combined 3D and sparse point cloud as a final product [31, 34], in most cases, Curr Forestry Rep (2019) 5:155–168 157 The camera poses and parameters obtained from SfM are then applied to generate a densified point cloud using MVS algo- rithms. Prior to the MVS densification, and for computational efficiency or even viability, images are clustered based on their location [37]. In this way, the dense point cloud of each cluster (i.e. group of images) is computed separately (Figs. 1 and 2). A dense point cloud, with colour/spectral information de- rived from the input images, represents the primary output of the SfM-MVS workflow. Subsequent processing steps (for ae- rial surveys) typically involve the derivation of a digital surface model (DSM) and an orthomosaic. A canopy height model (CHM) can be attained by height normalization (i.e. conversion from height above sea level to height above ground) with a pre- existing digital terrain model (DTM). When SfM-derived sur- face data are height normalized in such a way, this offers the calculation of forest metrics like those commonly derived from ALS (e.g. height, timber volume, biomass). Additionally, image metrics like radiance/reflectance values and texture may be ex- tracted [13� , 18� , 20, 38, 39]. Finally, rasterization can offer opportunities to explore the sensed information in more depth Fig. 1 Schematic workflow of the SfM-MVS process resulting in a dense when statistics are calculated for every cell (e.g. height percen- point cloud from image sets. The point cloud is georeferenced by provid- tiles, surface roughness, spectral indices) [40–42]. ing positional information for images and/or ground control points dense image matching algorithms, such as MVS, are used in a SfM Photogrammetry in Practice subsequent step to densify the point cloud. The whole process can thus be referred to as SfM-MVS. Figure 1 contains a With photogrammetry being a passive technique, results are schematic workflow of the whole SfM-MVS process, and highly influenced by the input image data. SfM photogram- Fig. 2 shows a graphic diagram of the main three steps. metry, employing an automated process to identify and match The SfM-MVS process starts with the automatic extraction of features by computer vision, is fundamentally dependent on keypoints (i.e. points or sets of pixels with distinctive contrast or image quality. Sensors, settings and acquisition designs texture) in the images. The keypoints are identified in all images should be considered with great care. and then tied (matched) across images where they appear. The In every circumstance, the camera settings need to be consid- scale-invariant feature transform (SIFT [35]) and its variations ered to ensure optimal image data is acquired given a set of con- are the most common algorithms for keypoint identification and straints, namely (i) those from the environment (lighting condi- matching in SfM [26]. SIFT produces numerical descriptors for tions), (ii) the platform (UAV, pole, tripod or handheld) and (iii) the each point in each image. These descriptors are invariant to scale camera and lens combination (the exposure triangle, focal length, and orientation, thus suitable for identifying points or objects in sensor size). Acquiring high-quality image data has been pictures taken from different perspectives and under different discussed in O’Connor et al. [43] and Mosbrucker et al. [ 44], conditions. Then, coherence of keypoint matches is checked with key rules-of-thumb including keeping the motion of the cam- using a coarse reconstruction of the geometry of the images era to a minimum, and increasing ISO (i.e. the sensors sensitivity) and the relative position of the keypoints on them (Figs. 1 and 2). to account for potential underexposure (Fig. 3). RAW image data Given a sufficient number of images and keypoint matches, is better to capture as it retains the raw pixel values acquired by the SfM performs bundle adjustments to simultaneously compute camera prior to quantization and compression [45]. camera poses and parameters, and a sparse 3D point cloud of Image network geometry has an impact on the quality of the scene (consisting of the position of keypoints matched in reproduction, and for every survey, a ‘convergent’ imaging different images). The bundle adjustment is solved using (i) geometry should be sought that where the principal axis (per- initialization values obtained from sequences of randomly se- pendicular to the image sensor) of the images used converge lected matched keypoints and, complementarily, parameters so that systematic error is minimized [46, 47]. In UAV imag- from the cameras and poses and (ii) a non-linear refinement ing, James and Robson [46] suggest surveying with gently [36]. Then, the outputs of SfM are scaled and georeferenced banked turns when using fixed wing UAVs, in order to based on ground control points (GCPs) and/or data from nav- achieve this. With rotary UAVs, a similar result can be igation devices from the camera or its platform (Figs. 1 and 2). achieved by angling the camera on the gimbal on which it is �� �� 158 Curr Forestry Rep (2019) 5:155–168 Fig. 2 The three key stages in a SfM-MVS workflow illustrated on two hypothetical images of a Canary Island pine forest: (1) keypoint identification and matching (e.g. SIFT), (2) SfM with camera parameters and a sparse point cloud as output and (3) the densified point cloud fol- lowing MVS mounted. For terrestrial imaging, a convergence of images on with overlapping images from multiple locations and angles AOIs is advised, as presented in Mosbrucker et al. [44]. (high overlap to increase redundancy and multiple viewing an- Within image acquisition and SfM photogrammetric gles of the same object to reduce occlusions and systematic workflows, users have many parameters which they can vary errors), (ii) any feature to be reconstructed should be visible in depending on the equipment and software used. For some, at least three images (five or six images for dense vegetation) users can have near full control (e.g. the ‘exposure triangle’; and the angular divergence between neighbouring images be- ISO, shutter speed and aperture), though there are several which tween should not exceed 10–20°, (iii) the scene is sufficiently will only be estimated prior to performing a survey (such as the illuminated (constant lighting is preferable, e.g. overcast or exact camera positions images will be acquired from). Other cloud-free conditions) and (iv) object of interest is fixed (pref- influential factors, which cannot be manipulated (e.g. light con- erably no movement from branches in wind). ditions), will have to be carefully considered when planning a SfM-based survey. The success of reconstruction is ultimately The Current Status of SfM in Forestry dependent on factors that can be broken down into five catego- ries, as presented in Table 1. The accuracy of the position and With the ability to produce highly detailed 3D information scale of a survey is then determined by the referencing approach from a set of images alone, SfM photogrammetry lays a pow- (e.g. GCPs, direct georeferencing, manual scaling). erful tool into the hands of anyone looking to collect their own To apply SfM photogrammetry in forestry, important aspects fit-for-purpose RS data. Owing particularly to the potential of to a successful survey are as follows: (i) the scene is covered using off-the-shelf cameras and the availability of affordable Curr Forestry Rep (2019) 5:155–168 159 Fig. 3 Image quality issues illustrated by simulated degradations on UAV c adds noise, which can rapidly degrade image quality at high-ISO values; image: a adds motion blur, which has negative impacts on the quality of d adds overexposure, where the image sensor was exposed for too long a photogrammetric reproduction (Sieberth et al., 2014); b adds JPEG time and e underexposure, where the image sensor was exposed for too compression, which smoothes subtle contrast changes across an image; little time user-friendly software, the application of SfM photogramme- acquisitions, also termed close-range photogrammetry try in physical geography has increased rapidly [26�� , 48]. (CRP), focus on the reconstruction of stems within sample With SfM photogrammetry being scale independent, images plots or the reconstruction of individual trees [15� ]. may be acquired from a multitude of platforms ranging from A further field of research is the assessment and monitoring ground-based, handheld or pole-mounted options, to un- of forest health condition. For SfM-based mapping of the can- manned aerial vehicles (UAVs) and manned aircraft. UAVs opy, hereby an aerial acquisition of image data, most com- have enabled geospatial data to be acquired in new ways. monly by UAV, with multispectral sensors prevails. Studies Flexibly deployed at scales from several hectares to square dealing with forest health often make use of the 3D informa- kilometres [49], they allow forest practitioners to collect their tion and derived 2D spectral products that SfM photogram- own aerial information. In fact, there is an increasing interest metry delivers [18� , 50]. The following sections describe the in UAV forest surveys that can arguably be attributed to SfM- research on SfM-based forest inventory and health assess- based photogrammetric processing [26�� ]. ments to date in more detail. The rapid adoption of SfM photogrammetry is indicated by a growing number of scientific publications in forestry that utilize Inventory this photogrammetric technique. We conducted a search for peer-reviewed studies indexed by the Web of Knowledge data- Forest inventory holds a central role in all of the forest re- base using the keywords ‘Structure from Motion’, ‘UAV’ and search. Sustainable management of forests relies on knowl- ‘Forestry’ (and their most common variations). The search re- edge of their structure, distribution and dynamics over time sults were manually filtered to retain only forestry-related stud- [51]. The collection of field data for inventory purposes is ies applying a SfM-based workflow. We further categorized labour intensive, time-consuming and expensive, and cannot results into research on aerial and terrestrial inventory, forest be applied to large areas, consequently drastically limiting the health and proof-of-concept studies. These results are presented number of field inventories that can be afforded [52, 53]. in Table 2 and reveal a steady rise of publications on forest Efforts to improve on the efficiency of inventory practices remote sensing with SfM photogrammetry. therefore drive research in this field [53]. Amongst RS tech- SfM photogrammetry applications aimed at forest invento- nologies, SfM photogrammetry offers a low-cost and flexible ry are currently the most studied (Table 2). Here, a distinction approach to collect information on forest structure, thus natu- between aerial and terrestrial approaches can be made. An rally there has been an increase in interest to use such data for aerial approach typically utilizes a canopy surface model de- forest inventory. rived from SfM and/or associated spectral properties to esti- Within the context of forest inventory, the main use of SfM mate inventory relevant parameters [11]. Terrestrial photogrammetry has been its application on UAV image data 160 Curr Forestry Rep (2019) 5:155–168 Table 1 Overview of variables influencing the results of a SfM survey Domain Variable Recommendation Scene Texture High surface contrast to allow for feature-point detection Pattern repetition Increase overlap and increase accuracy of geotags Moving features Avoid! Occlusions Increase overlap and viewing angles Lighting conditions Sun angle High! Solar noon is ideal Weather Overcast provides even lighting (ambient occlusion) for structural (RGB) surveys. For spectral surveys little atmospheric influence may be required, clear skies. Changing illumination Avoid! Camera parameters Focal length Wide but not too wide to minimize distortions. 28–35 mm is a good basis (James et al. 2012) Exposure Δ Well exposed - Aperture Small for max DOF*, f/8 an advisable default - Shutter speed High for reduced motion blur*, ground speed (m/s) * exposure time (s) = blurred pixel -ISO Length Low for min noise*, auto-ISO an advisable default *Ideal scenario, but will always be a compromise between these three parameters Pixel pitch As high as is practical. Physical pixel size positively influences dynamic range and sensitivity Survey characteristics Overlap High (> 80% forward and lateral) as rule of thumb for forests to increase redundancy and matchability in scenes with high pattern repetition, moving features and/or occlusions. As a rule of thumb, a UAV-SfM data acquisition should be planned so that each point will be visible at least in 4–5images. View angles Convergent for reduction of systematic errors (RGB) Parallel (Nadir) for spectral sensing (reflectance) Survey range With increasing distance to the object/scene (decreasing GSD) survey precision degrades. Increased GSD requires higher overlap. Processing parameters SfM—matching If matching is not successful at full image scale ½ or ¼ may promote matchability - Image scale - Keypoints The number may be reduced for large datasets to reduce processing time MVS—densification Densification may not always be required at full image scale/maximum point cloud density Secondary products Multitude of algorithms for meshing, gridding etc. (results will depend on specific method) to produce wall-to-wall auxiliary information in a similar fash- image data, in recent years, there has been an increasing effort ion to ALS data. As such, UAV-SfM data has been shown to in developing terrestrial SfM applications to replace or aug- be suitable for the estimation of inventory relevant biophysical ment field data collections. The focus of studies incorporating parameters such as height, density and biomass [11, 12, 34, CRP lies on estimating diameter at breast height (DBH), tree 54–56]. Even though SfM has mostly been applied to aerial position and stem curves. The following sub-sections Table 2 Number of publications 2010 2013 2014 2015 2016 2017 2018 02/ on SfM photogrammetry for forest/tree remote sensing per year with manually assigned sub- SfM in forestry 1 3 3 11 18 31 66 4 categories - Airborne inventory 0 1 2 4 7 22 24 1 - Terrestrial inventory 0 0 0 1 4 2 10 0 - Forest health 0 0 0 2 4 1 7 1 - Proof-of-concept 1 2 1 4 3 6 15 2 Results presented are based on a manually filtered search in the scientific publications database Web of Knowledge using the search terms: TS = (‘Structure from Motion’ OR ‘Structure-from-Motion’ OR ‘SfM’ OR ‘sfm’ OR ‘structure from motion’ OR ‘structure-from-motion’ OR photogrammetry OR UAS OR SfM OR UAV OR RPAS OR drone OR CRP OR ‘unmanned aerial’ OR ‘Unmanned Aerial’) AND TS = (forest OR forestry OR tree). The date of this search was 21 February 2019 Curr Forestry Rep (2019) 5:155–168 161 elaborate further on the developments