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

Learn More →

A 3D approach to reconstruct continuous optical images using lidar and MODIS

A 3D approach to reconstruct continuous optical images using lidar and MODIS Background: Monitoring forest health and biomass for changes over time in the global environment requires the provision of continuous satellite images. However, optical images of land surfaces are generally contaminated when clouds are present or rain occurs. Methods: To estimate the actual reflectance of land surfaces masked by clouds and potential rain, 3D simulations by the RAPID radiative transfer model were proposed and conducted on a forest farm dominated by birch and larch in Genhe City, DaXing’AnLing Mountain in Inner Mongolia, China. The canopy height model (CHM) from lidar data were used to extract individual tree structures (location, height, crown width). Field measurements related tree height to diameter of breast height (DBH), lowest branch height and leaf area index (LAI). Series of Landsat images were used to classify tree species and land cover. MODIS LAI products were used to estimate the LAI of individual trees. Combining all these input variables to drive RAPID, high-resolution optical remote sensing images were simulated and validated with available satellite images. Results: Evaluations on spatial texture, spectral values and directional reflectance were conducted to show comparable results. Conclusions: The study provides a proof-of-concept approach to link lidar and MODIS data in the parameterization of RAPID models for high temporal and spatial resolutions of image reconstruction in forest dominated areas. Keywords: Lidar; Optical; Temporal interpolation; 3D model; High resolution Background A few studies have been conducted to interpolate those Optical remote sensing images have been widely used contaminated images by rains or clouds. For example, in monitoring forest ecosystems. Spatial, temporal and Landsat images have been blended with MODerate- spectral resolutions are the three key indicators to be resolution Imaging Spectroradiometer (MODIS) data to considered during most applications. Spatial resolution create spatial and temporal fusion data (Gao et al. 2006; hadbeenimprovedfroma scaleof hundredsofmeters Hilker et al. 2009; Wu et al. 2012). Further, radiative (e.g. Landsat 8) to one of a half-meter (e.g. GeoEye-1 or transfer models have also been used to simulate a series Worldview-2) with only a slight increase in the number of high temporal resolution images for future space of spectral bands. However, in forested area, temporal earth observation missions (Inglada et al. 2011). How- resolution is generally reduced by frequent rains or ever, spatial resolution is generally moderate due to the cloud covers, which prevents users from continuously useofsimplehomogeneous radiativetransfermodels, acquiring clear optical remote sensing images. which are not able to deal with high resolution simula- Temporally continuous satellite images are important tion with diverse tree species and mountain shadows. for forest monitoring (Lunetta et al. 2004; Masek et al. In recent years, light detection and ranging (lidar) has 2008; Nitze et al. 2015) since forest reflectance varies been a widely used tool for forest studies (Adams et al. with seasonality (Kobayashi et al. 2007; Xu et al. 2013). 2012; Arno et al. 2013; Montesano et al. 2013). The greatest advantage of lidar is to provide direct measure- ments of very detailed 3D forest structures, so it can be * Correspondence: huaguo.huang@gmail.com used to reconstruct 3D trees to support the simulation Key Laboratory for Silviculture and Conservation of Ministry of Education, Beijing Forestry University, Beijing, China © 2015 Huang and Lian. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited. Huang and Lian Forest Ecosystems (2015) 2:20 Page 2 of 13 of radar remote sensing signals (Lucas et al. 2006) and The growing season typically begins in early May and to study how 3D structures affect the quality of optical senescence occurs in late September. In the summer of images (Barbier et al. 2011). By coupling lidar with high 2013, 18 field plots (45 m by 45 m) were established temporal data such as MODIS using 3D radiative trans- representing different combinations of forest types, density fer models, it will be possible to generate both high and leaf area index (LAI) (Table 1). The LAI, ranging from 2 −2 spatial and temporal resolution optical remote sensing 1.44 to 3.51 m ∙m , was measured using LAI-2000 images. However, very few studies were found using (LICOR Inc.) hemispheric data. The forest cover varies be- this approach. Therefore, we will test the possibility of tween 0.21 and 0.86. that approach to simulate high-resolution optical satellite Based on inventory data of individual tree structures images on an arbitrary day of the growing season in a for- in plots L1 to L9, the DBH and crown length (L) of trees ested area. were regressed on tree height (H), where heights were derived by lidar. For convenience, both crowns of larch Methods and birch are defined as spherical in shape. Study site Reflectance or transmittance spectra of leaves, branches The study site (Fig. 1), located in a 100 ha forested area and stems of birch and larch trees were measured in (50°54′ N, 121°54′ E) of the Genhe Forestry Reserve, the field using the integrating sphere of ASD (Analytical DaXing’AnLing Mountain in Inner Mongolia, China, be- Spectral Device, http://www.asdi.com/). Dry and wet soil longs to a boreal moist and cold temperature forest, with spectra were the default soil spectra in a PROSAIL model an elevation ranging from 784 to 1142 m. Annual average (Jacquemoud et al. 2009). The re-sampled spectral curves precipitation is 450 to 550 mm, with sixty per cent falling are shown in Fig. 2. in July and August. Annual average sunshine is 2594 h with a frost-free period of 80 days. Our study site occupied Airborne data 75 % of the total area. The forest is mainly composed of Small footprint full-waveform lidar data were acquired Dahurian Larch (Larix gmelinii) and White Birch (Betula from August 16 to September 25 in 2012 (Mu et al. 2015). platyphylla Suk.). The understory vegetation of the larch The system consisted of a Leica ALS60 with an integrated forest is a single layer of evergreen shrubs (normally Leica RCD105 camera. The CCD camera produced nat- Ledum palustre L. or Rhododendron dauricum L.). L. ural color mosaic images with 0.2 m resolution. As for palustre is generally a low shrub (less than 0.3 m), while lidar, the mean swath width was 1 km at a flying altitude the height of R. dauricum is around 1.5 m. Blueberries of approximately 2700 m (over rough terrain). The scan (Semen trigonellae) are widely distributed. The birch forest angle was less than 35°. Waveforms were digitized with a has a understory of grass or deciduous shrubs, such as frequency of 100 to 200 kHz. An average of eight reflected Rosa acicularis, Spiraea sericea Turcz., or Rubus L. pulses per m was obtained over the sample plots. Point Fig. 1 Location of the Genhe study site in DaXing’AnLing Mountain, Inner Mongolia, China Huang and Lian Forest Ecosystems (2015) 2:20 Page 3 of 13 Table 1 Plot variables at the study site 2 −2 Stand ID Tree stems (Birch: Larch) Slope (°) Shrub height (m) and cover Mean tree height (m) Mean DBH (cm) LAI (m ∙m ) A1 83:140 35 1.5, 70 % 11.6 11.8 2.68 ± 0.13 A2 97: 248 20 1.8, 60 % 12.1 11.6 2.88 ± 0.24 A3 41:269 25 1.5, 75 % 11.0 11.4 2.56 ± 0.09 A4 174:22 5 0.7, 3 % 16.0 16.1 3.31 ± 0.07 A5 142:103 10 2.0, 80 % 11.8 11.8 No data A6 