up to today regarding ALS data in terms of costs and accuracy. Goodbody et al. aerial and terrestrial SfM and highlight some of the key work [64] demonstrated the possibility to discriminate coniferous on using these photogrammetric data for inventory purposes. and deciduous species (overall accuracy of 86–95%). Puliti et al. [66� ] showed that UAV-SfM data could be used to accu- Aerial Inventory rately model stem density and height (RMSE% = 21.8% and 23.6%). Such results represent a substantial increase in accu- The use of SfM techniques applied to aerial image data for racy over ALS forest inventories and field assessment. forest inventory was pioneered by Dandois and Ellis in 2010 Furthermore, their study reported that data acquired using [54]. These authors were the first to use a series of unordered UAV-SfM techniques halved the amount of time required for but overlapping images acquired using a consumer-grade traditional field surveys that are commonly performed in re- camera mounted on a kite to produce a dense 3D point cloud generation stands. Thus, the use of UAV-SfM for regeneration representing the forest canopy. A first attempt to model forest forest may be particularly interesting since it allows a simul- biophysical properties using UAV-SfM data was done by taneous increase in the precision of the inventory while reduc- Dandois and Ellis in 2013 [34] and Lisein et al. in 2013 ing its costs. [55]. Both studies found that even though the results were Different methodological approaches have been applied to not consistent in all the studied areas, there was a correlation UAV-SfM data, similarly to ALS data. The methods can be between UAV-SfM data and variables such as dominant categorized into area-based approaches (ABA) [67] and indi- 2 2 height (R =0.07–0.91) or aboveground biomass (R =0.27– vidual tree crown (ITC) approaches [68, 69]. While in the 0.73). A more comprehensive evaluation of the possibilities to former case, the population units are represented by grid cells use UAV-SfM for forest inventories came with the studies by of area equal to that of the field plots; in the latter, they are Puliti et al. in 2015 [12] and Tuominen et al. in 2015 [56]who polygons representing single-tree crowns. In both cases, the extended their evaluation to the range of biophysical variables UAV-SfM data, corresponding either to the grid cells or the commonly used in forest management. Their results in terms single-tree crowns, are then linked to a sample of field obser- of RMSE% for dominant height (3.5%), Lorey’sheight vations either for field plots or for single trees through models. (13.3%–14.4%), stem density (38.6%), basal area (15.4– These models are then applied to all the population units either 23.9%) and timber volume (14.9–26.1%) were found to be for estimation of parameters for stand or forest level mapping. similar to errors associated with ALS-based forest inventories. The results of ABA methods have been presented in the pre- While these two studies set an important benchmark, they vious paragraph. The adoption of ITC approaches to UAV- were both conducted in even-aged managed boreal forests SfM has been found to be useful for detecting single trees with and thus provided limited information on how UAV-SfM 25–90% detection accuracy [63, 70, 71], to classify them ac- may perform in different forest types and forest developmental cording to tree species with overall accuracies up to 95% [71], stages. and measuring their height with RMSEs in the range of 0.5– Since the early days of UAV-SfM, the rapid growth in 2.84 m [55, 63]. In addition to rather large variability in the computing capabilities, availability of UAVs and SfM soft- accuracy of some of these variables, the results of UAV-SfM ware triggered increased interest in the scientific community ITC approaches vary according to forest types since they re- (see Table 1). This led to a widespread evaluation of UAV- main limited to the detection of the dominant tree layer, while SfM technology over a variety of forest types and forest de- smaller and dominated trees remain mostly undetected. velopmental stages. UAV-SfM data has been consistently proven to be useful for forest inventories in a large variety of Terrestrial Inventory forest types, including temperate European beech forests in Italy [13� ], mangrove forests in Malaysia [57], tropical forests Currently, terrestrial laser scanning (TLS) is the most accurate in Guyana [58], mixed conifer-broadleaved forest in Japan non-contact method of measurement to derive detailed forest [59], sparse sub-alpine coniferous forests in China [60], trop- inventory information at the plot level [15� ]. The main draw- ical woodlands in Malawi [41] and various plantations around backs of this technology are the high hardware costs [53], and the globe [61–63]. From these studies, a conclusion can be the time required for multiple scans mitigating occlusions drawn that the accuracy of UAV-SfM models is consistent along with post-processing to provide full coverage of a plot across many different forest types and on a similar scale to [72]. Mobile laser scanning systems reduce acquisition time ALS models. All of the aforementioned studies dealt with but high costs remain [73]. mature to nearly mature forest, while there has been little The deficiencies of traditional field data collection and the effort dedicated to estimating biophysical variables for forests need for reducing the cost of alternative laser scanning solu- under regeneration [64, 65, 66� ]. Nevertheless, the use of tions have encouraged the application of terrestrial photo- UAV-SfM data for regeneration forests may outperform alter- grammetry for forest inventory. Efforts to utilize terrestrial native data sources such as field assessments or the use of photogrammetric point clouds for deriving forest parameters 162 Curr Forestry Rep (2019) 5:155–168 derive from the low-cost of the equipment for the data collec- investigated by Liang et al. in 2014 and 2015 [74, 75]follow- tion, the automated SfM-based data processing and the poten- ed by Mokroš et al. in 2018 [78]. According to Mokroš et al. tially simple and fast data acquisition [74]. Requiring only a [78], the optimal acquisition solution resulted in portrait im- camera, typically handheld or mounted to a pole or tripod, ages, stop and go shooting mode and a path leading around the terrestrial SfM photogrammetry makes such a system highly plot with two diagonal paths through the plot. Differently, mobile, reducing the risk of occlusion yet providing a level of Liang et al. [75] concluded that the image matching results detail comparable to TLS [75]. of landscape images were optimal together with a Studies on terrestrial SfM for forestry purposes have be- photographing path based on inside and outside of an inner come more frequent in the last years (Table 1) and mainly circle (Fig. 4). For complex forest plots, Piermattei et al. [15� ] focus on linear rather than volumetric tree metrics. Studies found that the optimal acquisition path was a combination of vary according to (i) the scale of application, i.e. at plot level the solution found by Liang et al. [75] and Mokroš et al. [78]: and individual tree reconstruction; (ii) the measured forest landscape images, stop and go mode around the plot pointing parameters like tree position, DBH, height and stem curve; in, following by an inner circle pointing out of the plot and (iii) the resolution of the sensor, e.g. video, mobile phone two diagonals. This solution allowed reconstruction of stems and SLR camera; (iv) the camera configuration and photo- with an accuracy of few centimetres up to a few metres above graphic path and (v) the equipment used to acquire the images, ground. For low-density and medium-density forests, Liu e.g. pole, tripod, camera rig and backpack. Based on these et al. [77] proposed a system that combines two pole- aspects an overview of key work on terrestrial SfM applica- mounted cameras with a RTK GNSS for continuous captur- tions together with the obtained accuracies, acquisition meth- ing. Compared with total station measurements, their automat- od and geo-referencing approach are provided. ic determination of tree position, DBH and height achieved Most of the recent studies on photogrammetric measure- RMSEs of 0.16–0.2 m, 0.92–1.13 cm and 2.41–2.51 m ments of forest parameters are based on the single-camera respectively. technique, according to which overlapping images are ac- Most investigations on the use of terrestrial SfM were per- quired around the plot (Fig. 4). formed reconstructing single trees (i.e. not the entire forest Terrestrial photogrammetry has been evaluated in several plot) [16, 76, 79–83]. In those studies, DBH was the most studies in the past few years at plot scales [14, 15� , 16, 74, frequently estimated parameter and often compared with 76–78]. In these studies, DBH and tree locations were estimat- TLS data for accuracy assessment. Although sub-centimetre ed in circular plots with diameters ranging from 12 to 30 m. accuracy was achieved in all cases, the obtained RMSEs dif- The reported RMSE of the DBH ranged from 0.88 to 6.80 cm fered according to the approaches used, forest types and sur- compared with either field or TLS DBH measurements. Tree vey conditions, i.e. natural forest and controlled field settings. detection ranged between 60 and 98%. Results were influ- Not all the research studies report the time required for enced by the complexity of the forest plot, the acquisition path collecting the images. This can range from around 10 min to and mode. 2 h depending on the system used, parameters to be estimated, The impact of photographic path on the accuracy of forest plot size and survey configuration, by excluding the time to metrics derived from terrestrial SfM point clouds was firstly acquire scaling measurements. However, the accuracy of the Fig. 4 Example of a terrestrial SfM survey [51] in an open forest plot dense point cloud from the same point of view and c the dense point cloud showing a the configuration of camera positions and orientations, dense of a single stem without RGB colouring together with a 10-cm cross- point cloud and ground control points; b an example of an image and the section at 1.3 m (light blue point cloud) Curr Forestry Rep (2019) 5:155–168 163 scaling factor is crucial for forest plot and individual tree re- Näsi et al. [86] used hyperspectral image data combined construction [74]. To scale the photogrammetric point cloud, with the SfM-derived DSM for bark beetle damage detection most of the studies used targets surveyed by total station, at the individual tree level, achieving an overall classification implying additional equipment needs to be carried into the accuracy of 76% when distinguishing between healthy, field, consequently increasing the total acquisition time per infested and dead trees. In a follow -up study, Näsi et al. plot and reducing the portability of the entire surveying sys- [91] concluded that the individual tree-based approach, facil- tem. Aside from systems requiring a GNSS solution [77], itated by the combination of 3D and spectral data, provides a currently, only Liang et al. [74] tested natural reference ob- promising and cost-efficient alternative to field-based assess- jects, e.g. tree stems, for the determination of correct scale. ments of pest infestation. Minařík and Langhammer [92]also Their results showed that both natural reference objects and used a UAV-SfM-based mapping approach to map bark beetle artificial targets worked effectively. forest disturbance and found that bands from the red edge and NIR part of the spectrum were most suited for stress detection. Health Assessment and Monitoring These findings go in line with the results from Dash et al. [17], who assessed the potential of the commercially available mul- As part of a sustainable forest management, assessment and tispectral sensor, the Micasense RedEdge (Micasense Inc., monitoring of forest health condition play a crucial role. With Seattle, WA, USA), for the detection of early signs of stress threats to forest thought to increase globally [84], the identifi- during a simulated disease outbreak in a pine plantation. In the cation of declining forest health induced by biotic, abiotic and applied random forest classification of time-series data, nor- anthropogenic stress agents becomes imperative. RS ap- malized difference vegetation index (NDVI) showed to be the proaches offer rapid, spatially inclusive and objective ways best performing predictor variable to map physiological stress to monitor forest health when compared with field assess- symptoms along with the declining tree health. Further, late ments. With the aim of identifying and observing stress in examples of forest health monitoring are Baena et al. [93]and plants, multi- and hyperspectral sensors are capable of captur- Brovkina et al. [50], both successfully applying an OBIA ap- ing information outside the visible spectrum, which allow for proach on SfM-mapped NIR image data stemming from a estimation of biochemical plant traits like chlorophyll, leaf modified consumer RGB sensor to separate between dead pigments and canopy water content [85]. Spatially continuous and living trees. spectral mapping used to be exclusive to the manned airborne surveying domain. However, in recent years, lightweight sen- sors with discrete narrow spectral bands suitable for UAV Discussion mounting have become commercially available, allowing re- searchers to collect their own aerial spectral data [17, 86–88]. Used complementary to existing RS data (e.g. LiDAR) or by Such 2D spectral imagers may be used for SfM-based photo- itself, SfM photogrammetry has shown great potential for for- grammetric reconstruction and orthophoto generation similar- estry. Particularly attractive is the ability to use uncalibrated ly to RGB cameras, although they typically exhibit lower cameras paired with unstable or handheld platforms, enabling resolution. the use of low-cost and non-expert equipment. Ground and Opposed to the field of precision agriculture, where SfM- aerial SfM surveys can be carried out with high flexibility based processing of spectral image data is widely applied [89], offering the option for increased frequency RS surveys to in forestry there currently are only a few examples where incorporate, e.g. phenological changes in the analysis [59, SfM-derived mapped spectral products have found the appli- 65, 71]. The implementation of SfM algorithms in modern cation. Early UAV/SfM-based studies of forest health made photogrammetric software enables on-demand processing use of off-the-shelf RGB cameras, modified to capture near- with little required user input. SfM photogrammetry thus pre- infrared (NIR) images. Lehmann et al. [19] and Michez et al. sents a highly accessible and versatile solution to the acquisi- [18� ] used an object-based image analysis (OBIA) approach to tion of very high-resolution 3D data. In this regard, SfM em- segment and classify their scenes in order to identify declining powers common forestry practitioners to produce real-time tree health caused by biotic stress agents on both alder and data analytics with the minimum investment required for hard- oak. They achieved good overall classification accuracies ware and software. (79.5–84.1% and 82.5–85% at their respective study sites) Additional value in a SfM-based processing chain derives across five classes. However, they pointed out the limitations from the ability to provide multiple geospatial data products of NIR-modified RGB cameras, namely that visible and NIR (i.e. 3D models and orthomosaics) from a single sensor. spectra are not separable on the same sensor, spectral contam- Spectral information is inherently linked to the reconstructed ination due to broad and overlapping bands (see also Pauly structural data and derived products as these are generated [90]) as well as the inability to correct for changing light directly from the input imagery. Studies on forest health par- conditions (as downwelling irradiance is not captured). ticularly highlight the benefits of using the fused structural and 164 Curr Forestry Rep (2019) 5:155–168 spectral information that SfM-based processing of UAV image to further apply the method by Giannetti et al. [13� ]in a data provides [18� , 50]. wider variety of forest types and response variables. Point clouds generated from high-resolution images can Furthermore, the greater complexity of DTM- exhibit point densities greater than LiDAR, providing higher independent variables over more traditional explanatory detail information on the visible surface of forests. The in- variables could limit the transferability of the models creased spectral variation stemming from such high- through space and time. resolution data may hereby provide another valuable source 3) Lack of acquisition and processing protocols: of information, namely texture, such as the case in an OBIA The success of a SfM-based photogrammetric acquisi- approach [94]. Alongside the computational analysis, high- tion is largely based on the sensor used, the photographic resolution SfM-generated models appear visually realistic, path and viewing angles along with the chosen image providing experts a near true depiction of the scene. Intuitive overlap as well as the composition of a scene. to understand, SfM models thus hold an important advantage Adjustments to the acquisition approach to ensure quality over coarser remote sensing methods by enabling the rapid data are currently undertaken based on the surveyor’s visual assessment and/or validation. experience. Here protocols that enable certainty for SfM As is the case with all RS data, these will only ever be an outputs across forest types and phenological stages, yet approximation of the Earth’s surface and some limitations minimizing acquisition efforts, need to be established. always remain. With SfM photogrammetry being a new tech- Eltner et al. [24] suggested a protocol for the collection nology, the boundaries of these limitations are not fully tested of image data in geoscientific studies, which should be yet. Some of the main challenges with SfM photogrammetry extended to take into account forestry-specific factors. for forest applications that we are facing nowadays relate to Additional research is required on the parametrization of the following: SfM-based photogrammetric software for vegetated scenes specifically. Processing protocols designed to de- 1) Reproducibility: liver data adequate to the research question and to opti- With SfM photogrammetry enabling frequent surveys, mize processing speed are needed. variations in illumination, atmospheric and seasonal con- 4) Image matching issues: ditions are inevitable between acquisitions. Being a pas- Forests may prove to be challenging scenes for the sive sensing technique, these variations are directly feature matching algorithms underlying a SfM workflow. reflected on the data thus on the replicability of analyses. Their fine uniform texture, repeating patterns and poten- To ensure the use of SfM data on demand, allowing ac- tial movement (e.g. branches in wind) can introduce un- quisitions at different times of the year, it is therefore certainty in matching, consequently leading to incomplete reconstruction and/or noisy point clouds. In such cases, crucial to develop protocols for varying conditional sce- narios and models that account for variations in the data. the likelihood of identifying visual similarities in overlap- 2) Availability of accurate DTMs: ping images is promoted by increasing the distance to the Most airborne inventory studies presented here area of interest (AOI), thus increasing the number of fea- adopted highly accurate DTMs (e.g. ALS-based DTMs) tures per image and decreasing perspective distortions. to normalize UAV-SfM data and these are relatively rare Coarser ground sampling distances (GSDs) and higher around the globe, thus potentially limiting the area of image overlaps were shown to positively influence image application of UAV-SfM. To overcome this issue, some matching [95, 96]. The overlap should thus be increased authors proposed the use of DTMs generated from the when decreasing the GSD (i.e. images with finer detail). UAV-SfM data themselves [41] orthe use ofcoarsereso- Other potential mitigation strategies for reconstruction er- lution global DTMs such as shuttle radar topography mis- rors, like the use of high-accuracy position and orientation sion data (SRTM) [41]. While the former type of DTM is information for reduction of matching uncertainty, have obtainable only in open forests, the latter source was yet to be studied. found to be unsuitable for estimation of aboveground bio- mass. A conceptually novel approach came with the study To widely employ SfM photogrammetry in operational for- by Giannetti et al. [13� ] who, to overcome any of the estry, future research needs to tackle the abovementioned hur- abovementioned limitations, proposed the use of UAV- dles. It is essential to develop a consensus on acquisition pro- SfM data-derived variables without prior normalization tocols and parametrization of SfM photogrammetry software (i.e. DTM-independent variables). Their results showed that is set to answer specific research questions across forest that models fitted raw UAV-SfM data alone predicted types and environmental conditions. We have started to gain stem volume with similar accuracy to ALS data, even in some understanding of how image quality, overlap, GSD and the highly productive broadleaf forest in steep terrain. photographic path are influencing SfM-based reconstruction Despite such encouraging results, it remains fundamental [15� , 43�� , 95–97]. However, prior to processing, uncertainty Curr Forestry Rep (2019) 5:155–168 165 Human and Animal Rights and Informed Consent This article does not remains in predicting the completeness of these photogram- contain any studies with human or animal subjects performed by any of metric models. More in-depth work on these influential pa- the authors. rameters is needed in conjunction with the development of Open Access This article is distributed under the terms of the Creative methods that allow for reliable quality estimation of SfM- Commons Attribution 4.0 International License (http:// based outputs. Towards the quantification of data quality, creativecommons.org/licenses/by/4.0/), which permits unrestricted use, James et al. [98] presented a method for estimating the preci- distribution, and reproduction in any medium, provided you give appro- sion of each point produced within the SfM pipeline by re- priate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. peatedly running bundle adjustments on a set of input images. These ‘precision maps’ allow practitioners to describe the spatial variability of precision within SfM-derived products and gain insight into limitations in a given survey (such as image quality or control-point measurements). To our knowl- References edge, ‘precision maps’ have not been applied in the context of forested scenes. In forestry, future studies would benefit from Papers of particular interest, published recently, have been this method to objectively describe the data quality of SfM- highlighted as: derived products and thereby reduce uncertainty in subsequent � Of importance analysis. �� Of major importance 1. Mcroberts R, Tomppo E. Remote sensing support for national forest inventories. Remote Sens Environ. 2007;110:412–9. https://doi. org/10.1016/j.rse.2006.09.034. Conclusions 2. Wulder M. Optical remote-sensing techniques for the assessment of forest inventory and biophysical parameters. Prog Phys Geogr. A camera and a computer are the basic requirements for 1998;22:449–76. SfM photogrammetry. With the examples given here, 3. Korpela I. Individual tree measurements by means of digital aerial and in terms of what valuable data may be extracted photogrammetry. Silva Fennica Monographs 2004;3:93. 4. Lim K, Treitz P, Wulder M, St-Onge B, Flood M. LiDAR remote from SfM-derived data by analysis, SfM photogramme- sensing of forest structure. Prog Phys Geogr Earth Environ. try shows great potential for forest practitioners and 2003;27:88–106. researchers. Adding the power of UAVs for the acquisi- 5. Næsset E. Airborne laser scanning as a method in operational forest tion of aerial image data, the canopy of a forest can be inventory: status of accuracy assessments accomplished in mapped nearly in real time, responding rapidly to man- Scandinavia. Scand J For Res. 2007;22:433–42. 6. Kangas A, Gobakken T, Puliti S, Hauglin M, Næsset E. Value of agement needs. The temporal and spatial dimension that airborne laser scanning and digital aerial photogrammetry data in can be provided with SfM photogrammetry enables as- forest decision making. Silva Fennica 2018;52:1–19. sessment and monitoring of forests in an economical 7. Ginzler C, Hobi ML. Countrywide stereo-image matching for way that has not existed before. updating digital surface models in the framework of the swiss na- However, constraints linked to the fundamental principles tional forest inventory. Remote Sens. 2015;7:4343–70. 8. Koch B. Remote sensing supporting national forest inventories. In of SfM photogrammetry being a passive optical method will Knowledge Reference for National Forest Assessments; FAO. remain. Influencing factors like viewing geometry, lighting 2013:1–18. and the availability of static texture have to be carefully con- 9. White JC, Wulder MA, Vastaranta M, Coops NC, Pitt D, Woods M. sidered prior to every survey. We suggest that, in order to The utility of image-based point clouds for forest inventory: a com- parison with airborne laser scanning. Forests. 2013;4:518–36. enable the collection of fit-for-purpose data with predictable 10. Leberl F, Irschara A, Pock T, Meixner P, Gruber M, Scholz S, et al. quality, further work is needed on acquisition and processing Point clouds: Lidar versus 3D vision. Photogramm Eng Remote protocols specific to forestry. Further progress in these areas Sens. 2010;76:1123–34. will facilitate the move away from proof-of-concept studies 11. Goodbody TRH, Coops NC, White JC. Digital aerial photogram- and towards the operational application of SfM photogram- metry for updating area-based forest inventories: a review of oppor- tunities, challenges, and future directions. Curr For Rep. 2019;5: metry in forestry. 55–75. https://doi.org/10.1007/s40725-019-00087-2. 12. Puliti S, Ørka HO, Gobakken T, Naesset E. Inventory of small Funding Jakob Iglhaut is funded by the Royal Society Research Grant forest areas using an unmanned aerial system. Remote Sens. RG140494. Stefano Puliti is funded by NIBIO. 2015;7:9632–54. 13.� Giannetti F, Chirici G, Gobakken T, Næsset E, Travaglini D, Puliti Compliance with Ethical Standards S. A new approach with DTM-independent metrics for forest grow- ing stock prediction using UAV photogrammetric data. Remote Conflict of Interest Jakob Iglhaut, Carlos Cabo, Stefano Puliti, Livia Sens Environ. 2018;213:195–205. By omitting the prerequisite of DTM availability their approach indicates forward-looking Piermattei, James O’Connor, and Jacqueline Rosette declare that they potential for inventories based solely on SfM photogrammetry. have no conflict of interest. 166 Curr Forestry Rep (2019) 5:155–168 14. Mikita T, Janata P, Surový P. Forest stand inventory based on com- 31. Fonstad MA, Dietrich JT, Courville BC, Jensen JL, Carbonneau bined aerial and terrestrial close-range photogrammetry. Forests. PE. Topographic structure from motion: a new development in 2016;7:1–14. photogrammetric measurement. Earth Surf Process Landf. 2013;38:421–30. 15.� Piermattei L, Karel W, Wang D, Wieser M, Mokroš M, Koreň M, 32. Micheletti N, Chandler JH, Lane SN. Structure from motion (SfM) et al. Terrestrial Structure from Motion photogrammetry for deriv- photogrammetry. Br Soc Geomorphol. 2015. ing forest inventory data. Remote Sens. 2019. https://doi.org/10. 3390/rs11080950. This paper deals with the detection of tree 33. Skarlatos D, Kiparissi S. Comparison of laser scanning, photogram- metry and Sfm-Mvs pipeline applied in structures and artificial characteristics from close-range SfM photogrammetry and surfaces. ISPRS Ann Photogramm Remote Sens Spat Inf Sci. TLS in four case studies. The two techniques with respective 2012;I–3:299–304. acquisition and post-processing phases and their source re- 34. Dandois JP, Ellis EC. High spatial resolution three-dimensional quirements are discussed thoroughly. mapping of vegetation spectral dynamics using computer vision. 16. Mokroš M, Výbošťok J, Tomaštík J, Grznárová A, Valent P, Slavík Remote Sens Environ. 2013;136:259–76. https://doi.org/10.1016/ M, et al. High precision individual tree diameter and perimeter j.rse.2013.04.005. estimation from close-range photogrammetry. Forests. 2018;9. 35. Lowe DG. Distinctive image features from Scale-Invariant https://doi.org/10.3390/f9110696. Keypoints. Int J Comput Vis. 2004;60:91–110. 17. Dash JP, Watt MS, Pearse GD, Heaphy M, Dungey HS. Assessing 36. Shum H-Y, Ke Q, Zhang Z. Efficient bundle adjustment with virtual very high resolution UAV imagery for monitoring forest health key frames: a hierarchical approach to multi-frame structure from during a simulated disease outbreak. ISPRS J Photogramm motion. In: Proceedings. 1999 IEEE Comput. Soc. Conf. Comput. Remote Sens. 2017;131:1–14. Vis. Pattern Recognit. (Cat. No PR00149). IEEE Comput. Soc, pp 18.� Michez A, Piégay H, Lisein J, Claessens H, Lejeune P. 538–543. Classification of riparian forest species and health condition using 37. Ahmadabadian AH, Robson S, Boehm J, Shortis M, Wenzel K, multi-temporal and hyperspatial imagery from unmanned aerial Fritsch D. A comparison of dense matching algorithms for scaled system. Environ Monit Assess. 