91:173 45 0.3, 80 % 11.2 11.5 No data A7 111:102 30 1.5, 50 % 14.2 16.2 2.06 ± 0.13 A8 156:238 10 No data 11.0 9.5 2.51 ± 0.20 A9 65:180 45 1.3, 70 % 14.5 15.4 2.36 ± 0.17 L1 131:220 5 1.2, 20 % 8.4 10.0 2.71 ± 0.25 L2 89:421 <5 1.2, 5 % 9.0 8.7 3.06 ± 0.22 L3 11:71 5 0.7, 3 % 8.0 7.8 2.42 ± 0.50 L4 294:79 5 1.2, 15 % 11.7 11.6 3.51 ± 0.20 L5 1:327 7 1.0, 5 % 12.5 12.9 2.43 ± 0.12 L6 0:173 5 0.5, 20 % 17.2 22.3 1.44 ± 0.06 L7 11:90 7 1.0, 80 % 9.7 12.5 1.50 ± 0.15 L8 60:118 12 1.3, 30 % 13.1 15.4 2.70 ± 0.12 L9 0:585 5 0.5, 50 % 8.4 8.7 2.28 ± 0.09 clouds were first classified by the TerraScan software (see Optical satellite images www.terrasolid.com) to separate the ground points from Several scales of geo-referenced satellite images were used, other points. We used Delaunay triangulation and bilinear including SPOT-6 (1 m), Landsat (30 m) and MODIS interpolation method to generate a digital elevation model (250 m). Due to frequent cloud cover, SPOT and Landsat (DEM) from ground returns. DSM (digital surface model) were not able to capture clear land surface images during was created using a maximum value in a window size of rainy days, which happened mostly in the growing season 0.5 m. CHM (canopy height model) was calculated as the (May to September). There was only one cloud-free difference between DSM and DEM (Fig. 3). SPOT-6 image, obtained on October 10, 2013. For the Individual tree crowns were segmented from CHM same reason, only three Landsat images were clear (May using the “TreeVaW” tool (Popescu and Wynne 2004), 5, 2013; Sep 9, 2013; Sep 29, 2013) in 2013. As well, we which uses a circular window filter to segment trees and collected a Landsat 8 image on May 24, 2014. produces the location of each tree (x , y ), height (H ) and A Gram-Schmidt Spectral Sharpening image fusion tech- i i i crown radius (R ). nique in ENVI 5.1 (ITT Exelis) was applied to produce Fig 2 Component reflectance in 18 bands Huang and Lian Forest Ecosystems (2015) 2:20 Page 4 of 13 Fig. 3 Lidar derived CHM and DEM (1 km): a CHM, gray color representing tree height 0–30 m; b DEM, gray color representing elevation 770–895 m pan-sharpened Landsat 7 or 8 multi-spectral images with a at MODIS pixel scales (250 to 1000 m). The main input resolution of 15 m. This pan-sharpening method was se- parameters of RAPID consist of 3D structures of the lected because it preserves the original spectral information ground, trees, buildings and rivers, as well as reflectance of the image and can be simultaneously applied to multi- and transmittance of leaves, branches, walls, water bodies spectral bands. The Landsat image pixel values (in digital and roads under a few sun and sensor angles. The main numbers) were converted to top-of-atmosphere (TOA) outputs are BRF curves and land surface reflectance im- spectral radiance, which was further converted to land ages with defined spatial resolution (default 0.5 m). surface reflectance using the Fast Line-of-sight Atmos- pheric Analysis of Hypercubes (FLAASH) atmospheric Simulation framework correction model with the atmospheric visibility parameter Figure 4 shows the 3D simulation framework, with inte- estimated from the MODIS aerosol product. grated parameters extracted from lidar data, field plots MODIS 16-day 250 m NDVI images and the 500 m data, Landsat images and MODIS images managed into LAI products from Jan 1 to Dec 31, 2013 were down- the RAPID model to simulate optical images of a virtual loaded. Because a maximum value filtering method was sensor with several view angles, 18 spectral bands and a used, NDVI and LAI products had significantly fewer half-meter spatial resolution. cloud cover problems. NDVI data were used to deter- The sensor is an advanced version of the Compact High mine the phenology of the boreal forest, including birch, Resolution Imaging Spectrometer (CHRIS/PROBA). CHRIS larch and understory, which allow interpolation of Land- is the only multi-angular sensor launched with both high sat images ranging from clear to contaminated days. spatial (17 m) and spectral resolution (20–40 nm) (Rautiai- MODIS LAI products were utilized to determine the leaf nen et al. 2008; García Millán et al. 2014). For any selected area of each tree. Despite its low resolution, it is the only target, five images with different viewing angles (−55°, −36°, continuous global leaf area product, but with acceptable 0°, 36° and 55°) were made within a short span of 2.5 min. accuracy (Ahl et al. 2006). The virtual sensor was placed above the canopy under clear MODIS Bidirectional Reflectance Factor (BRF) products sky conditions. in May of 2013 were collected for validation. The BRF A large number of input parameters needed to be set curves were reconstructed from the kernel coefficients in order to simulate seasonal variation. A few parameters, using the Algorithm for Model Bidirectional Reflectance such as LAI and soil moisture, vary considerably over Anisotropies of the Land Surface (AMBRALS) (Wanner the growing season, while other parameters remain et al. 1995; Huang et al. 2013b; Sharma et al. 2013). relatively stable. Given our relatively limited data source, we defined five basic assumptions to reduce the number RAPID model of unknowns: RAPID is a 3D radiative transfer model, able to simulate reflectance images over complex 3D natural scenes at 1) The DTM (digital terrain model) remained large scales (30 to 1000 m) with great efficiency (Huang unchanged, a reasonable assumption for forested et al. 2013a), implying that RAPID can simulate images areas; Huang and Lian Forest Ecosystems (2015) 2:20 Page 5 of 13 Fig. 4 Simulation framework to generate time series of optical images 2) Tree crowns were ellipsoid or cone shaped, similar content) while fixing others, seasonal leaf reflectance with geometric optic models (Schaaf and Strahlerl and transmittance could be simulated (Barry et al. 2009). 1994; Chen et al. 2012); Previous studies have shown that the amount of leaf 3) Individual tree LAI (LAI ) was predicted from tree chlorophyll is correlated with NDVI (Wu et al. 2008; tree height (Xiao et al. 2006); Rulinda et al. 2011; Croft et al. 2013; Feng and Niu 2014). 4) We accepted a spherical leaf angle distribution Therefore, we used a linear relationship between MODIS (LAD ) for all trees due to missing measurements; NDVI products (0.1 to 1.0) and the amount of leaf chloro- tree −2 5) Reflectance of non-vegetation objects, such as walls, phyll (10 to 100 μg∙cm ). water bodies and roads remained constant, following precedents set in existing literature (Wang et al. 2008) Background reflectance or in the ENVI spectral library. Seasonal variation in background reflectance was com- plex. However, soil moisture played a major role (Muller With these assumptions, there were two types of input and Décamps 2001; Weidong et al. 2002; Whiting et al. parameters: fixed or dynamic. Fixed parameters were 2004), which was then derived from TVDI (temperature DEM, land cover map, individual tree map (coordinates, vegetation dryness index) inferred from temperature and DBH, height, crown radius, crown length). DEM and NDVI (Sandholt et al. 2002; Liang et al. 2014). TVDI is land cover map were re-sampled to a resolution of 1 m. highly correlated with soil moisture (Holzman et al. 2014). Land cover maps were generated using a decision tree Therefore, we estimated the soil reflectance as the weighted method with six classes: bare soil, road, birch forest, average of dry soil and wet soil reflectance, where the larch forest, water surface and buildings. Decision rules weights were TVDI and (1-TVDI) respectively. The back- were