2016;188:146. A great example surface reconstruction using stereo camera rigs. ISPRS J of forest health monitoring at the individual tree level based on Photogramm Remote Sens. 2013;78:157–67. SfM-derived structural, spectral and textural variables. 38. Gini R, Sona G, Ronchetti G, Passoni D, Pinto L. Improving tree 19. Lehmann JRK, Nieberding F, Prinz T, Knoth C. Analysis of un- species classification using UAS multispectral images and texture manned aerial system-based CIR images in forestry-a new perspec- measures. ISPRS Int J Geo-Information. 2018;7:315. tive to monitor pest infestation levels. Forests. 2015;6:594–612. 39. Puliti S. Tree-stump detection, segmentation, classification, and https://doi.org/10.3390/f6030594. measurement using unmanned aerial vehicle (UAV) imagery. 20. Lisein J, Michez A, Claessens H, Lejeune P. Discrimination of Forests. 2018. https://doi.org/10.3390/f9030102 . deciduous tree species from time series of unmanned aerial system 40. Puliti S, Gobakken T, Ørka HO, Næsset E. Assessing 3D point imagery. PLoS One. 2015;10:1–20. clouds from aerial photographs for species-specific forest invento- 21. Alonzo M, Andersen H-E, Morton D, Cook B. Quantifying boreal ries. Scand J For Res. 2017;32:68–79. forest structure and composition using UAV structure from motion. 41. Kachamba D, Ørka H, Gobakken T, Eid T, Mwase W, Kachamba Forests. 2018;9:119. DJ, et al. Biomass estimation using 3D data from unmanned aerial 22. Saarinen N, Vastaranta M, Näsi R, Rosnell T, Hakala T, Honkavaara vehicle imagery in a tropical woodland. Remote Sens. 2016;8:968. E, et al. Assessing biodiversity in boreal forests with UAV-based 42. White JC, Wulder MA, Varhola A, Vastaranta M, Coops NC, Cook photogrammetric point clouds and hyperspectral imaging. Remote BD, et al. A best practices guide for generating forest inventory Sens. 2018;10. https://doi.org/10.3390/rs10020338. attributes from airborne laser scanning data using an area-based 23. Smith MW, Carrivick JL, Quincey DJ. Structure from motion pho- approach. Canadian Forest Service: Information Report. 2017. togrammetry in physical geography. Prog Phys Geogr. 2015;40: https://doi.org/10.5558/tfc2013-132 . 247–75. 43.�� O’Connor J, Smith MJ, James MR. Cameras and settings for aerial 24. Eltner A, Kaiser A, Castillo C, Rock G, Neugirg F, Abellán A. surveys in the geosciences: optimising image data. Prog Phys Image-based surface reconstruction in geomorphometry-merits, Geogr. 2017;41:325–44 This paper gives a very clear introduc- limits and developments. Earth Surf Dyn. 2016;4:359–89. https:// tion to aerial image capture. Foundational principles essential doi.org/10.5194/esurf-4-359-2016. to capturing high quality imagery with digital sensors are pro- 25. Bemis SP, Micklethwaite S, Turner D, James MR, Akciz S, Thiele vided. Cameras and settings are discussed with worked S, et al. Ground-based and UAV-based photogrammetry: a multi- examples. scale, high-resolution mapping tool for structural geology and 44. Mosbrucker AR, Major JJ, Spicer KR, Pitlick J. Camera system paleoseismology. J Struct Geol. 2014;69:163–78. considerations for geomorphic applications of SfM photogramme- 26.�� Carrivick J, Smith MJ. Quincey D. Structure from motion in the try. Earth Surf Process Landf. 2017;42:969–86. geosciences. 2016. This book further elaborates on the concepts 45. Verhoeven GJJ. It’s all about the format – unleashing the power of and methods of SfM photogrammetry applied in physical ge- RAW aerial photography. Int J Remote Sens. 2010;31:2009–42. ography. Technical aspects are described with great clarity. 46. James MR, Robson S. Mitigating systematic error in topographic 27. Bolles RC, Baker HH, Marimont DH. Epipolar-plane image anal- models derived from UAVand ground-based image networks. Earth ysis: an approach to determining structure from motion. Int J Surf Process Landf. 2014;39:1413–20. Comput Vis. 1987;1:7–55. 47. Wackrow R, Chandler JH. Minimising systematic error surfaces in 28. Ullman S. The interpretation of structure from motion. Proc R Soc digital elevation models using oblique convergent imagery. Lond Ser B Biol Sci. 1979;203:405–26. Photogramm Rec. 2011;26:16–31. 29. Koenderink JJ, van Doorn AJ. Affine structure from motion. J Opt 48. Anderson K, Westoby MJ, James MR. Low-budget topographic Soc Am A. 1991;8:377–85. surveying comes of age: structure from motion photogrammetry 30. Westoby MJ, Brasington J, Glasser NF, Hambrey MJ, Reynolds in geography and the geosciences. Prog Phys Geogr Earth Environ. 2019;43:163–73. JM. “Structure-from-motion” photogrammetry: a low-cost, effec- tive tool for geoscience applications. Geomorphology. 2012;179: 49. Anderson K, Gaston KJ. Lightweight unmanned aerial vehicles will 300–14 revolutionize spatial ecology. Front Ecol Environ. 2013;11:138–46 . . Curr Forestry Rep (2019) 5:155–168 167 50. Brovkina O, Cienciala E, Surový P, Janata P. Unmanned aerial UAV-SfM data can outperform ALS and traditional field data in forest inventory. vehicles (UAV) for assessment of qualitative classification of Norway spruce in temperate forest stands. Geo-Spat Inf Sci. 67. Næsset E, Økland T. Estimating tree height and tree crown proper- 2018;21:12–20. ties using airborne scanning laser in a boreal nature reserve. Remote Sens Environ. 2002;79:105–15. 51. Liang X, Kukko A, Hyyppä J, Lehtomäki M, Pyörälä J, Yu X, et al. 68. Brandtberg T. Automatic individual tree based analysis of high In-situ measurements from mobile platforms: an emerging ap- spatial resolution aerial images on naturally regenerated boreal for- proach to address the old challenges associated with forest inven- ests. Can J For Res. 1999;29:1464–78. tories. ISPRS J Photogramm Remote Sens. 2018;143:97–107. 69. Hyyppä J, Inkinen M. Detecting and estimating attributes for single 52. Thompson ID, Maher SC, Rouillard DP, Fryxell JM, Baker JA. trees using laser scanner. Photogramm J Finl. 1999;16:27–42. Accuracy of forest inventory mapping: some implications for boreal 70. Mohan M, Silva C, Klauberg C, Jat P, Catts G, Cardil A, et al. forest management. For Ecol Manag. 2007;252:208–21. ndividual tree detection from unmanned aerial vehicle (UAV) de- 53. White JC, Coops NC, Wulder MA, Vastaranta M, Hilker T, rived canopy height model in an open canopy mixed conifer forest. Tompalski P. Remote sensing technologies for enhancing forest Forests. 2017;8:340. inventories: a review. Can J Remote Sens. 2016;42:619–41. 71. Nevalainen O, Honkavaara E, Tuominen S, Viljanen N, Hakala T, 54. Dandois JP, Ellis EC. Remote sensing of vegetation structure using Yu X, et al. Individual tree detection and classification with UAV- computer vision. Remote Sens. 2010;2:1157–76. based photogrammetric point clouds and hyperspectral imaging. 55. Lisein J, Pierrot-Deseilligny M, Bonnet S, Lejeune P. A photogram- Remote Sens. 2017;9:185. metric workflow for the creation of a forest canopy height model 72. Liang X, Kankare V, Hyyppä J, Wang Y, Kukko A, Haggrén H, from small unmanned aerial system imagery. Forests. 2013;4:922– et al. Terrestrial laser scanning in forest inventories. ISPRS J Photogramm Remote Sens. 2016;115:63–77. 56. Tuominen S, Balazs A, Saari H, Pölönen I, Sarkeala J, Viitala R. 73. Wulder MA, Bater CW, Coops NC, Hilker TH, White JC. The role Unmanned aerial system imagery and photogrammetric canopy of LiDAR in sustainable forest management. For Chron. 2008;84: height data in area-based estimation of forest variables. Silva 807–826. Fenn. 2015. https://doi.org/10.14214/sf.1348. 74. Liang X, Jaakkola A, Wang Y, Hyyppä J, Honkavaara E, Liu J, et al. 57. Otero V, Van De Kerchove R, Satyanarayana B, Martínez-Espinosa The use of a hand-held camera for individual tree 3D mapping in C, Bin FMA, Bin IMR, et al. Managing mangrove forests from the forest sample plots. Remote Sens. 2014;6:6587–603. sky: forest inventory using field data and unmanned aerial vehicle 75. Liang X, Wang Y, Jaakkola A, Kukko A, Kaartinen H, Hyyppä J, (UAV) imagery in the Matang Mangrove Forest Reserve, peninsu- et al. Forest data collection using terrestrial image-based point lar Malaysia. For Ecol Manag. 2018;411:35–45. clouds from a handheld camera compared to terrestrial and personal 58. Roşca S, Suomalainen J, Bartholomeus H, Herold M. Comparing laser scanning. IEEE Trans Geosci Remote Sens. 2015;53:5117– terrestrial laser scanning and unmanned aerial vehicle structure 32. https://doi.org/10.1109/TGRS.2015.2417316. from motion to assess top of canopy structure in tropical forests. 76. Surový P, Yoshimoto A, Panagiotidis D. Accuracy of reconstruction Interface Focus. 2018;8:20170038. of the tree stem surface using terrestrial close-range photogramme- 59. Jayathunga S, Owari T, Tsuyuki S. Evaluating the performance of try. Remote Sens. 2016;8:1–13. photogrammetric products using fixed-wing UAV imagery over a 77. Liu J, Feng Z, Yang L, Mannan A, Khan T, Zhao Z, et al. Extraction mixed conifer-broadleaf forest: comparison with airborne laser of sample plot parameters from 3D point cloud reconstruction based scanning. Remote Sens. 2018;10. https://doi.org/10.3390/ on combined RTK and CCD continuous photography. Remote rs10020187. Sens. 2018;10:1299. 60. Lin J, Wang M, Ma M, Lin Y, Lin J, Wang M, et al. Aboveground 78. Mokroš M, Liang X, Surový P, Valent P, Čerňava J, Chudý F. tree biomass estimation of sparse subalpine coniferous forest with Evaluation of close-range photogrammetry image collection UAV oblique photography. Remote Sens. 2018;10:1849. methods for estimating tree diameters. Int J Geo-Inf. 2018;7:93. 61. Shin P, Sankey T, Moore M, Thode A, Shin P, Sankey T, et al. 79. Morgenroth J, Gomez C. Assessment of tree structure using a 3D Evaluating unmanned aerial vehicle images for estimating forest image analysis technique-a proof of concept. Urban For Urban canopy fuels in a ponderosa pine stand. Remote Sens. 2018;10: Green. 2014;13:198–203. 80. Miller JM. Estimation of individual tree metrics using structure- 62. Iizuka K, Yonehara T, Itoh M, Kosugi Y. Estimating tree height and from-motion photogrammetry. MSc Thesis. 2015. diameter at breast height (DBH) from digital surface models and 81. Bauwens S, Fayolle A, Gourlet-Fleury S, Ndjele LM, Mengal C, orthophotos obtained with an unmanned aerial system for a Lejeune P. Terrestrial photogrammetry: a non-destructive method Japanese Cypress (Chamaecyparis obtusa) Forest. Remote Sens. for modelling irregularly shaped tropical tree trunks. Methods Ecol 2018;10. https://doi.org/10.3390/rs10010013. Evol. 2017;8:460–71. 63. Guerra-Hernández J, Cosenza DN, Rodriguez LCE, Silva M, Tomé 82. Berveglieri A, Tommaselli AMG, Liang X, Honkavaara E. Vertical M, Díaz-Varela RA, et al. Comparison of ALS- and UAV(SfM)- optical scanning with panoramic vision for tree trunk reconstruc- derived high-density point clouds for individual tree detection in tion. Sensors (Switzerland). 2017;17:1–13. Eucalyptus plantations. Int J Remote Sens. 2018;39:5211–35. 83. Huang H, Zhang H, Chen C, Tang L. Three-dimensional digitiza- 64. Goodbody TRH, Coops NC, Hermosilla T, Tompalski P, Crawford tion of the arid land plant Haloxylon ammodendron using a P. Assessing the status of forest regeneration using digital aerial consumer-grade camera. Ecol Evol. 2018;8:5891–9. photogrammetry and unmanned aerial systems. Int J Remote 84. LauschA,ErasmiS,KingD,MagdonP, Heurich M. Sens. 2018;39:5246–64. Understanding forest health with remote sensing -part I—areview 65. Feduck C, McDermid G, Castilla G, Feduck C, McDermid GJ, of spectral traits, processes and remote-sensing characteristics. Castilla G. Detection of coniferous seedlings in UAV imagery. Remote Sens. 2016;8:1029. Forests. 2018;9:432. 85. Stone C, Mohammed C. Application of remote sensing technolo- 66.� Puliti S, Solberg S, Granhus A. Use of UAV photogrammetric data gies for assessing planted forests damaged by insect pests and fun- gal pathogens: a review. Curr For Rep. 2017;3:75–92. for estimation of biophysical properties in forest stands under re- generation. Remote Sens. 2019;11:233. https://doi.org/10.3390/ 86. Näsi R, Honkavaara E, Lyytikäinen-Saarenmaa P, Blomqvist M, rs11030233 . This paper demonstrates that high resolution Litkey P, Hakala T, et al. Using UAV-based photogrammetry and 168 Curr Forestry Rep (2019) 5:155–168 hyperspectral imaging for mapping bark beetle damage at tree-lev- 94. Blaschke T. Object based image analysis for remote sensing. ISPRS J Photogramm Remote Sens. 2010;65:2–16. el. Remote Sens. 2015;7:15467–93. 95. Frey J, Kovach K, Stemmler S, Koch B. UAV photogrammetry of 87. Minařík R, Langhammer J. Use of a multispectral UAV photogram- forests as a vulnerable process. A sensitivity analysis for a structure metry for detection and tracking of forest disturbance dynamics. Int from motion RGB-image pipeline. Remote Sens. 2018;10. https:// Arch Photogramm Remote Sens Spat Inf Sci. 2016;41:711–8. doi.org/10.3390/rs10060912. 88. Aasen H, Bareth G. Spectral and 3D nonspectral approaches to crop 96. Ni W, Sun G, Pang Y, Zhang Z, Liu J, Yang A, et al. Mapping three- trait estimation using ground and UAV sensing. In Biophys. dimensional structures of forest canopy using UAV stereo imagery: Biochem. Charact. Plant Species Stud: CRC Press; 2019;103–32. evaluating impacts of forward overlaps and image resolutions with 89. Hunt ER, Daughtry CST. What good are unmanned aircraft systems LiDAR data as reference. IEEE J Sel Top Appl Earth Obs Remote for agricultural remote sensing and precision agriculture? Int J Sens. 2018;11:3578–89. Remote Sens. 2018;39:5345–76. 97. Dandois JP, Olano M, Ellis EC. Optimal altitude, overlap, and 90. Pauly K. Towards calibrated vegetation indices from UAS-derived weather conditions for computer vision uav estimates of forest orthomosaics. 13th Int Conf Precis Agric. 2016. https://doi.org/10. structure. Remote Sens. 2015;7:13895–920. https://doi.org/10. 13140/RG.2.2.21842.35524. 3390/rs71013895. 91. Näsi R, Honkavaara E, Blomqvist M, Lyytikäinen-Saarenmaa P, 98. James MR, Robson S, Smith MW. 3-D uncertainty-based topo- Hakala T, Viljanen N, et al. Remote sensing of bark beetle damage graphic change detection with structure-from-motion photogram- in urban forests at individual tree level using a novel hyperspectral metry: precision maps for ground control and directly camera from UAV and aircraft. Urban For Urban Green. 2018;30: georeferenced surveys. Earth Surf Process Landf. 2017;42:1769– 72–83. 88. https://doi.org/10.1002/esp.4125. 92. Minařík R, Langhammer J. Use of a multispectral Uav photogram- metry for detection and tracking of forest disturbance dynamics. Int Publisher’sNote Springer Nature remains neutral with regard to Arch Photogramm Remote Sens Spat Inf Sci. 2016;XLI-B8:711–8. jurisdictional claims in published maps and institutional affiliations. 93. Baena S, Moat J, Whaley O, Boyd DS. Identifying species from the air: UAVs and the very high resolution challenge for plant conser- vation. PLoS One. 2017;12:1–21. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Current Forestry Reports Springer Journals

Structure from Motion Photogrammetry in Forestry: a Review

Loading next page...
 
/lp/springer-journals/structure-from-motion-photogrammetry-in-forestry-a-review-00ZFQ9O03F

References (107)

Publisher
Springer Journals
Copyright
Copyright © 2019 by The Author(s)
Subject
Environment; Sustainable Development; Environmental Management; Nature Conservation; Forestry; Forestry Management; Ecology
eISSN
2198-6436
DOI
10.1007/s40725-019-00094-3
Publisher site
See Article on Publisher Site

Abstract

Purpose of Review The adoption of Structure from Motion photogrammetry (SfM) is transforming the acquisition of three-dimensional (3D) remote sensing (RS) data in forestry. SfM photogrammetry enables surveys with little cost and technical expertise. We present the theoretical principles and practical considerations of this technology and show opportunities that SfM photogrammetry offers for forest practitioners and researchers. Recent Findings Our examples of key research indicate the successful application of SfM photogrammetry in forestry, in an operational context and in research, delivering results that are comparable to LiDAR surveys. Reviewed studies have identified possibilities for the extraction of biophysical forest parameters from airborne and terrestrial SfM point clouds and derived 2D data in area-based approaches (ABA) and individual tree approaches. Additionally, increases in the spatial and spectral resolution of sensors available for SfM photogrammetry enable forest health assessment and mon- itoring. The presented research reveals that coherent 3D data and spectral information, as provided by the SfM workflow, promote opportunities to derive both structural and physiological attributes at the individual tree crown (ITC) as well as stand levels. Summary We highlight the potential of using unmanned aerial vehicles (UAVs) and consumer-grade cameras for terres- trial SfM-based surveys in forestry. Offering several spatial products from a single sensor, the SfM workflow enables foresters to collect their own fit-for-purpose RS data. With the broad availability of non-expert SfM software, we provide important practical considerations for the collection of quality input image data to enable successful photogrammetric surveys. . . . . . Keywords SfM Point cloud UAV Close-range photogrammetry (CRP) Forest inventory Forest health This article is part of the Topical Collection on Remote Sensing * Jakob Iglhaut CETEMAS, Centro Tecnológico y Forestal de la Madera, Área de j.iglhaut.919076@swansea.ac.uk Desarrollo Forestal Sostenible, Pumarabule, Carbayín Bajo s/n 33936, Siero, Spain Carlos Cabo Department of Mining Exploitation and Prospecting, Polytechnic carloscabo.uniovi@gmail.com School of Mieres, University of Oviedo, Campus de Mieres, C/ Gonzalo Gutiérrez Quirós s/n, 33600 Mieres, Spain Stefano Puliti stefano.puliti@nibio.no Department of National Forest Inventory, Division of Forestry and Livia Piermattei Forest Resources, Norwegian Institute of Bioeconomy Research livia.piermattei@geo.tuwien.ac.at (NIBIO), P.O. Box 115, 1431 Ås, Norway James O’Connor Department of Geodesy and Geoinformation, TU Wien, james.oconnor@kingston.ac.uk Gußhausstraße 27-29/E120, 1040 Vienna, Austria Jacqueline Rosette 6 Physical Geography Catholic University of Eichstätt-Ingolstadt, j.a.rosette@swansea.ac.uk Ostenstraße 26, 85072 Eichstätt, Germany 1 7 Department of Geography, Swansea University, Singleton Park, Department of Natural and Built Environment, Kingston University, Swansea SA2 8PP, UK Penrhyn Road, Kingston upon Thames, Surrey KT2 1EE, UK 156 Curr Forestry Rep (2019) 5:155–168 Introduction spectral RS data has found wide application and acceptance in physical geography [23]: Structure from Motion (SfM), paired The use of remotely sensed (RS) data in forestry is motivated by with multi-view stereo (MVS) algorithms (SfM-MVS, com- efforts to increase cost efficiency, precision and timeliness of monly abbreviated to just SfM). SfM is based on computer forest information [1]. Differently, to traditional field-based sam- vision and facilitates the photogrammetric reconstruction from pling, the availability of full-coverage RS data enables the pro- images alone. Contrary to traditional stereophotogrammetry, duction of maps of key forestry variables, which are useful for 3D information can be computed from overlapping images, forest management purposes. First examples of aerial imagery without the need for prerequisite information on camera loca- usage for forestry purposes date back to the beginning of the tion and orientation, camera calibration and/or surveyed refer- 1920s [2, 3]. Over the past century, there has been tremendous ence points in the scene. This allows the use of inexpensive growth in the number of RS data sources available for the as- imaging platforms, both for aerial or terrestrial applications. sessment and monitoring of forests. Three-dimensional (3D) RS SfM photogrammetry has been comprehensively reviewed data, which can describe tree or canopy height, have shown in the geosciences [24, 25, 26�� ], where it has been gaining great potential for forest inventory [4]. In the past 20 years, the prominence for topographical surveys. We complement these use of airborne laser scanning (ALS) has been widely used for findings with a summary of SfM photogrammetry use specific forest inventory purposes and has become the standard data to forestry. We present an overview of the theoretical princi- source for operational forest inventories in many countries ples of a SfM-MVS workflow and its applications in forestry around the world [5–7]. Nevertheless, the acquisition of ALS by reviewing a representative sample of key research in this data requires a degree of planning and investment, making these field. Challenges and technical considerations are discussed, data sources cost-effective only on a relatively large scale [8]. concluding with opportunities and practical implications for Up to the beginning of 2010, there were no cost-effective means operational use of SfM by forest practitioners. of acquiring high-resolution 3D RS data for smaller areas, such as single forest properties or forest stands. Furthermore, in those Structure from Motion: Theoretical Principles cases, where ALS-based forest management is implemented, surveys are carried out infrequently, e.g. at intervals of 10– Traditional stereophotogrammetry methods are based on an anal- 20 years [5]. Hence, for some forest stands, the information ogy of the binocular human vision. Depth can be perceived from may be too unreliable for decision-making. Timeliness is a key two points whose relative position is known. However, depth, requirement to enable the adoption of precision forestry prac- volumes or 3D features can also be perceived from a single tices. This is especially true when the forest structure is changing observing point if either the observer or the object is moving rapidly, as is the case in fast-growing regeneration forests, or [27, 28]. SfM is a photogrammetric technique that is based on when growth is hindered by biotic or abiotic disturbances. both these principles: (i) the binocular vision and (ii) the chang- Photogrammetric approaches to obtain 3D information on ing vision of an object that is moving or observed from a moving forest structure have become popular, offering substantial cost point [29]. SfM is used for estimating 3D models from se- savings in the case of aerial photogrammetry compared with quences of overlapping 2D images. It gained popularity in recent ALS [9, 10]. Photogrammetry is limited to the reconstruction years due to its ability to deal with sets of unordered and hetero- of surfaces visible in the image data, providing ground informa- geneous images without prior knowledge of the camera param- tion only where large vegetation gaps exist. However, photo- eters [30]. SfM differs from traditional photogrammetry mainly grammetricdatacanbecombinedwithpre-existinggrounddata, in three aspects: (i) features can be automatically identified and derived from light detection and ranging data (LiDAR) for ex- matched in images at differing scales, viewing angles and orien- ample. This data synergy has been thoroughly discussed by tations, which is of particular benefit when small unstable plat- Goodbody et al. [11], indicating the potential for cost-efficient forms are considered; (ii) the equations used in the algorithm can forest inventory updates. Similarly, Kangas et al. [6] suggest an be solved without information of camera positions or ground equal value of photogrammetric and ALS data in forest manage- control points, although both can be added and used and (iii) ment planning, given that a ALS ground information is available camera calibration can be automatically solved or refined during from previous campaigns. Additional to the proven complemen- the process. SfM can thus automatically deliver photogrammet- tary use of LiDAR and photogrammetric data [9, 11, 12], recent ric models without requiring rigorous homogeneity in overlap- attempts at deriving inventory relevant forest metrics from pho- ping images, camera poses and calibrations [31–33]. togrammetric data alone show potential for aerial [13� , 14]and ‘SfM’ photogrammetry is commonly used to define the terrestrial [15� , 16] acquisitions. Further standalone use of photo- entire reconstruction workflow, from image set to dense point grammetry was shown for forest health monitoring [17, 18� , 19] cloud; however, strictly speaking, SfM only refers to a specific species classification [20] and biodiversity assessments [21, 22]. step in the workflow that provides camera parameters and a In the last decade, a photogrammetric approach offering sparse point cloud (see Fig. 1). Although some studies use the flexible and cost-effective acquisition of combined 3D and sparse point cloud as a final product [31, 34], in most cases, Curr Forestry Rep (2019) 5:155–168 157 The camera poses and parameters obtained from SfM are then applied to generate a densified point cloud using MVS algo- rithms. Prior to the MVS densification, and for computational efficiency or even viability, images are clustered based on their location [37]. In this way, the dense point cloud of each cluster (i.e. group of images) is computed separately (Figs. 1 and 2). A dense point cloud, with colour/spectral information de- rived from the input images, represents the primary output of the SfM-MVS workflow. Subsequent processing steps (for ae- rial surveys) typically involve the derivation of a digital surface model (DSM) and an orthomosaic. A canopy height model (CHM) can be attained by height normalization (i.e. conversion from height above sea level to height above ground) with a pre- existing digital terrain model (DTM). When SfM-derived sur- face data are height normalized in such a way, this offers the calculation of forest metrics like those commonly derived from ALS (e.g. height, timber volume, biomass). Additionally, image metrics like radiance/reflectance values and texture may be ex- tracted [13� , 18� , 20, 38, 39]. Finally, rasterization can offer opportunities to explore the sensed information in more depth Fig. 1 Schematic workflow of the SfM-MVS process resulting in a dense when statistics are calculated for every cell (e.g. height percen- point cloud from image sets. The point cloud is georeferenced by provid- tiles, surface roughness, spectral indices) [40–42]. ing positional information for images and/or ground control points dense image matching algorithms, such as MVS, are used in a SfM Photogrammetry in Practice subsequent step to densify the point cloud. The whole process can thus be referred to as SfM-MVS. Figure 1 contains a With photogrammetry being a passive technique, results are schematic workflow of the whole SfM-MVS process, and highly influenced by the input image data. SfM photogram- Fig. 2 shows a graphic diagram of the main three steps. metry, employing an automated process to identify and match The SfM-MVS process starts with the automatic extraction of features by computer vision, is fundamentally dependent on keypoints (i.e. points or sets of pixels with distinctive contrast or image quality. Sensors, settings and acquisition designs texture) in the images. The keypoints are identified in all images should be considered with great care. and then tied (matched) across images where they appear. The In every circumstance, the camera settings need to be consid- scale-invariant feature transform (SIFT [35]) and its variations ered to ensure optimal image data is acquired given a set of con- are the most common algorithms for keypoint identification and straints, namely (i) those from the environment (lighting condi- matching in SfM [26]. SIFT produces numerical descriptors for tions), (ii) the platform (UAV, pole, tripod or handheld) and (iii) the each point in each image. These descriptors are invariant to scale camera and lens combination (the exposure triangle, focal length, and orientation, thus suitable for identifying points or objects in sensor size). Acquiring high-quality image data has been pictures taken from different perspectives and under different discussed in O’Connor et al. [43] and Mosbrucker et al. [ 44], conditions. Then, coherence of keypoint matches is checked with key rules-of-thumb including keeping the motion of the cam- using a coarse reconstruction of the geometry of the images era to a minimum, and increasing ISO (i.e. the sensors sensitivity) and the relative position of the keypoints on them (Figs. 1 and 2). to account for potential underexposure (Fig. 3). RAW image data Given a sufficient number of images and keypoint matches, is better to capture as it retains the raw pixel values acquired by the SfM performs bundle adjustments to simultaneously compute camera prior to quantization and compression [45]. camera poses and parameters, and a sparse 3D point cloud of Image network geometry has an impact on the quality of the scene (consisting of the position of keypoints matched in reproduction, and for every survey, a ‘convergent’ imaging different images). The bundle adjustment is solved using (i) geometry should be sought that where the principal axis (per- initialization values obtained from sequences of randomly se- pendicular to the image sensor) of the images used converge lected matched keypoints and, complementarily, parameters so that systematic error is minimized [46, 47]. In UAV imag- from the cameras and poses and (ii) a non-linear refinement ing, James and Robson [46] suggest surveying with gently [36]. Then, the outputs of SfM are scaled and georeferenced banked turns when using fixed wing UAVs, in order to based on ground control points (GCPs) and/or data from nav- achieve this. With rotary UAVs, a similar result can be igation devices from the camera or its platform (Figs. 1 and 2). achieved by angling the camera on the gimbal on which it is �� �� 158 Curr Forestry Rep (2019) 5:155–168 Fig. 2 The three key stages in a SfM-MVS workflow illustrated on two hypothetical images of a Canary Island pine forest: (1) keypoint identification and matching (e.g. SIFT), (2) SfM with camera parameters and a sparse point cloud as output and (3) the densified point cloud fol- lowing MVS mounted. For terrestrial imaging, a convergence of images on with overlapping images from multiple locations and angles AOIs is advised, as presented in Mosbrucker et al. [44]. (high overlap to increase redundancy and multiple viewing an- Within image acquisition and SfM photogrammetric gles of the same object to reduce occlusions and systematic workflows, users have many parameters which they can vary errors), (ii) any feature to be reconstructed should be visible in depending on the equipment and software used. For some, at least three images (five or six images for dense vegetation) users can have near full control (e.g. the ‘exposure triangle’; and the angular divergence between neighbouring images be- ISO, shutter speed and aperture), though there are several which tween should not exceed 10–20°, (iii) the scene is sufficiently will only be estimated prior to performing a survey (such as the illuminated (constant lighting is preferable, e.g. overcast or exact camera positions images will be acquired from). Other cloud-free conditions) and (iv) object of interest is fixed (pref- influential factors, which cannot be manipulated (e.g. light con- erably no movement from branches in wind). ditions), will have to be carefully considered when planning a SfM-based survey. The success of reconstruction is ultimately The Current Status of SfM in Forestry dependent on factors that can be broken down into five catego- ries, as presented in Table 1. The accuracy of the position and With the ability to produce highly detailed 3D information scale of a survey is then determined by the referencing approach from a set of images alone, SfM photogrammetry lays a pow- (e.g. GCPs, direct georeferencing, manual scaling). erful tool into the hands of anyone looking to collect their own To apply SfM photogrammetry in forestry, important aspects fit-for-purpose RS data. Owing particularly to the potential of to a successful survey are as follows: (i) the scene is covered using off-the-shelf cameras and the availability of affordable Curr Forestry Rep (2019) 5:155–168 159 Fig. 3 Image quality issues illustrated by simulated degradations on UAV c adds noise, which can rapidly degrade image quality at high-ISO values; image: a adds motion