largely based on the Ratio Vegetation Index (RVI), ground was defined as soil covered by a homogeneous the Normalized Difference Water Index (NDWI) and shrub layer with a LAI of 0.5. Shrub leaves were assumed CHM. Dynamic parameters determined the seasonal to have the same optical parameters as birch leaves. change of reflectance, such as component reflectances, LAI and sun position, obtained mainly from time series Growing season of MODIS products, including NDVI, LAI and land sur- From the Landsat classification map, pure birch and larch face temperature (LST). pixels were selected to determine the beginning and final day (DOY) of the annual growing season, using the follow- Leaf reflectance ing phenology analysis. Leaf reflectance and transmittance were measured only First, MODIS time series NDVI data were fitted using once, which was not sufficient to represent optical features a harmonic analysis (Jonsson and Eklundh 2004) to re- for the entire growing season. Thanks to the PROSPECT move random noises. We referred to the maximum and model and changing the most sensitive input (leaf chlorophyll minimum values of NDVI as NDVI and NDVI . max min Huang and Lian Forest Ecosystems (2015) 2:20 Page 6 of 13 The starting date was defined as the DOY when NDVI equation. Both species showed very similar phenology in exceeded 20 % of (NDVI – NDVI ) between DOY the spring without a difference on the average (Delbart max min 1 and DOY 180. Similarly, the final date was the DOY et al. 2005), so we used the MODIS LAI to calibrate the when NDVI exceeded 20 % of (NDVI – NDVI ) g value for both species: max min between DOY 180 and DOY 365. gðÞ DOY f  H  πR ¼ LAI  Area i MODIS MODIS i i Temporal leaf area index of individual trees LAI  Area MODIS MODIS It was difficult to calculate LAI precisely. Instead, it ⇒gðÞ DOY¼ tree f  H  πR was possible to allocate the MODIS LAI into individual trees. Based on assumption (3), LAI is linearly related ð3Þ tree to tree height (H ) for each species. Therefore, LAI tree tree is a function of both species and DOY (see Equation 1). LAI ¼ fðÞ species gðÞ DOY H ð1Þ tree tree Evaluation method Since the TreeVaW (Popescu and Wynne 2004) had not where f is a coefficient relating H to LAI , which is tree tree been tested in our study site, we manually segmented a constant for each species and g is a temporal correction few tree crowns in nine sub-plots with different tree factor. Plot LAI and individual tree height in field plots densities in order to evaluate the accuracy of the ex- were used to calibrate f values for both birch or larch: tracted number of trees, height, location and crown f  H  πR ¼ LAI  Area radius. i plot plot i i LAI  Area We carried out four types of evaluations: (a) CCD plot plot ð2Þ ⇒f ¼ image was used to check the pattern of simulated half- H  πR meter images; (b) Landsat images were used to check For birch trees, we calibrated f as 0.25 and 0.20 for reflectance values of nadir images at the same date; (c) larch. From previous studies (Li et al. 2009; Liu and Jin MODIS BRF products were used to compare simulated 2013), we determined that the LAI of birch and larch BRFs and(d) finally, weusedfourdates of Landsatim- varied with DOY and could be fitted with a polynomial ages to evaluate temporal simulations. Fig. 5 Comparisons of tree segmentation between manual operation and TreeVaW in a sparse subplot (a-b) and a dense subplot (c-d); (a) and (c) are manual results; (b) and (d) are TreeVaW results Huang and Lian Forest Ecosystems (2015) 2:20 Page 7 of 13 LðÞ birch¼ 0:5475H−0:0118 R ¼ 0:87 ð6Þ LðÞ larch ¼ 0:6551H−0:0731 R ¼ 0:55 ð7Þ Growing season Figure 6 shows the smoothed NDVI curves for birch and larch-dominated forests. The starting date, final date and length of the growing season were estimated as DOY 140 (May 20), 273 (Sep 3) and 130 days. During the growing season, the birch forest had higher NDVI values than larch forests, and the larch forests in the flat wetland area had significant lower NDVI values than those in mountain areas. Fig. 6 Smoothed MODIS 16-day 250 m NDVI products in 2013 Land cover classification The Landsat 8 image on May 24, 2014 was used to pro- Results duce a 15 m classification map, given a suitable growing Tree structure season and good image quality to distinguish birch and Compared to manual segmentation, TreeVaW detected larch (Fig. 7a). Compared to the old forest map (Fig. 7b), 88 % of the number of trees in sparse plots (Fig. 5a, b), the southern regions (1 and 2) visually matched much but only 74 % in dense plots (Fig. 5c, d). Crown radii ob- better than the northern regions (3 and 4). Fortunately, the tained from TreeVaW ranged from 0.59 to 0.71 m, lower major study area was located in regions 1 and 2, where the than those from manual segmentation. The mean tree accuracy (around 75 %) was calculated from random sam- height error and location bias of detected trees was 0.88 pling points. Major rules of the decision tree were the fol- and 0.91 m. lowing: (1) forest vegetations = (RVI > 0 and NDWI > 0 Based on regression analysis of plot data, both tree and CHM > 2 m); (2) shrubs or grasses = (RVI > 0 and DBH and crown length (L) were well predicted from tree NDWI > 0 and CHM ≤ 2 m); (3) birch = ((1) and RVI > height (H) with coefficients of determination larger than 7.0); (4) larch = ((1) and RVI ≤ 5.0); (5) mixed forests = ((1) 0.80: and RVI > 5.0 and RVI ≤ 7.0). Forest understory in the Genhe Reserve was complex 1:5652 2 DBHðÞ birch¼ 0:2466H R ¼ 0:89 ð4Þ but of considerable value in identifying forest types (see Table 1). Some shrubs were evergreen, while grasses shed 1:8704 2 DBHðÞ larch ¼ 0:1639H R ¼ 0:82 ð5Þ leaves. Therefore, the vegetation detected in SPOT-6 AB Fig. 7 Vegetation map: (a) classified image with the lightest greenness representing birch forests; (b) forest map with white color representing birch forests Huang and Lian Forest Ecosystems (2015) 2:20 Page 8 of 13 Fig. 8 Determining vegetation as evergreen bush in winter season: (a) evergreen understory on Spot-6 image (red color); (b) CCD image; (c) CHM image image (1 m, October) was used to define shrubs as Comparisons of BRF evergreen vegetation because only evergreens had green Five pixels around the central study area showed variation leaves at that time of the year (Fig. 8). in the BRF curves, used as a reference to evaluate the RAPID BRF results (Fig. 11). Generally, the simulated BRF matches the shape of MODIS BRF in spite of absolute Comparisons of nadir images biases in a few view directions. First, the simulated red Simulated nadir images (0.5 m resolution) were compared BRF is higher than all MODIS BRFs when the view zenith to the CCD image in Fig. 9. The spatial texture and land angle (VZA) is between −50° and 40°. Second, in both red cover difference are consistent, but the simulated forests and NIR bands, the backscattering BRF when VZA larger look sparser. than 50°, is lower than the MODIS BRF. The spectral results were compared with Landsat 8 re- flectance images on May 24, 2014 (Fig. 10). Both simulation and Landsat images showed typical vegetation reflectance Temporal results spectra (low red reflectance and high near infrared (NIR) Four Landsat images were used to check the simulation reflectance). Simulation results are significantly lower in ability of temporal variations; the dynamic parameters of blue bands (0.02 to 0.06). birch and larch trees are shown in Table 2. Fig. 9 Comparing nadir image (0.5 m) with CCD: (a) simulated image (R = Near infrared (NIR), G = red, B = green); (b) airborne CCD mosaic image from multiple days Huang and Lian Forest Ecosystems (2015) 2:20 Page 9 of 13 Fig. 10 Comparison of nadir top of canopy (TOC) reflectance image with a Landsat 8 image using linear stretch (0 to 0.3): (a) simulated image (0.5 m, R = NIR, G = red, B = green); (b) re-sampled 15 m image from (a); (c) Landsat 8 (15 m) on May 24, 2014 (R = NIR, G = red, B = green); (d) Spectral curves of dense and sparse canopies Figure 12 compares the results between simulated Discussion and real images in stripes of birch and larch forest Our main objective was to create and test how to couple stands (600 m by 600 m). The resolutions were 15 m lidar data and temporal optical data MODIS in order to except forthe LandsatTMimage (30m)onSeptember simulate high-resolution optical satellite images. A frame- 5, 2013. The birch bands (marked as A) showed signifi- work was built and tested at the Genhe Forest Farm. In cant variation in reflectance from brown color (bare spite of some biases or errors, the approach successfully soil), red color (green canopy), pink color (dense can- produced temporal images with high spatial, spectral and opy) to mixed color (discoloring canopy), reconstructed angular resolutions, which confirmed the possibility to from simulated images in spite of slightly different fuse lidar and MODIS data. colors. In the lower part of the Landsat ETM+ image, a black no-data area showed up, due to a sensor error Major contributors on simulation (SLC-OFF). The results on Sept 5, 2013 showed larger The framework included four main data sources: lidar, discrepancies. Landsat, MODIS and field data. To drive a 3D model, Huang and Lian Forest Ecosystems (2015) 2:20 Page 10 of 13 Fig. 11 Comparisons between MODIS BRF product and RAPID simulations: (a) red band (0.620–0.670 nm); (b) NIR band (0.841–0.876 nm) the most important inputs were 3D scenes and the inside Major errors reflectance and transmittance of 3D objects. Lidar was the Despite the fact that three types of evaluation on re- first contributor to providing 3D structures of individual flectance, i.e., spatial texture, BRF and Landsat simula- trees and background. Lidar-derived 3D structures were tion demonstrated the capability to simulate temporal normally static, but 3D scenic objects, especially their LAI, images, quantitative validation was still missing due to were dynamic. Therefore, we used an allocation method lots of uncertainties in the entire workflow. We tried to to downscale MODIS LAI into each tree, a technique not address major error sources and assess their uncertainty: found in previous studies. Landsat images were used to classify birch and larch, supporting the generation of 1) 3D structure errors: 3D trees. The optical parameters of 3D objects were collected in It has to be admitted that suppressed trees and irregular the field or obtained from existing references; these were tree crowns are hard to detect from CHM. A previous also dynamic. Therefore, MODIS NDVI data were used to study has shown that TreeVaW method can identify more calibrate leaf chlorophyll for the PROSPECT model, which than 95 % of the trees in planted forests but only 70 % in then simulated dynamic leaf reflectance and transmittance. natural forests (Antonarakis et al. 2008). Although other Background soil reflectance varied over time and was diffi- detection algorithms may help improve the accuracy, the cult to obtain. An alternative is to use TVDI to adjust soil inter-comparisons between detection methods found that reflectance, which is a more recent idea and needs to be the correct percentage of the number of trees was gener- evaluated in any future research. ally between 50 to 90 % (Kaartinen et al. 2012). In our Table 2 Dynamic parameters of birch or larch forests −2 2 −2 2 −2 Date DOY Chlorophyll content (μg∙cm ) LAI (m ∙m ) LAI (m ∙m ) TVDI larch birch May 5, 2013 125 41 0.16 × H 0.13 × H 0.61 tree tree May 24, 2014 144 52 0.17 × H 0.18 × H 0.68 tree tree June 27, 2012 179 85 0.20 × H 0.23 × H 0.91 tree tree Sept 5, 2013 248 72 0.19 × H 0.11 × H 0.62 tree tree Huang and Lian Forest Ecosystems (2015) 2:20 Page 11 of 13 Fig. 12 Comparison between simulated and Landsat images with false color composition (RGB = [NIR, RED, GREEN]); A and B represent birch and larch trees, respectively study, the percentage of the correct number was between season will be low. Continuous field observations are 74 to 90 %, which is consistent with results above. The strongly suggested. high level of missed detection leads to a higher clumping effect and sparser forests (Figs. 7 and 8), which then re- 4) MODIS data uncertainty sulted in higher reflectance biases induced by background uncertainties. The most recent MODIS LAI product is Collection 5 (this a version code), which has uncertainties around +/−1.0 2) Unknown background: for relatively pure pixels (Fang et al. 2012). However, considering the low resolution of MODIS pixels, the Although we classified evergreen bush, its background uncertainties of inversed LAIs are even larger for mixed type and dynamic reflectance were almost unknown. pixels. The image matching between MODIS (1 km) Therefore, a very rude LAI of 0.5 was assumed for all and CHM (0.5 m) sounds tricky. However, as the only understories. In fact, it is possible to retrieve forest available product, it was used in our simulation frame- background reflectance from satellite data (Canisius and work. In any future study, we will use Landsat images Chen 2007; Pisek and Chen 2009; Pisek et al. 2010; to bridge the gap of higher resolution of LAI products Tuanmu et al. 2010; Rautiainen et al. 2011; Pisek et al. (Gao et al. 2014). The BRF biases between simulation 2012). We will try these methods to inverse background and MODIS can be partially attributed to the limitations reflectance in later studies. We were able to validate the of MODIS BRF in reconstructing higher and narrower TVDI-adjusted soil reflectance, although it should have hotspots (Huang et al. 2013b). directional effects. Actually, we used isotropic soil reflect- ance, which may explain the BRF biases with large angles 5) Landsat data uncertainty in backward view directions. Landsat images were used to compare simulated nadir 3) Leaf discoloring reflectance and image textures. Figure 10 shows sig- nificant differences in the blue band, which can be In September, the leaves of both birch and larch chan- largely explained by an atmospheric correction error ged color. However, these changes varied even for trees because blue band reflectance should be lower after a of the same species, probably an effect of age, elevation correct removal of aerosol scatter. This atmospheric or density, making it difficult to identify individual trees. correction was carried out by using the FLAASH mod- Therefore, the accuracy in discoloration during the growing ule of the ENVI 5.1 software, where the aerosol optical Huang and Lian Forest Ecosystems (2015) 2:20 Page 12 of 13 depth and water content were only estimated from scale optical image dataset will be useful to support the images. understanding of scale problems. Conclusions Efficiency problems We presented a simulation framework which links lidar RAPID is relatively fast with 3D models, but running with optical images to produce series of temporal im- one case (1 km) still needs four to six hours at individual ages. The study provides a proof-of-concept approach to tree scale on a workstation (using 10 CPU cores). We link lidar data in the parameterization of a RAPID model are of the opinion that it is not feasible for the gener- for temporal image reconstruction in forest dominated ation of operational products. However, for some scien- areas. Demonstrations were applied at the Genhe Forest tific use, focused on local areas, it may be worthwhile to Farm, a remote forest reserve in China. Evaluations on obtain images with very high resolutions (spatial, spec- nadir reflectance, spatial textures and BRF confirmed tral and angular) for research within an acceptable time that 3D simulation provides an insight look into how im- frame and cost structure. Because RAPID can run at ages vary over time. Many uncertainties were identified, scalable resolutions, 3D scenes of dense forests can be which can be expected to be reduced in any future study. up-scaled to regular grids with a medium resolution (e.g. Strategies to improve efficiency are possible and discussed. 