blur, which has negative impacts on the quality of d adds overexposure, where the image sensor was exposed for too long a photogrammetric reproduction (Sieberth et al., 2014); b adds JPEG time and e underexposure, where the image sensor was exposed for too compression, which smoothes subtle contrast changes across an image; little time user-friendly software, the application of SfM photogramme- acquisitions, also termed close-range photogrammetry try in physical geography has increased rapidly [26�� , 48]. (CRP), focus on the reconstruction of stems within sample With SfM photogrammetry being scale independent, images plots or the reconstruction of individual trees [15� ]. may be acquired from a multitude of platforms ranging from A further field of research is the assessment and monitoring ground-based, handheld or pole-mounted options, to un- of forest health condition. For SfM-based mapping of the can- manned aerial vehicles (UAVs) and manned aircraft. UAVs opy, hereby an aerial acquisition of image data, most com- have enabled geospatial data to be acquired in new ways. monly by UAV, with multispectral sensors prevails. Studies Flexibly deployed at scales from several hectares to square dealing with forest health often make use of the 3D informa- kilometres [49], they allow forest practitioners to collect their tion and derived 2D spectral products that SfM photogram- own aerial information. In fact, there is an increasing interest metry delivers [18� , 50]. The following sections describe the in UAV forest surveys that can arguably be attributed to SfM- research on SfM-based forest inventory and health assess- based photogrammetric processing [26�� ]. ments to date in more detail. The rapid adoption of SfM photogrammetry is indicated by a growing number of scientific publications in forestry that utilize Inventory this photogrammetric technique. We conducted a search for peer-reviewed studies indexed by the Web of Knowledge data- Forest inventory holds a central role in all of the forest re- base using the keywords ‘Structure from Motion’, ‘UAV’ and search. Sustainable management of forests relies on knowl- ‘Forestry’ (and their most common variations). The search re- edge of their structure, distribution and dynamics over time sults were manually filtered to retain only forestry-related stud- [51]. The collection of field data for inventory purposes is ies applying a SfM-based workflow. We further categorized labour intensive, time-consuming and expensive, and cannot results into research on aerial and terrestrial inventory, forest be applied to large areas, consequently drastically limiting the health and proof-of-concept studies. These results are presented number of field inventories that can be afforded [52, 53]. in Table 2 and reveal a steady rise of publications on forest Efforts to improve on the efficiency of inventory practices remote sensing with SfM photogrammetry. therefore drive research in this field [53]. Amongst RS tech- SfM photogrammetry applications aimed at forest invento- nologies, SfM photogrammetry offers a low-cost and flexible ry are currently the most studied (Table 2). Here, a distinction approach to collect information on forest structure, thus natu- between aerial and terrestrial approaches can be made. An rally there has been an increase in interest to use such data for aerial approach typically utilizes a canopy surface model de- forest inventory. rived from SfM and/or associated spectral properties to esti- Within the context of forest inventory, the main use of SfM mate inventory relevant parameters [11]. Terrestrial photogrammetry has been its application on UAV image data 160 Curr Forestry Rep (2019) 5:155–168 Table 1 Overview of variables influencing the results of a SfM survey Domain Variable Recommendation Scene Texture High surface contrast to allow for feature-point detection Pattern repetition Increase overlap and increase accuracy of geotags Moving features Avoid! Occlusions Increase overlap and viewing angles Lighting conditions Sun angle High! Solar noon is ideal Weather Overcast provides even lighting (ambient occlusion) for structural (RGB) surveys. For spectral surveys little atmospheric influence may be required, clear skies. Changing illumination Avoid! Camera parameters Focal length Wide but not too wide to minimize distortions. 28–35 mm is a good basis (James et al. 2012) Exposure Δ Well exposed - Aperture Small for max DOF*, f/8 an advisable default - Shutter speed High for reduced motion blur*, ground speed (m/s) * exposure time (s) = blurred pixel -ISO Length Low for min noise*, auto-ISO an advisable default *Ideal scenario, but will always be a compromise between these three parameters Pixel pitch As high as is practical. Physical pixel size positively influences dynamic range and sensitivity Survey characteristics Overlap High (> 80% forward and lateral) as rule of thumb for forests to increase redundancy and matchability in scenes with high pattern repetition, moving features and/or occlusions. As a rule of thumb, a UAV-SfM data acquisition should be planned so that each point will be visible at least in 4–5images. View angles Convergent for reduction of systematic errors (RGB) Parallel (Nadir) for spectral sensing (reflectance) Survey range With increasing distance to the object/scene (decreasing GSD) survey precision degrades. Increased GSD requires higher overlap. Processing parameters SfM—matching If matching is not successful at full image scale ½ or ¼ may promote matchability - Image scale - Keypoints The number may be reduced for large datasets to reduce processing time MVS—densification Densification may not always be required at full image scale/maximum point cloud density Secondary products Multitude of algorithms for meshing, gridding etc. (results will depend on specific method) to produce wall-to-wall auxiliary information in a similar fash- image data, in recent years, there has been an increasing effort ion to ALS data. As such, UAV-SfM data has been shown to in developing terrestrial SfM applications to replace or aug- be suitable for the estimation of inventory relevant biophysical ment field data collections. The focus of studies incorporating parameters such as height, density and biomass [11, 12, 34, CRP lies on estimating diameter at breast height (DBH), tree 54–56]. Even though SfM has mostly been applied to aerial position and stem curves. The following sub-sections Table 2 Number of publications 2010 2013 2014 2015 2016 2017 2018 02/ on SfM photogrammetry for forest/tree remote sensing per year with manually assigned sub- SfM in forestry 1 3 3 11 18 31 66 4 categories - Airborne inventory 0 1 2 4 7 22 24 1 - Terrestrial inventory 0 0 0 1 4 2 10 0 - Forest health 0 0 0 2 4 1 7 1 - Proof-of-concept 1 2 1 4 3 6 15 2 Results presented are based on a manually filtered search in the scientific publications database Web of Knowledge using the search terms: TS = (‘Structure from Motion’ OR ‘Structure-from-Motion’ OR ‘SfM’ OR ‘sfm’ OR ‘structure from motion’ OR ‘structure-from-motion’ OR photogrammetry OR UAS OR SfM OR UAV OR RPAS OR drone OR CRP OR ‘unmanned aerial’ OR ‘Unmanned Aerial’) AND TS = (forest OR forestry OR tree). The date of this search was 21 February 2019 Curr Forestry Rep (2019) 5:155–168 161 elaborate further on the developments up to today regarding ALS data in terms of costs and accuracy. Goodbody et al. aerial and terrestrial SfM and highlight some of the key work [64] demonstrated the possibility to discriminate coniferous on using these photogrammetric data for inventory purposes. and deciduous species (overall accuracy of 86–95%). Puliti et al. [66� ] showed that UAV-SfM data could be used to accu- Aerial Inventory rately model stem density and height (RMSE% = 21.8% and 23.6%). Such results represent a substantial increase in accu- The use of SfM techniques applied to aerial image data for racy over ALS forest inventories and field assessment. forest inventory was pioneered by Dandois and Ellis in 2010 Furthermore, their study reported that data acquired using [54]. These authors were the first to use a series of unordered UAV-SfM techniques halved the amount of time required for but overlapping images acquired using a consumer-grade traditional field surveys that are commonly performed in re- camera mounted on a kite to produce a dense 3D point cloud generation stands. Thus, the use of UAV-SfM for regeneration representing the forest canopy. A first attempt to model forest forest may be particularly interesting since it allows a simul- biophysical properties using UAV-SfM data was done by taneous increase in the precision of the inventory while reduc- Dandois and Ellis in 2013 [34] and Lisein et al. in 2013 ing its costs. [55]. Both studies found that even though the results were Different methodological approaches have been applied to not consistent in all the studied areas, there was a correlation UAV-SfM data, similarly to ALS data. The methods can be between UAV-SfM data and variables such as dominant categorized into area-based approaches (ABA) [67] and indi- 2 2 height (R =0.07–0.91) or aboveground biomass (R =0.27– vidual tree crown (ITC) approaches [68, 69]. While in the 0.73). A more comprehensive evaluation of the possibilities to former case, the population units are represented by grid cells use UAV-SfM for forest inventories came with the studies by of area equal to that of the field plots; in the latter, they are Puliti et al. in 2015 [12] and Tuominen et al. in 2015 [56]who polygons representing single-tree crowns. In both cases, the extended their evaluation to the range of biophysical variables UAV-SfM data, corresponding either to the grid cells or the commonly used in forest management. Their results in terms single-tree crowns, are then linked to a sample of field obser- of RMSE% for dominant height (3.5%), Lorey’sheight vations either for field plots or for single trees through models. (13.3%–14.4%), stem density (38.6%), basal area (15.4– These models are then applied to all the population units either 23.9%) and timber volume (14.9–26.1%) were found to be for estimation of parameters for stand or forest level mapping. similar to errors associated with ALS-based forest inventories. The results of ABA methods have been presented in the pre- While these two studies set an important benchmark, they vious paragraph. The adoption of ITC approaches to UAV- were both conducted in even-aged managed boreal forests SfM has been found to be useful for detecting single trees with and thus provided limited information on how UAV-SfM 25–90% detection accuracy [63, 70, 71], to classify them ac- may perform in different forest types and forest developmental cording to tree species with overall accuracies up to 95% [71], stages. and measuring their height with RMSEs in the range of 0.5– Since the early days of UAV-SfM, the rapid growth in 2.84 m [55, 63]. In addition to rather large variability in the computing capabilities, availability of UAVs and SfM soft- accuracy of some of these variables, the results of UAV-SfM ware triggered increased interest in the scientific community ITC approaches vary according to forest types since they re- (see Table 1). This led to a widespread evaluation of UAV- main limited to the detection of the dominant tree layer, while SfM technology over a variety of forest types and forest de- smaller and dominated trees remain mostly undetected. velopmental stages. UAV-SfM data has been consistently proven to be useful for forest inventories in a large variety of Terrestrial Inventory forest types, including temperate European beech forests in Italy [13� ], mangrove forests in Malaysia [57], tropical forests Currently, terrestrial laser scanning (TLS) is the most accurate in Guyana [58], mixed conifer-broadleaved forest in Japan non-contact method of measurement to derive detailed forest [59], sparse sub-alpine coniferous forests in China [60], trop- inventory information at the plot level [15� ]. The main draw- ical woodlands in Malawi [41] and various plantations around backs of this technology are the high hardware costs [53], and the globe [61–63]. From these studies, a conclusion can be the time required for multiple scans mitigating occlusions drawn that the accuracy of UAV-SfM models is consistent along with post-processing to provide full coverage of a plot across many different forest types and on a similar scale to [72]. Mobile laser scanning systems reduce acquisition time ALS models. All of the aforementioned studies dealt with but high costs remain [73]. mature to nearly mature forest, while there has been little The deficiencies of traditional field data collection and the effort dedicated to estimating biophysical variables for forests need for reducing the cost of alternative laser scanning solu- under regeneration [64, 65, 66� ]. Nevertheless, the use of tions have encouraged the application of terrestrial photo- UAV-SfM data for regeneration forests may outperform alter- grammetry for forest inventory. Efforts to utilize terrestrial native data sources such as field assessments or the use of photogrammetric point clouds for deriving forest parameters 162 Curr Forestry Rep (2019) 5:155–168 derive from the low-cost of the equipment for the data collec- investigated by Liang et al. in 2014 and 2015 [74, 75]follow- tion, the automated SfM-based data processing and the poten- ed by Mokroš et al. in 2018 [78]. According to Mokroš et al. tially simple and fast data acquisition [74]. Requiring only a [78], the optimal acquisition solution resulted in portrait im- camera, typically handheld or mounted to a pole or tripod, ages, stop and go shooting mode and a path leading around the terrestrial SfM photogrammetry makes such a system highly plot with two diagonal paths through the plot. Differently, mobile, reducing the risk of occlusion yet providing a level of Liang et al. [75] concluded that the image matching results detail comparable to TLS [75]. of landscape images were optimal together with a Studies on terrestrial SfM for forestry purposes have be- photographing path based on inside and outside of an inner come more frequent in the last years (Table 1) and mainly circle (Fig. 4). For complex forest plots, Piermattei et al. [15� ] focus on linear rather than volumetric tree metrics. Studies found that the optimal acquisition path was a combination of vary according to (i) the scale of application, i.e. at plot level the solution found by Liang et al. [75] and Mokroš et al. [78]: and individual tree reconstruction; (ii) the measured forest landscape images, stop and go mode around the plot pointing parameters like tree position, DBH, height and stem curve; in, following by an inner circle pointing out of the plot and (iii) the resolution of the sensor, e.g. video, mobile phone two diagonals. This solution allowed reconstruction of stems and SLR camera; (iv) the camera configuration and photo- with an accuracy of few centimetres up to a few metres above graphic path and (v) the equipment used to acquire the images, ground. For low-density and medium-density forests, Liu e.g. pole, tripod, camera rig and backpack. Based on these et al. [77] proposed a system that combines two pole- aspects an overview of key work on terrestrial SfM applica- mounted cameras with a RTK GNSS for continuous captur- tions together with the obtained accuracies, acquisition meth- ing. Compared with total station measurements, their automat- od and geo-referencing approach are provided. ic determination of tree position, DBH and height achieved Most of the recent studies on photogrammetric measure- RMSEs of 0.16–0.2 m, 0.92–1.13 cm and 2.41–2.51 m ments of forest parameters are based on the single-camera respectively. technique, according to which overlapping images are ac- Most investigations on the use of terrestrial SfM were per- quired around the plot (Fig. 4). formed reconstructing single trees (i.e. not the entire forest Terrestrial photogrammetry has been evaluated in several plot) [16, 76, 79–83]. In those studies, DBH was the most studies in the past few years at plot scales [14, 15� , 16, 74, frequently estimated parameter and often compared with 76–78]. In these studies, DBH and tree locations were estimat- TLS data for accuracy assessment. Although sub-centimetre ed in circular plots with diameters ranging from 12 to 30 m. accuracy was achieved in all cases, the obtained RMSEs dif- The reported RMSE of the DBH ranged from 0.88 to 6.80 cm fered according to the approaches used, forest types and sur- compared with either field or TLS DBH measurements. Tree vey conditions, i.e. natural forest and controlled field settings. detection ranged between 60 and 98%. Results were influ- Not all the research studies report the time required for enced by the complexity of the forest plot, the acquisition path collecting the images. This can range from around 10 min to and mode. 2 h depending on the system used, parameters to be estimated, The impact of photographic path on the accuracy of forest plot size and survey configuration, by excluding the time to metrics derived from terrestrial SfM point clouds was firstly acquire scaling measurements. However, the accuracy of the Fig. 4 Example of a terrestrial SfM survey [51] in an open forest plot dense point cloud from the same point of view and c the dense point cloud showing a the configuration of camera positions