5 to 10 m), which significantly improves calculation effi- ciency (less than 30 min) without much loss in accuracy. Competing interests Furthermore, we can create a reference table of 3D scenes, The authors declare that they have no competing interests. classifying a study area into fewer categories with possible combinations of DEM, understory, tree locations, tree Authors’ contributions HH is the major and contact author for most work. JL processed a small part heights and tree LAI. The corresponding reflectance of the data. Both authors read and approved the final manuscript. images will be simulated and stored as an image database. Once the database is created, a quick search method can be Acknowledgements used to pick up desired images based on input parameters The authors gratefully acknowledge the Chinese National Basic Research Program (2013CB733401) and the Chinese Natural Science Foundation such as DEM, understory, tree distribution and LAI. In the Project (41171278). current framework, we only dealt with the capability of coupling simulation. Improvements will be presented Received: 28 February 2015 Accepted: 18 June 2015 in our next study. References Adams T, Beets P, Parrish C (2012) Extracting more data from LiDAR in forested Scale issues areas by analyzing waveform shape. Remote Sens 4(3):682–702. Scale effect and scaling have been big issues in remote doi:10.3390/rs4030682 sensing community. When models or algorithms at small Ahl DE, Gower ST, Burrows SN, Shabanov NV, Myneni RB, Knyazikhin Y (2006) Monitoring spring canopy phenology of a deciduous broadleaf forest using scales are used at large scales, they may produce certain MODIS. Remote Sens Environ 104(1):88–95. doi:10.1016/j.rse.2006.05.003 errors, especially for non-linear models (Tao et al. 2009). Antonarakis AS, Richards KS, Brasington J, Bithell M, Muller E (2008) Retrieval of Scale issues constrain the accuracy of retrieval and limit vegetative fluid resistance terms for rigid stems using airborne lidar. J Geophys Res 113(G2):G02S07. doi:10.1029/2007JG000543 the development of remote sensing applications. In this Arno J, Escola A, Valles JM, Llorens J, Sanz R, Masip J, Palacin J, Rosell-Polo JR study, MODIS LAI products at a 1 km scale were down- (2013) Leaf area index estimation in vineyards using a ground-based LiDAR scaled to the level of LAI of individual trees, using lidar at scanner. Precision Agric 14(3):290–306. doi:10.1007/s11119-012-9295-0 Barbier N, Proisy C, Vega C, Sabatier D, Couteron P (2011) Bidirectional texture a scale of only a few meters, used to coincide with the function of high resolution optical images of tropical forest: An approach RAPID model scale. Assuming that MODIS LAI products using LiDAR hillshade simulations. Remote Sens Environ 115(1):167–179. are scale-corrected, this scaling does not change total leaf doi:10.1016/j.rse.2010.08.015 Barry KM, Newnham GJ, Stone C (2009) Estimation of chlorophyll content in area. Since LAI was found to be correlated with tree tree Eucalyptus globulus foliage with the leaf reflectance model PROSPECT. Agr Forest height, the down-scaling model should also be linear with- Meteorol 149(6–7):1209–1213, http://dx.doi.org/10.1016/j.agrformet.2009.01.005 out model scale issues. In comparison with satellite images, Canisius F, Chen JM (2007) Retrieving forest background reflectance in a boreal region from Multi-anglo Imaging SpectroRadiometer (MISR) data. Remote the simulated RAPID images should also be re-sampled to Sens Environ 107(1–2):312–321. doi:10.1016/j.rse.2006.07.023 coincide with the satellite scale, including the MODIS and Chen G, Wulder MA, White JC, Hilker T, Coops NC (2012) Lidar calibration and Landsat scales. Fortunately, reflectance or radiance images validation for geometric-optical modeling with Landsat imagery. Remote Sens Environ 124(0):384–393, http://dx.doi.org/10.1016/j.rse.2012.05.026 are scale independent. Croft H, Chen JM, Zhang Y, Simic A (2013) Modelling leaf chlorophyll content in Using MODIS and lidar data to reconstruct 3D scenes broadleaf and needle leaf canopies from ground, CASI, Landsat TM 5 at an individual tree scale, the RAPID model is capable and MERIS reflectance data. Remote Sens Environ 133(0):128–140, http://dx.doi.org/10.1016/j.rse.2013.02.006 of simulating time series of images at spatial scales from Delbart N, Kergoat L, Le Toan T, Lhermitte J, Picard G (2005) Determination of 0.5 m to 1 km and temporal scales of a few days. Although phenological dates in boreal regions using normalized difference water the 3D scene estimation is not perfect, this type of multi- index. Remote Sens Environ 97(1):26–38. doi:10.1016/j.rse.2005.03.011 Huang and Lian Forest Ecosystems (2015) 2:20 Page 13 of 13 Fang H, Wei S, Liang S (2012) Validation of MODIS and CYCLOPES LAI products Muller E, Décamps H (2001) Modeling soil moisture–reflectance. Remote Sens using global field measurement data. Remote Sens Environ 119:43–54. Environ 76(2):173–180, http://dx.doi.org/10.1016/S0034-4257(00)00198-X doi:10.1016/j.rse.2011.12.006 Nitze I, Barrett B, Cawkwell F (2015) Temporal optimisation of image acquisition Feng M, Niu Z (2014) Chlorophyll content retrieve of vegetation using Hyperion for land cover classification with Random Forest and MODIS time-series. Int J data based on empirical models. Remote Sens Land Resour 26(1):71–77. Appl Earth Observ Geoinformation 34:136–146. doi:10.1016/j.jag.2014.08.001 doi:10.6046/gtzyyg.2014.01.13 Pisek J, Chen JM (2009) Mapping forest background reflectivity over North Gao F, Anderson MC, Kustas WP, Houborg R (2014) Retrieving leaf area index America with Multi-angle Imaging SpectroRadiometer (MISR) data. Remote from landsat using MODIS LAI products and field measurements. IEEE Geosci Sens Environ 113(11):2412–2423. doi:10.1016/j.rse.2009.07.003 Remote Sens Lett 11(4):773–777. doi:10.1109/lgrs.2013.2278782 Pisek J, Chen JM, Miller JR, Freemantle JR, Peltoniemi JI, Simic A (2010) Mapping Gao F, Masek J, Schwaller M, Hall F (2006) On the blending of the Landsat and forest background reflectance in a boreal region using multiangle compact MODIS surface reflectance: predicting daily Landsat surface reflectance. IEEE airborne spectrographic imager data. IEEE Trans Geosci Remote Sens Trans Geosci Remote Sens 44(8):2207–2218. doi:10.1109/tgrs.2006.872081 48(1):499–510. doi:10.1109/tgrs.2009.2024756 Pisek J, Rautiainen M, Heiskanen J, Mottus M (2012) Retrieval of seasonal García Millán VE, Sánchez-Azofeifa A, Málvarez García GC, Rivard B (2014) Quantifying dynamics of forest understory reflectance in a Northern European boreal tropical dry forest succession in the Americas using CHRIS/PROBA. Remote Sens forest from MOD'S BRDF data. Remote Sens Environ 117:464–468. Environ 144(0):120–136, http://dx.doi.org/10.1016/j.rse.2014.01.010 doi:10.1016/j.rse.2011.09.012 Hilker T, Wulder MA, Coops NC, Seitz N, White JC, Gao F, Masek JG, Stenhouse G Popescu SC, Wynne RH (2004) Seeing the trees in the forest: using lidar and (2009) Generation of dense time series synthetic Landsat data through data multispectral data fusion with local filtering and variable window size for blending with MODIS using a spatial and temporal adaptive reflectance fusion estimating tree height. Photogrammetric Eng Remote Sens 70(5):589–604 model. Remote Sens Environ 113(9):1988–1999. doi:10.1016/j.rse.2009.05.011 Rautiainen M, Lang M, Mõttus M, Kuusk A, Nilson T, Kuusk J, Lükk T (2008) Holzman ME, Rivas R, Bayala M (2014) Subsurface soil moisture estimation by Multi-angular reflectance properties of a hemiboreal forest: An analysis VI-LST method. IEEE Geosci Remote Sens Lett 11(11):1951–1955. using CHRIS PROBA data. Remote Sens Environ 112(5):2627–2642, doi:10.1109/lgrs.2014.2314617 http://dx.doi.org/10.1016/j.rse.2007.12.005 Huang H, Qin W, Liu Q (2013a) RAPID: a radiosity applicable to porous individual Rautiainen M, Mottus M, Heiskanen J, Akujarvi A, Majasalmi T, Stenberg P (2011) objects for directional reflectance over complex vegetated scenes. Remote Seasonal reflectance dynamics of common understory types in a northern Sens Environ 132:221–237. doi:10.1016/j.rse.2013.01.013 European boreal forest. Remote Sens Environ 115(12):3020–3028. Huang X, Jiao Z, Dong Y, Zhang H, Li X (2013b) Analysis of BRDF and albedo retrieved doi:10.1016/j.rse.2011.06.005 by kernel-driven models using field measurements. IEEE J Selected Topics Appl Rulinda CM, Bijker W, Stein A (2011) The chlorophyll variability in Meteosat Earth Observ Remote Sens 6(1):149–161. doi:10.1109/jstars.2012.2208264 derived NDVI in a context of drought monitoring. Procedia Environ Sci Inglada J, Hagolle O, Dedieu G (2011) A framework for the simulation of high 3(0):32–37, http://dx.doi.org/10.1016/j.proenv.2011.02.007 temporal resolution image series. Paper presented at the IEEE International Sandholt I, Rasmussen K, Andersen J (2002) A simple interpretation of the surface Geoscience and Remote Sensing Symposium (IGARSS), 24–29 July 2011 temperature/vegetation index space for assessment of surface moisture Jacquemoud S, Verhoef W, Baret F, Bacour C, Zarco-Tejada PJ, Asner GP, François status. Remote Sens Environ 79(2–3):213–224, http://dx.doi.org/10.1016/ C, Ustin SL (2009) PROSPECT + SAIL models: a review of use for vegetation S0034-4257(01)00274-7 characterization. Remote Sens Environ 113(Supplement 1 (0)):S56–S66, Schaaf CB, Strahlerl AH (1994) Validation of bidirectional and hemispherical http://dx.doi.org/10.1016/j.rse.2008.01.026 reflectances from a geometric-optical model using ASAS imagery and Jonsson P, Eklundh L (2004) TIMESAT - a program for analyzing time-series of satellite pyranometer measurements of a spruce forest. Remote Sens Environ sensor data. Comput Geosci 30(8):833–845. doi:10.1016/j.cageo.2004.05.006 49(2):138–144, http://dx.doi.org/10.1016/0034-4257(94)90050-7 Kaartinen H, Hyyppä J, Yu X, Vastaranta M, Hyyppä H, Kukko A, Holopainen M, Sharma RC, Kajiwara K, Honda Y (2013) Estimation of forest canopy structural Heipke C, Hirschmugl M, Morsdorf F, Næsset E, Pitkänen J, Popescu S, Solberg S, parameters using kernel-driven bi-directional reflectance model based Wolf BM, Wu J-C (2012) An international comparison of individual tree detection multi-angular vegetation indices. ISPRS J Photogrammetry Remote Sens and extraction using airborne laser scanning. Remote Sensing 4(4):950–974 78:50–57. doi:10.1016/j.isprsjprs.2012.12.006 Kobayashi H, Suzuki R, Kobayashi S (2007) Reflectance seasonality and its relation Tao X, Yan B, Wang K, Wu D, Fan W, Xu X, Liang S (2009) Scale transformation to the canopy leaf area index in an eastern Siberian larch forest: Multi-satellite of Leaf Area Index product retrieved from multiresolution remotely sensed data and radiative transfer analyses. Remote Sens Environ 106(2):238–252. data: analysis and case studies. Int J Remote Sens 30(20):5383–5395. doi:10.1016/j.rse.2006.08.011 doi:10.1080/01431160903130978 Li GZ, Wang HX, Zhu JJ (2009) Monthly changes of leaf area index and canopy Tuanmu M-N, Vina A, Bearer S, Xu W, Ouyang Z, Zhang H, Liu J (2010) Mapping openness of Larix olgensis in mountainous regions in east Liaoning province. understory vegetation using phenological characteristics derived from remotely J Northeast Forestry Univ 37(7):13 sensed data. Remote Sens Environ 114(8):1833–1844. doi:10.1016/j.rse.2010.03.008 Liang L, Zhao S-h, Qin Z-h, He K-x, Chen C, Luo Y-x, Zhou X-d (2014) Drought Wang JD, Zhang LX, Liu QH, Zhang B, Yin Q (2008) The spectrum knowledge change trend using MODIS TVDI and its relationship with climate factors database of typical land surface objects in China. Science Press, Beijing in china from 2001 to 2010. J Integr Agric 13(7):1501–1508, Wanner W, Li X, Strahler AH (1995) On the derivation of kernels for kernel-driven http://dx.doi.org/10.1016/S2095-3119(14)60813-3 models of bidirectional reflectance. J Geophys Res 100(D10):21077–21089. Liu ZL, Jin GZ (2013) Estimation of leaf area index of secondary Betula platyphylla doi:10.1029/95JD02371 forest in Xiaoxing' an Mountains. Acta Ecologica Sinica 33(8):9 Weidong L, Baret F, Xingfa G, Qingxi T, Lanfen Z, Bing Z (2002) Relating soil Lucas RM, Lee AC, Williams ML (2006) Enhanced simulation of radar backscatter surface moisture to reflectance. Remote Sens Environ 81(2–3):238–246, from forests using LiDAR and optical data. IEEE Trans Geosci Remote Sens http://dx.doi.org/10.1016/S0034-4257(01)00347-9 44(10):2736–2754. doi:10.1109/tgrs.2006.881802 Whiting ML, Li L, Ustin SL (2004) Predicting water content using Gaussian model on Lunetta RS, Johnson DM, Lyon JG, Crotwell J (2004) Impacts of imagery temporal soil spectra. Remote Sens Environ 89(4):535–552, http://dx.doi.org/10.1016/ frequency on land-cover change detection monitoring. Remote Sens Environ j.rse.2003.11.009 89(4):444–454. doi:10.1016/j.rse.2003.10.022 Wu C, Niu Z, Tang Q, Huang W (2008) Estimating chlorophyll content from Masek JG, Huang C, Wolfe R, Cohen W, Hall F, Kutler J, Nelson P (2008) North hyperspectral vegetation indices: Modeling and validation. Agr Forest Meteorol American forest disturbance mapped from a decadal Landsat record. Remote 148(8–9):1230–1241, http://dx.doi.org/10.1016/j.agrformet.2008.03.005 Sens Environ 112(6):2914–2926. doi:10.1016/j.rse.2008.02.010 Wu M-Q, Wang J, Niu Z, Zhao Y-Q, Wang C-Y (2012) A model for spatial and temporal Montesano P, Cook B, Sun G, Simard M, Nelson R, Ranson K, Zhang Z, Luthcke S data fusion. J Infrared Millimeter Waves 31(1):80–84 (2013) Achieving accuracy requirements for forest biomass mapping: a Xiao C-W, Janssens IA, Curiel Yuste J, Ceulemans R (2006) Variation of specific leaf spaceborne data fusion method for estimating forest biomass and LiDAR area and upscaling to leaf area index in mature Scots pine. Trees 20(3):304–310. sampling error. Remote Sens Environ 130:153–170. doi:10.1016/ doi:10.1007/s00468-005-0039-x j.rse.2012.11.016 Xu G-c, Pang Y, Li Z-y, Zhao K-r, Liu L-x (2013) The changes of forest canopy Mu X, Zhang Q, Liu Q, Pang Y, Hu K (2015) A study on typical forest biomass spectral reflectance with seasons in Xiaoxing'anling. Spectrosc Spectral Anal mapping technology of great khingan using airborne laser scanner data. 33(12):3303–3307 Remote Sens Technol Appl 30(2):220–225 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png "Forest Ecosystems" Springer Journals

A 3D approach to reconstruct continuous optical images using lidar and MODIS

Huang, HuaGuo; Lian, Jun
"Forest Ecosystems" , Volume 2 (1): 13 – Dec 1, 2015
Download PDF
13 pages

Loading next page...