and orientations, dense of a single stem without RGB colouring together with a 10-cm cross- point cloud and ground control points; b an example of an image and the section at 1.3 m (light blue point cloud) Curr Forestry Rep (2019) 5:155–168 163 scaling factor is crucial for forest plot and individual tree re- Näsi et al. [86] used hyperspectral image data combined construction [74]. To scale the photogrammetric point cloud, with the SfM-derived DSM for bark beetle damage detection most of the studies used targets surveyed by total station, at the individual tree level, achieving an overall classification implying additional equipment needs to be carried into the accuracy of 76% when distinguishing between healthy, field, consequently increasing the total acquisition time per infested and dead trees. In a follow -up study, Näsi et al. plot and reducing the portability of the entire surveying sys- [91] concluded that the individual tree-based approach, facil- tem. Aside from systems requiring a GNSS solution [77], itated by the combination of 3D and spectral data, provides a currently, only Liang et al. [74] tested natural reference ob- promising and cost-efficient alternative to field-based assess- jects, e.g. tree stems, for the determination of correct scale. ments of pest infestation. Minařík and Langhammer [92]also Their results showed that both natural reference objects and used a UAV-SfM-based mapping approach to map bark beetle artificial targets worked effectively. forest disturbance and found that bands from the red edge and NIR part of the spectrum were most suited for stress detection. Health Assessment and Monitoring These findings go in line with the results from Dash et al. [17], who assessed the potential of the commercially available mul- As part of a sustainable forest management, assessment and tispectral sensor, the Micasense RedEdge (Micasense Inc., monitoring of forest health condition play a crucial role. With Seattle, WA, USA), for the detection of early signs of stress threats to forest thought to increase globally [84], the identifi- during a simulated disease outbreak in a pine plantation. In the cation of declining forest health induced by biotic, abiotic and applied random forest classification of time-series data, nor- anthropogenic stress agents becomes imperative. RS ap- malized difference vegetation index (NDVI) showed to be the proaches offer rapid, spatially inclusive and objective ways best performing predictor variable to map physiological stress to monitor forest health when compared with field assess- symptoms along with the declining tree health. Further, late ments. With the aim of identifying and observing stress in examples of forest health monitoring are Baena et al. [93]and plants, multi- and hyperspectral sensors are capable of captur- Brovkina et al. [50], both successfully applying an OBIA ap- ing information outside the visible spectrum, which allow for proach on SfM-mapped NIR image data stemming from a estimation of biochemical plant traits like chlorophyll, leaf modified consumer RGB sensor to separate between dead pigments and canopy water content [85]. Spatially continuous and living trees. spectral mapping used to be exclusive to the manned airborne surveying domain. However, in recent years, lightweight sen- sors with discrete narrow spectral bands suitable for UAV Discussion mounting have become commercially available, allowing re- searchers to collect their own aerial spectral data [17, 86–88]. Used complementary to existing RS data (e.g. LiDAR) or by Such 2D spectral imagers may be used for SfM-based photo- itself, SfM photogrammetry has shown great potential for for- grammetric reconstruction and orthophoto generation similar- estry. Particularly attractive is the ability to use uncalibrated ly to RGB cameras, although they typically exhibit lower cameras paired with unstable or handheld platforms, enabling resolution. the use of low-cost and non-expert equipment. Ground and Opposed to the field of precision agriculture, where SfM- aerial SfM surveys can be carried out with high flexibility based processing of spectral image data is widely applied [89], offering the option for increased frequency RS surveys to in forestry there currently are only a few examples where incorporate, e.g. phenological changes in the analysis [59, SfM-derived mapped spectral products have found the appli- 65, 71]. The implementation of SfM algorithms in modern cation. Early UAV/SfM-based studies of forest health made photogrammetric software enables on-demand processing use of off-the-shelf RGB cameras, modified to capture near- with little required user input. SfM photogrammetry thus pre- infrared (NIR) images. Lehmann et al. [19] and Michez et al. sents a highly accessible and versatile solution to the acquisi- [18� ] used an object-based image analysis (OBIA) approach to tion of very high-resolution 3D data. In this regard, SfM em- segment and classify their scenes in order to identify declining powers common forestry practitioners to produce real-time tree health caused by biotic stress agents on both alder and data analytics with the minimum investment required for hard- oak. They achieved good overall classification accuracies ware and software. (79.5–84.1% and 82.5–85% at their respective study sites) Additional value in a SfM-based processing chain derives across five classes. However, they pointed out the limitations from the ability to provide multiple geospatial data products of NIR-modified RGB cameras, namely that visible and NIR (i.e. 3D models and orthomosaics) from a single sensor. spectra are not separable on the same sensor, spectral contam- Spectral information is inherently linked to the reconstructed ination due to broad and overlapping bands (see also Pauly structural data and derived products as these are generated [90]) as well as the inability to correct for changing light directly from the input imagery. Studies on forest health par- conditions (as downwelling irradiance is not captured). ticularly highlight the benefits of using the fused structural and 164 Curr Forestry Rep (2019) 5:155–168 spectral information that SfM-based processing of UAV image to further apply the method by Giannetti et al. [13� ]in a data provides [18� , 50]. wider variety of forest types and response variables. Point clouds generated from high-resolution images can Furthermore, the greater complexity of DTM- exhibit point densities greater than LiDAR, providing higher independent variables over more traditional explanatory detail information on the visible surface of forests. The in- variables could limit the transferability of the models creased spectral variation stemming from such high- through space and time. resolution data may hereby provide another valuable source 3) Lack of acquisition and processing protocols: of information, namely texture, such as the case in an OBIA The success of a SfM-based photogrammetric acquisi- approach [94]. Alongside the computational analysis, high- tion is largely based on the sensor used, the photographic resolution SfM-generated models appear visually realistic, path and viewing angles along with the chosen image providing experts a near true depiction of the scene. Intuitive overlap as well as the composition of a scene. to understand, SfM models thus hold an important advantage Adjustments to the acquisition approach to ensure quality over coarser remote sensing methods by enabling the rapid data are currently undertaken based on the surveyor’s visual assessment and/or validation. experience. Here protocols that enable certainty for SfM As is the case with all RS data, these will only ever be an outputs across forest types and phenological stages, yet approximation of the Earth’s surface and some limitations minimizing acquisition efforts, need to be established. always remain. With SfM photogrammetry being a new tech- Eltner et al. [24] suggested a protocol for the collection nology, the boundaries of these limitations are not fully tested of image data in geoscientific studies, which should be yet. Some of the main challenges with SfM photogrammetry extended to take into account forestry-specific factors. for forest applications that we are facing nowadays relate to Additional research is required on the parametrization of the following: SfM-based photogrammetric software for vegetated scenes specifically. Processing protocols designed to de- 1) Reproducibility: liver data adequate to the research question and to opti- With SfM photogrammetry enabling frequent surveys, mize processing speed are needed. variations in illumination, atmospheric and seasonal con- 4) Image matching issues: ditions are inevitable between acquisitions. Being a pas- Forests may prove to be challenging scenes for the sive sensing technique, these variations are directly feature matching algorithms underlying a SfM workflow. reflected on the data thus on the replicability of analyses. Their fine uniform texture, repeating patterns and poten- To ensure the use of SfM data on demand, allowing ac- tial movement (e.g. branches in wind) can introduce un- quisitions at different times of the year, it is therefore certainty in matching, consequently leading to incomplete reconstruction and/or noisy point clouds. In such cases, crucial to develop protocols for varying conditional sce- narios and models that account for variations in the data. the likelihood of identifying visual similarities in overlap- 2) Availability of accurate DTMs: ping images is promoted by increasing the distance to the Most airborne inventory studies presented here area of interest (AOI), thus increasing the number of fea- adopted highly accurate DTMs (e.g. ALS-based DTMs) tures per image and decreasing perspective distortions. to normalize UAV-SfM data and these are relatively rare Coarser ground sampling distances (GSDs) and higher around the globe, thus potentially limiting the area of image overlaps were shown to positively influence image application of UAV-SfM. To overcome this issue, some matching [95, 96]. The overlap should thus be increased authors proposed the use of DTMs generated from the when decreasing the GSD (i.e. images with finer detail). UAV-SfM data themselves [41] orthe use ofcoarsereso- Other potential mitigation strategies for reconstruction er- lution global DTMs such as shuttle radar topography mis- rors, like the use of high-accuracy position and orientation sion data (SRTM) [41]. While the former type of DTM is information for reduction of matching uncertainty, have obtainable only in open forests, the latter source was yet to be studied. found to be unsuitable for estimation of aboveground bio- mass. A conceptually novel approach came with the study To widely employ SfM photogrammetry in operational for- by Giannetti et al. [13� ] who, to overcome any of the estry, future research needs to tackle the abovementioned hur- abovementioned limitations, proposed the use of UAV- dles. It is essential to develop a consensus on acquisition pro- SfM data-derived variables without prior normalization tocols and parametrization of SfM photogrammetry software (i.e. DTM-independent variables). Their results showed that is set to answer specific research questions across forest that models fitted raw UAV-SfM data alone predicted types and environmental conditions. We have started to gain stem volume with similar accuracy to ALS data, even in some understanding of how image quality, overlap, GSD and the highly productive broadleaf forest in steep terrain. photographic path are influencing SfM-based reconstruction Despite such encouraging results, it remains fundamental [15� , 43�� , 95–97]. However, prior to processing, uncertainty Curr Forestry Rep (2019) 5:155–168 165 Human and Animal Rights and Informed Consent This article does not remains in predicting the completeness of these photogram- contain any studies with human or animal subjects performed by any of metric models. More in-depth work on these influential pa- the authors. rameters is needed in conjunction with the development of Open Access This article is distributed under the terms of the Creative methods that allow for reliable quality estimation of SfM- Commons Attribution 4.0 International License (http:// based outputs. Towards the quantification of data quality, creativecommons.org/licenses/by/4.0/), which permits unrestricted use, James et al. [98] presented a method for estimating the preci- distribution, and reproduction in any medium, provided you give appro- sion of each point produced within the SfM pipeline by re- priate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. peatedly running bundle adjustments on a set of input images. These ‘precision maps’ allow practitioners to describe the spatial variability of precision within SfM-derived products and gain insight into limitations in a given survey (such as image quality or control-point measurements). To our knowl- References edge, ‘precision maps’ have not been applied in the context of forested scenes. In forestry, future studies would benefit from Papers of particular interest, published recently, have been this method to objectively describe the data quality of SfM- highlighted as: derived products and thereby reduce uncertainty in subsequent � Of importance analysis. �� Of major importance 1. Mcroberts R, Tomppo E. Remote sensing support for national forest inventories. Remote Sens Environ. 2007;110:412–9. https://doi. org/10.1016/j.rse.2006.09.034. Conclusions 2. Wulder M. Optical remote-sensing techniques for the assessment of forest inventory and biophysical parameters. Prog Phys Geogr. A camera and a computer are the basic requirements for 1998;22:449–76. SfM photogrammetry. With the examples given here, 3. Korpela I. Individual tree measurements by means of digital aerial and in terms of what valuable data may be extracted photogrammetry. Silva Fennica Monographs 2004;3:93. 4. Lim K, Treitz P, Wulder M, St-Onge B, Flood M. LiDAR remote from SfM-derived data by analysis, SfM photogramme- sensing of forest structure. Prog Phys Geogr Earth Environ. try shows great potential for forest practitioners and 2003;27:88–106. researchers. Adding the power of UAVs for the acquisi- 5. Næsset E. Airborne laser scanning as a method in operational forest tion of aerial image data, the canopy of a forest can be inventory: status of accuracy assessments accomplished in mapped nearly in real time, responding rapidly to man- Scandinavia. Scand J For Res. 2007;22:433–42. 6. Kangas A, Gobakken T, Puliti S, Hauglin M, Næsset E. Value of agement needs. The temporal and spatial dimension that airborne laser scanning and digital aerial photogrammetry data in can be provided with SfM photogrammetry enables as- forest decision making. Silva Fennica 2018;52:1–19. sessment and monitoring of forests in an economical 7. Ginzler C, Hobi ML. Countrywide stereo-image matching for way that has not existed before. updating digital surface models in the framework of the swiss na- However, constraints linked to the fundamental principles tional forest inventory. Remote Sens. 2015;7:4343–70. 8. Koch B. Remote sensing supporting national forest inventories. In of SfM photogrammetry being a passive optical method will Knowledge Reference for National Forest Assessments; FAO. remain. Influencing factors like viewing geometry, lighting 2013:1–18. and the availability of static texture have to be carefully con- 9. White JC, Wulder MA, Vastaranta M, Coops NC, Pitt D, Woods M. sidered prior to every survey. We suggest that, in order to The utility of image-based point clouds for forest inventory: a com- parison with airborne laser scanning. Forests. 2013;4:518–36. enable the collection of fit-for-purpose data with predictable 10. Leberl F, Irschara A, Pock T, Meixner P, Gruber M, Scholz S, et al. quality, further work is needed on acquisition and processing Point clouds: Lidar versus 3D vision. Photogramm Eng Remote protocols specific to forestry. Further progress in these areas Sens. 2010;76:1123–34. will facilitate the move away from proof-of-concept studies 11. Goodbody TRH, Coops NC, White JC. Digital aerial photogram- and towards the operational application of SfM photogram- metry for updating area-based forest inventories: a review of oppor- tunities, challenges, and future directions. Curr For Rep. 2019;5: metry in forestry. 55–75. https://doi.org/10.1007/s40725-019-00087-2. 12. Puliti S, Ørka HO, Gobakken T, Naesset E. Inventory of small Funding Jakob Iglhaut is funded by the Royal Society Research Grant forest areas using an unmanned aerial system. Remote Sens. RG140494. Stefano Puliti is funded by NIBIO. 2015;7:9632–54. 13.� Giannetti F, Chirici G, Gobakken T, Næsset E, Travaglini D, Puliti Compliance with Ethical Standards S. A new approach with DTM-independent metrics for forest grow- ing stock prediction using UAV photogrammetric data. Remote Conflict of Interest Jakob Iglhaut, Carlos Cabo, Stefano Puliti, Livia Sens Environ. 2018;213:195–205. By omitting the prerequisite of DTM availability their approach indicates forward-looking Piermattei, James O’Connor, and Jacqueline Rosette declare that they potential for inventories based solely on SfM photogrammetry. have no conflict of interest. 166 Curr Forestry Rep (2019) 5:155–168 14. Mikita T, Janata P, Surový P. Forest stand inventory based on com- 31. Fonstad MA, Dietrich JT, Courville BC, Jensen JL, Carbonneau bined aerial and terrestrial close-range photogrammetry. Forests. PE. Topographic structure from motion: a new development in 2016;7:1–14. photogrammetric measurement. Earth Surf Process Landf. 2013;38:421–30. 15.� Piermattei L, Karel W, Wang D, Wieser M, Mokroš M, Koreň M, 32. Micheletti N, Chandler JH, Lane SN. Structure from motion (SfM) et al. Terrestrial Structure from Motion photogrammetry for deriv- photogrammetry. Br Soc Geomorphol. 2015. ing forest inventory data. Remote Sens. 2019. https://doi.org/10. 3390/rs11080950. This paper deals with the detection of tree 33. Skarlatos D, Kiparissi S. Comparison of laser scanning, photogram- metry and Sfm-Mvs pipeline applied in structures and artificial characteristics from close-range SfM photogrammetry and surfaces. ISPRS Ann Photogramm Remote Sens Spat Inf Sci. TLS in four case studies. The two techniques with respective 2012;I–3:299–304. acquisition and post-processing phases and their source re- 34. Dandois JP, Ellis EC. High spatial resolution three-dimensional quirements are discussed thoroughly. mapping of vegetation spectral dynamics using computer vision. 16. Mokroš M, Výbošťok J, Tomaštík J, Grznárová A, Valent P, Slavík Remote Sens Environ. 2013;136:259–76. https://doi.org/10.1016/ M, et al. High precision individual tree diameter and perimeter j.rse.2013.04.005. estimation from close-range photogrammetry. Forests. 2018;9. 35. Lowe DG. Distinctive image features from Scale-Invariant https://doi.org/10.3390/f9110696. Keypoints. Int J Comput Vis. 2004;60:91–110. 17. Dash JP, Watt MS, Pearse GD, Heaphy M, Dungey HS. Assessing 36. Shum H-Y, Ke Q, Zhang Z. Efficient bundle adjustment with virtual very high resolution UAV imagery for monitoring forest health key frames: a hierarchical approach to multi-frame structure from during a simulated disease outbreak. ISPRS J Photogramm motion. In: Proceedings. 1999 IEEE Comput. Soc. Conf. Comput. Remote Sens. 2017;131:1–14. Vis. Pattern Recognit. (Cat. No PR00149). IEEE Comput. Soc, pp 18.� Michez A, Piégay H, Lisein J, Claessens H, Lejeune P. 538–543. Classification of riparian forest species and health condition using 37. Ahmadabadian AH, Robson S, Boehm J, Shortis M, Wenzel K, multi-temporal and hyperspatial imagery from unmanned aerial Fritsch D. A comparison of dense matching algorithms for scaled system. Environ Monit Assess. 2016;188:146. A great example surface reconstruction using stereo camera rigs. ISPRS J of forest health monitoring at the individual tree level based on Photogramm Remote Sens. 2013;78:157–67. SfM-derived structural, spectral and textural variables. 38. Gini R, Sona G, Ronchetti G, Passoni D, Pinto L. Improving tree 19. Lehmann JRK, Nieberding F, Prinz T, Knoth C. Analysis of un- species classification using UAS multispectral images and texture manned aerial system-based CIR images in forestry-a new perspec- measures. ISPRS Int J Geo-Information. 2018;7:315. tive to monitor pest infestation levels. Forests. 2015;6:594–612. 39. Puliti S. Tree-stump detection, segmentation, classification, and https://doi.org/10.3390/f6030594. measurement using unmanned aerial vehicle (UAV) imagery. 20. Lisein J, Michez A, Claessens H, Lejeune P. Discrimination of Forests. 2018. https://doi.org/10.3390/f9030102 . deciduous tree species from time series of unmanned aerial system 40. Puliti S, Gobakken T, Ørka HO, Næsset E. Assessing 3D point imagery. PLoS One. 2015;10:1–20. clouds from aerial photographs for species-specific forest invento- 21. Alonzo M, Andersen H-E, Morton D, Cook B. Quantifying boreal ries. Scand J For Res. 2017;32:68–79. forest structure and composition using UAV structure from motion. 41. Kachamba D, Ørka H, Gobakken T, Eid T, Mwase W, Kachamba Forests. 2018;9:119. DJ, et al. Biomass estimation using 3D data from unmanned aerial 22. Saarinen N, Vastaranta M, Näsi R, Rosnell T, Hakala T, Honkavaara vehicle imagery in a tropical woodland. Remote Sens. 2016;8:968. E, et al. Assessing biodiversity in boreal forests with UAV-based 42. White JC, Wulder MA, Varhola A, Vastaranta M, Coops NC, Cook photogrammetric point clouds and hyperspectral imaging. Remote BD, et al. A best practices guide for generating forest inventory Sens. 2018;10. https://doi.org/10.3390/rs10020338. attributes from airborne laser scanning data using an area-based 23. Smith MW, Carrivick JL, Quincey DJ. Structure from motion pho- approach. Canadian Forest Service: Information Report. 2017. togrammetry in physical geography. Prog Phys Geogr. 2015;40: https://doi.org/10.5558/tfc2013-132 . 247–75. 43.�� O’Connor J, Smith MJ, James MR. Cameras and settings for aerial 24. Eltner A, Kaiser A, Castillo C, Rock G, Neugirg F, Abellán A. surveys in the geosciences: optimising image data. Prog Phys Image-based surface reconstruction in geomorphometry-merits, Geogr. 2017;41:325–44 This paper gives a very clear introduc- limits and developments. Earth Surf Dyn. 2016;4:359–89. https:// tion to aerial image capture. Foundational principles essential doi.org/10.5194/esurf-4-359-2016. to capturing high quality imagery with digital sensors are pro- 25. Bemis SP, Micklethwaite S, Turner D, James MR, Akciz S, Thiele vided. Cameras and settings are discussed with worked S, et al. Ground-based and UAV-based photogrammetry: a multi- examples. scale, high-resolution mapping tool for structural geology and 44. Mosbrucker AR, Major JJ, Spicer KR, Pitlick J. Camera system paleoseismology. J Struct Geol. 2014;69:163–78. considerations for geomorphic applications of SfM photogramme- 26.�� Carrivick J, Smith MJ. Quincey D. Structure from motion in the try. Earth Surf Process Landf. 2017;42:969–86. geosciences. 2016. This book further elaborates on the concepts 45. Verhoeven GJJ. It’s all about the format – unleashing the power of and methods of SfM photogrammetry applied in physical ge- RAW aerial photography. Int J Remote Sens. 2010;31:2009–42. ography. Technical aspects are described with great clarity. 46. James MR, Robson S. Mitigating systematic error in topographic 27. Bolles RC, Baker HH, Marimont DH. Epipolar-plane image anal- models derived from UAVand ground-based image networks. Earth ysis: an approach to determining structure from motion. Int J Surf Process Landf. 2014;39:1413–20. Comput Vis. 1987;1:7–55. 47. Wackrow R, Chandler JH. Minimising systematic error surfaces in 28. Ullman S. The interpretation of structure from motion. Proc R Soc digital elevation models using oblique convergent imagery. Lond Ser B Biol Sci. 1979;203:405–26. Photogramm Rec. 2011;26:16–31. 29. Koenderink JJ, van Doorn AJ. Affine structure from motion. J Opt 48. Anderson K, Westoby MJ, James MR. Low-budget topographic Soc Am A. 1991;8:377–85. surveying comes of age: structure from motion photogrammetry 30. Westoby MJ, Brasington J, Glasser NF, Hambrey MJ, Reynolds in geography and the geosciences. Prog Phys Geogr Earth Environ. 2019;43:163–73. JM. “Structure-from-motion” photogrammetry: a low-cost, effec- tive tool for geoscience applications. Geomorphology. 2012;179: 49. Anderson K, Gaston KJ. Lightweight unmanned aerial vehicles will 300–14 revolutionize spatial ecology. Front Ecol Environ. 2013;11:138–46 . . Curr Forestry Rep (2019) 5:155–168 167 50. Brovkina O, Cienciala E, Surový P, Janata P. Unmanned aerial UAV-SfM data can outperform ALS and traditional field data in forest inventory. vehicles (UAV) for assessment of qualitative classification of Norway spruce in temperate forest stands. Geo-Spat Inf Sci. 67. Næsset E, Økland T. Estimating tree height and tree crown proper- 2018;21:12–20. ties using airborne scanning laser in a boreal nature reserve. Remote Sens Environ. 2002;79:105–15. 51. Liang X, Kukko A, Hyyppä J, Lehtomäki M, Pyörälä J, Yu X, et al. 68. Brandtberg T. Automatic individual tree based analysis of high In-situ measurements from mobile platforms: an emerging ap- spatial resolution aerial images on naturally regenerated boreal for- proach to address the old challenges associated with forest inven- ests. Can J For Res. 1999;29:1464–78. tories. ISPRS J Photogramm Remote Sens. 2018;143:97–107. 69. Hyyppä J, Inkinen M. Detecting and estimating attributes for single 52. Thompson ID, Maher SC, Rouillard DP, Fryxell JM, Baker JA. trees using laser scanner. Photogramm J Finl. 1999;16:27–42. Accuracy of forest inventory mapping: some implications for boreal 70. Mohan M, Silva C, Klauberg C, Jat P, Catts G, Cardil A, et al. forest management. For Ecol Manag. 2007;252:208–21. ndividual tree detection from unmanned aerial vehicle (UAV) de- 53. White JC, Coops NC, Wulder MA, Vastaranta M, Hilker T, rived canopy height model in an open canopy mixed conifer forest. Tompalski P. Remote sensing technologies for enhancing forest Forests. 2017;8:340. inventories: a review. Can J Remote Sens. 2016;42:619–41. 71. Nevalainen O, Honkavaara E, Tuominen S, Viljanen N, Hakala T, 54. Dandois JP, Ellis EC. Remote sensing of vegetation structure using Yu X, et al. Individual tree detection and classification with UAV- computer vision. Remote Sens. 2010;2:1157–76. based photogrammetric point clouds and hyperspectral imaging. 55. Lisein J, Pierrot-Deseilligny M, Bonnet S, Lejeune P. A photogram- Remote Sens. 2017;9:185. metric workflow for the creation of a forest canopy height model 72. Liang X, Kankare V, Hyyppä J, Wang Y, Kukko A, Haggrén H, from small unmanned aerial system imagery. Forests. 2013;4:922– et al. Terrestrial laser scanning in forest inventories. ISPRS J Photogramm Remote Sens. 2016;115:63–77. 56. Tuominen S, Balazs A, Saari H, Pölönen I, Sarkeala J, Viitala R. 73. Wulder MA, Bater CW, Coops NC, Hilker TH, White JC. The role Unmanned aerial system imagery and photogrammetric canopy of LiDAR in sustainable forest management. For Chron. 2008;84: height data in area-based estimation of forest variables. Silva 807–826. Fenn. 2015. https://doi.org/10.14214/sf.1348. 74. Liang X, Jaakkola A, Wang Y, Hyyppä J, Honkavaara E, Liu J, et al. 57. Otero V, Van De Kerchove R, Satyanarayana B, Martínez-Espinosa The use of a hand-held camera for individual tree 3D mapping in C, Bin FMA, Bin IMR, et al. Managing mangrove forests from the forest sample plots. Remote Sens. 2014;6:6587–603. sky: forest inventory using field data and unmanned aerial vehicle 75. Liang X, Wang Y, Jaakkola A, Kukko A, Kaartinen H, Hyyppä J, (UAV) imagery in the Matang Mangrove Forest Reserve, peninsu- et al. Forest data collection using terrestrial image-based point lar Malaysia. For Ecol Manag. 2018;411:35–45. clouds from a handheld camera compared to terrestrial and personal 58. Roşca S, Suomalainen J, Bartholomeus H, Herold M. Comparing laser scanning. IEEE Trans Geosci Remote Sens. 2015;53:5117– terrestrial laser scanning and unmanned aerial vehicle structure 32. https://doi.org/10.1109/TGRS.2015.2417316. from motion to assess top of canopy structure in tropical forests. 76. Surový P, Yoshimoto A, Panagiotidis D. Accuracy of reconstruction Interface Focus. 2018;8:20170038. of the tree stem surface using terrestrial close-range photogramme- 59. Jayathunga S, Owari T, Tsuyuki S. Evaluating the performance of try. Remote Sens. 2016;8:1–13. photogrammetric products using fixed-wing UAV imagery over a 77. Liu J, Feng Z, Yang L, Mannan A, Khan T, Zhao Z, et al. Extraction mixed conifer-broadleaf forest: comparison with airborne laser of sample plot parameters from 3D point cloud reconstruction based scanning. Remote Sens. 2018;10. https://doi.org/10.3390/ on combined RTK and CCD continuous photography. Remote rs10020187. Sens. 2018;10:1299. 60. Lin J, Wang M, Ma M, Lin Y, Lin J, Wang M, et al. Aboveground 78. Mokroš M, Liang X, Surový P, Valent P, Čerňava J, Chudý F. tree biomass estimation of sparse subalpine coniferous forest with Evaluation of close-range photogrammetry image collection UAV oblique photography. Remote Sens. 2018;10:1849. methods for estimating tree diameters. Int J Geo-Inf. 2018;7:93. 61. Shin P, Sankey T, Moore M, Thode A, Shin P, Sankey T, et al. 79. Morgenroth J, Gomez C. Assessment of tree structure using a 3D Evaluating unmanned aerial vehicle images for estimating forest image analysis technique-a proof of concept. Urban For Urban canopy fuels in a ponderosa pine stand. Remote Sens. 2018;10: Green. 2014;13:198–203. 80. Miller JM. Estimation of individual tree metrics using structure- 62. Iizuka K, Yonehara T, Itoh M, Kosugi Y. Estimating tree height and from-motion photogrammetry. MSc Thesis. 2015. diameter at breast height (DBH) from digital surface models and 81. Bauwens S, Fayolle A, Gourlet-Fleury S, Ndjele LM, Mengal C, orthophotos obtained with an unmanned aerial system for a Lejeune P. Terrestrial photogrammetry: a non-destructive method Japanese Cypress (Chamaecyparis obtusa) Forest. Remote Sens. for modelling irregularly shaped tropical tree trunks. Methods Ecol 2018;10. https://doi.org/10.3390/rs10010013. Evol. 2017;8:460–71. 63. Guerra-Hernández J, Cosenza DN, Rodriguez LCE, Silva M, Tomé 82. Berveglieri A, Tommaselli AMG, Liang X, Honkavaara E. Vertical M, Díaz-Varela RA, et al. Comparison of ALS- and UAV(SfM)- optical scanning with panoramic vision for tree trunk reconstruc- derived high-density point clouds for individual tree detection in tion. Sensors (Switzerland). 2017;17:1–13. Eucalyptus plantations. Int J Remote Sens. 2018;39:5211–35. 83. Huang H, Zhang H, Chen C, Tang L. Three-dimensional digitiza- 64. Goodbody TRH, Coops NC, Hermosilla T, Tompalski P, Crawford tion of the arid land plant Haloxylon ammodendron using a P. Assessing the status of forest regeneration using digital aerial consumer-grade camera. Ecol Evol. 2018;8:5891–9. photogrammetry and unmanned aerial systems. Int J Remote 84. LauschA,ErasmiS,KingD,MagdonP, Heurich M. Sens. 2018;39:5246–64. Understanding forest health with remote sensing -part I—areview 65. Feduck C, McDermid G, Castilla G, Feduck C, McDermid GJ, of spectral traits, processes and remote-sensing characteristics. Castilla G. Detection of coniferous seedlings in UAV imagery. Remote Sens. 2016;8:1029. Forests. 2018;9:432. 85. Stone C, Mohammed C. Application of remote sensing technolo- 66.� Puliti S, Solberg S, Granhus A. Use of UAV photogrammetric data gies for assessing planted forests damaged by insect pests and fun- gal pathogens: a review. Curr For Rep. 2017;3:75–92. for estimation of biophysical properties in forest stands under re- generation. Remote Sens. 2019;11:233. https://doi.org/10.3390/ 86. Näsi R, Honkavaara E, Lyytikäinen-Saarenmaa P, Blomqvist M, rs11030233 . This paper demonstrates that high resolution Litkey P, Hakala T, et al. Using UAV-based photogrammetry and 168 Curr Forestry Rep (2019) 5:155–168 hyperspectral imaging for mapping bark beetle damage at tree-lev- 94. Blaschke T. Object based image analysis for remote sensing. ISPRS J Photogramm Remote Sens. 2010;65:2–16. el. Remote Sens. 2015;7:15467–93. 95. Frey J, Kovach K, Stemmler S, Koch B. UAV photogrammetry of 87. Minařík R, Langhammer J. Use of a multispectral UAV photogram- forests as a vulnerable process. A sensitivity analysis for a structure metry for detection and tracking of forest disturbance dynamics. Int from motion RGB-image pipeline. Remote Sens. 2018;10. https:// Arch Photogramm Remote Sens Spat Inf Sci. 2016;41:711–8. doi.org/10.3390/rs10060912. 88. Aasen H, Bareth G. Spectral and 3D nonspectral approaches to crop 96. Ni W, Sun G, Pang Y, Zhang Z, Liu J, Yang A, et al. Mapping three- trait estimation using ground and UAV sensing. In Biophys. dimensional structures of forest canopy using UAV stereo imagery: Biochem. Charact. Plant Species Stud: CRC Press; 2019;103–32. evaluating impacts of forward overlaps and image resolutions with 89. Hunt ER, Daughtry CST. What good are unmanned aircraft systems LiDAR data as reference. IEEE J Sel Top Appl Earth Obs Remote for agricultural remote sensing and precision agriculture? Int J Sens. 2018;11:3578–89. Remote Sens. 2018;39:5345–76. 97. Dandois JP, Olano M, Ellis EC. Optimal altitude, overlap, and 90. Pauly K. Towards calibrated vegetation indices from UAS-derived weather conditions for computer vision uav estimates of forest orthomosaics. 13th Int Conf Precis Agric. 2016. https://doi.org/10. structure. Remote Sens. 2015;7:13895–920. https://doi.org/10. 13140/RG.2.2.21842.35524. 3390/rs71013895. 91. Näsi R, Honkavaara E, Blomqvist M, Lyytikäinen-Saarenmaa P, 98. James MR, Robson S, Smith MW. 3-D uncertainty-based topo- Hakala T, Viljanen N, et al. Remote sensing of bark beetle damage graphic change detection with structure-from-motion photogram- in urban forests at individual tree level using a novel hyperspectral metry: precision maps for ground control and directly camera from UAV and aircraft. Urban For Urban Green. 2018;30: georeferenced surveys. Earth Surf Process Landf. 2017;42:1769– 72–83. 88. https://doi.org/10.1002/esp.4125. 92. Minařík R, Langhammer J. Use of a multispectral Uav photogram- metry for detection and tracking of forest disturbance dynamics. Int Publisher’sNote Springer Nature remains neutral with regard to Arch Photogramm Remote Sens Spat Inf Sci. 2016;XLI-B8:711–8. jurisdictional claims in published maps and institutional affiliations. 93. Baena S, Moat J, Whaley O, Boyd DS. Identifying species from the air: UAVs and the very high resolution challenge for plant conser- vation. PLoS One. 2017;12:1–21.

Journal

Current Forestry ReportsSpringer Journals

Published: Jul 16, 2019

There are no references for this article.