 
/lp/springer-journals/a-3d-approach-to-reconstruct-continuous-optical-images-using-lidar-and-k2nVY92j0i

References (54)

Publisher
Springer Journals
Copyright
2015 Huang and Lian.
eISSN
2197-5620
DOI
10.1186/s40663-015-0044-5
Publisher site
See Article on Publisher Site

Abstract

Background: Monitoring forest health and biomass for changes over time in the global environment requires the provision of continuous satellite images. However, optical images of land surfaces are generally contaminated when clouds are present or rain occurs. Methods: To estimate the actual reflectance of land surfaces masked by clouds and potential rain, 3D simulations by the RAPID radiative transfer model were proposed and conducted on a forest farm dominated by birch and larch in Genhe City, DaXing’AnLing Mountain in Inner Mongolia, China. The canopy height model (CHM) from lidar data were used to extract individual tree structures (location, height, crown width). Field measurements related tree height to diameter of breast height (DBH), lowest branch height and leaf area index (LAI). Series of Landsat images were used to classify tree species and land cover. MODIS LAI products were used to estimate the LAI of individual trees. Combining all these input variables to drive RAPID, high-resolution optical remote sensing images were simulated and validated with available satellite images. Results: Evaluations on spatial texture, spectral values and directional reflectance were conducted to show comparable results. Conclusions: The study provides a proof-of-concept approach to link lidar and MODIS data in the parameterization of RAPID models for high temporal and spatial resolutions of image reconstruction in forest dominated areas. Keywords: Lidar; Optical; Temporal interpolation; 3D model; High resolution Background A few studies have been conducted to interpolate those Optical remote sensing images have been widely used contaminated images by rains or clouds. For example, in monitoring forest ecosystems. Spatial, temporal and Landsat images have been blended with MODerate- spectral resolutions are the three key indicators to be resolution Imaging Spectroradiometer (MODIS) data to considered during most applications. Spatial resolution create spatial and temporal fusion data (Gao et al. 2006; hadbeenimprovedfroma scaleof hundredsofmeters Hilker et al. 2009; Wu et al. 2012). Further, radiative (e.g. Landsat 8) to one of a half-meter (e.g. GeoEye-1 or transfer models have also been used to simulate a series Worldview-2) with only a slight increase in the number of high temporal resolution images for future space of spectral bands. However, in forested area, temporal earth observation missions (Inglada et al. 2011). How- resolution is generally reduced by frequent rains or ever, spatial resolution is generally moderate due to the cloud covers, which prevents users from continuously useofsimplehomogeneous radiativetransfermodels, acquiring clear optical remote sensing images. which are not able to deal with high resolution simula- Temporally continuous satellite images are important tion with diverse tree species and mountain shadows. for forest monitoring (Lunetta et al. 2004; Masek et al. In recent years, light detection and ranging (lidar) has 2008; Nitze et al. 2015) since forest reflectance varies been a widely used tool for forest studies (Adams et al. with seasonality (Kobayashi et al. 2007; Xu et al. 2013). 2012; Arno et al. 2013; Montesano et al. 2013). The greatest advantage of lidar is to provide direct measure- ments of very detailed 3D forest structures, so it can be * Correspondence: huaguo.huang@gmail.com used to reconstruct 3D trees to support the simulation Key Laboratory for Silviculture and Conservation of Ministry of Education, Beijing Forestry University, Beijing, China © 2015 Huang and Lian. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited. Huang and Lian Forest Ecosystems (2015) 2:20 Page 2 of 13 of radar remote sensing signals (Lucas et al. 2006) and The growing season typically begins in early May and to study how 3D structures affect the quality of optical senescence occurs in late September. In the summer of images (Barbier et al. 2011). By coupling lidar with high 2013, 18 field plots (45 m by 45 m) were established temporal data such as MODIS using 3D radiative trans- representing different combinations of forest types, density fer models, it will be possible to generate both high and leaf area index (LAI) (Table 1). The LAI, ranging from 2 −2 spatial and temporal resolution optical remote sensing 1.44 to 3.51 m ∙m , was measured using LAI-2000 images. However, very few studies were found using (LICOR Inc.) hemispheric data. The forest cover varies be- this approach. Therefore, we will test the possibility of tween 0.21 and 0.86. that approach to simulate high-resolution optical satellite Based on inventory data of individual tree structures images on an arbitrary day of the growing season in a for- in plots L1 to L9, the DBH and crown length (L) of trees ested area. were regressed on tree height (H), where heights were derived by lidar. For convenience, both crowns of larch Methods and birch are defined as spherical in shape. Study site Reflectance or transmittance spectra of leaves, branches The study site (Fig. 1), located in a 100 ha forested area and stems of birch and larch trees were measured in (50°54′ N, 121°54′ E) of the Genhe Forestry Reserve, the field using the integrating sphere of ASD (Analytical DaXing’AnLing Mountain in Inner Mongolia, China, be- Spectral Device, http://www.asdi.com/). Dry and wet soil longs to a boreal moist and cold temperature forest, with spectra were the default soil spectra in a PROSAIL model an elevation ranging from 784 to 1142 m. Annual average (Jacquemoud et al. 2009). The re-sampled spectral curves precipitation is 450 to 550 mm, with sixty per cent falling are shown in Fig. 2. in July and August. Annual average sunshine is 2594 h with a frost-free period of 80 days. Our study site occupied Airborne data 75 % of the total area. The forest is mainly composed of Small footprint full-waveform lidar data were acquired Dahurian Larch (Larix gmelinii) and White Birch (Betula from August 16 to September 25 in 2012 (Mu et al. 2015). platyphylla Suk.). The understory vegetation of the larch The system consisted of a Leica ALS60 with an integrated forest is a single layer of evergreen shrubs (normally Leica RCD105 camera. The CCD camera produced nat- Ledum palustre L. or Rhododendron dauricum L.). L. ural color mosaic images with 0.2 m resolution. As for palustre is generally a low shrub (less than 0.3 m), while lidar, the mean swath width was 1 km at a flying altitude the height of R. dauricum is around 1.5 m. Blueberries of approximately 2700 m (over rough terrain). The scan (Semen trigonellae) are widely distributed. The birch forest angle was less than 35°. Waveforms were digitized with a has a understory of grass or deciduous shrubs, such as frequency of 100 to 200 kHz. An average of eight reflected Rosa acicularis, Spiraea sericea Turcz., or Rubus L. pulses per m was obtained over the sample plots. Point Fig. 1 Location of the Genhe study site in DaXing’AnLing Mountain, Inner Mongolia, China Huang and Lian Forest Ecosystems (2015) 2:20 Page 3 of 13 Table 1 Plot variables at the study site 2 −2 Stand ID Tree stems (Birch: Larch) Slope (°) Shrub height (m) and cover Mean tree height (m) Mean DBH (cm) LAI (m ∙m ) A1 83:140 35 1.5, 70 % 11.6 11.8 2.68 ± 0.13 A2 97: 248 20 1.8, 60 % 12.1 11.6 2.88 ± 0.24 A3 41:269 25 1.5, 75 % 11.0 11.4 2.56 ± 0.09 A4 174:22 5 0.7, 3 % 16.0 16.1 3.31 ± 0.07 A5 142:103 10 2.0, 80 % 11.8 11.8 No data A6 91:173 45 0.3, 80 % 11.2 11.5 No data A7 111:102 30 1.5, 50 % 14.2 16.2 2.06 ± 0.13 A8 156:238 10 No data 11.0 9.5 2.51 ± 0.20 A9 65:180 45 1.3, 70 % 14.5 15.4 2.36 ± 0.17 L1 131:220 5 1.2, 20 % 8.4 10.0 2.71 ± 0.25 L2 89:421 <5 1.2, 5 % 9.0 8.7 3.06 ± 0.22 L3 11:71 5 0.7, 3 % 8.0 7.8 2.42 ± 0.50 L4 294:79 5 1.2, 15 % 11.7 11.6 3.51 ± 0.20 L5 1:327 7 1.0, 5 % 12.5 12.9 2.43 ± 0.12 L6 0:173 5 0.5, 20 % 17.2 22.3 1.44 ± 0.06 L7 11:90 7 1.0, 80 % 9.7 12.5 1.50 ± 0.15 L8 60:118 12 1.3, 30 % 13.1 15.4 2.70 ± 0.12 L9 0:585 5 0.5, 50 % 8.4 8.7 2.28 ± 0.09 clouds were first classified by the TerraScan software (see Optical satellite images www.terrasolid.com) to separate the ground points from Several scales of geo-referenced satellite images were used, other points. We used Delaunay triangulation and bilinear including SPOT-6 (1 m), Landsat (30 m) and MODIS interpolation method to generate a digital elevation model (250 m). Due to frequent cloud cover, SPOT and Landsat (DEM) from ground returns. DSM (digital surface model) were not able to capture clear land surface images during was created using a maximum value in a window size of rainy days, which happened mostly in the growing season 0.5 m. CHM (canopy height model) was calculated as the (May to September). There was only one cloud-free difference between DSM and DEM (Fig. 3). SPOT-6 image, obtained on October 10, 2013. For the Individual tree crowns were segmented from CHM same reason, only three Landsat images were clear (May using the “TreeVaW” tool (Popescu and Wynne 2004), 5, 2013; Sep 9, 2013; Sep 29, 2013) in 2013. As well, we which uses a circular window filter to segment trees and collected a Landsat 8 image on May 24, 2014. produces the location of each tree (x , y ), height (H ) and A Gram-Schmidt Spectral Sharpening image fusion tech- i i i crown radius (R ). nique in ENVI 5.1 (ITT Exelis) was applied to produce Fig 2 Component reflectance in 18 bands Huang and Lian Forest Ecosystems (2015) 2:20 Page 4 of 13 Fig. 3 Lidar derived CHM and DEM (1 km): a CHM, gray color representing tree height 0–30 m; b DEM, gray color representing elevation 770–895 m pan-sharpened Landsat 7 or 8 multi-spectral images with a at MODIS pixel scales (250 to 1000 m). The main input resolution of 15 m. This pan-sharpening method was se- parameters of RAPID consist of 3D structures of the lected because it preserves the original spectral information ground, trees, buildings and rivers, as well as reflectance of the image and can be simultaneously applied to multi- and transmittance of leaves, branches, walls, water bodies spectral bands. The Landsat image pixel values (in digital and roads under a few sun and sensor angles. The main numbers) were converted to top-of-atmosphere (TOA) outputs are BRF curves and land surface reflectance im- spectral radiance, which was further converted to land ages with defined spatial resolution (default 0.5 m). surface reflectance using the Fast Line-of-sight Atmos- pheric Analysis of Hypercubes (FLAASH) atmospheric Simulation framework correction model with the atmospheric visibility parameter Figure 4 shows the 3D simulation framework, with inte- estimated from the MODIS aerosol product. grated parameters extracted from lidar data, field plots MODIS 16-day 250 m NDVI images and the 500 m data, Landsat images and MODIS images managed into LAI products from Jan 1 to Dec 31, 2013 were down- the RAPID model to simulate optical images of a virtual loaded. Because a maximum value filtering method was sensor with several view angles, 18 spectral bands and a used, NDVI and LAI products had significantly fewer half-meter spatial resolution. cloud cover problems. NDVI data were used to deter- The sensor is an advanced version of the Compact High mine the phenology of the boreal forest, including birch, Resolution Imaging Spectrometer (CHRIS/PROBA). CHRIS larch and understory, which allow interpolation of Land- is the only multi-angular sensor launched with both high sat images ranging from clear to contaminated days. spatial (17 m) and spectral resolution (20–40 nm) (Rautiai- MODIS LAI products were utilized to determine the leaf nen et al. 2008; García Millán et al. 2014). For any selected area of each tree. Despite its low resolution, it is the only target, five images with different viewing angles (−55°, −36°, continuous global leaf area product, but with acceptable 0°, 36° and 55°) were made within a short span of 2.5 min. accuracy (Ahl et al. 2006). The virtual sensor was placed above the canopy under clear MODIS Bidirectional Reflectance Factor (BRF) products sky conditions. in May of 2013 were collected for validation. The BRF A large number of input parameters needed to be set curves were reconstructed from the kernel coefficients in order to simulate seasonal variation. A few parameters, using the Algorithm for Model Bidirectional Reflectance such as LAI and soil moisture, vary considerably over Anisotropies of the Land Surface (AMBRALS) (Wanner the growing season, while other parameters remain et al. 1995; Huang et al. 2013b; Sharma et al. 2013). relatively stable. Given our relatively limited data source, we defined five basic assumptions to reduce the number RAPID model of unknowns: RAPID is a 3D radiative transfer model, able to simulate reflectance images over complex 3D natural scenes at 1) The DTM (digital terrain model) remained large scales (30 to 1000 m) with great efficiency (Huang unchanged, a reasonable assumption for forested et al. 2013a), implying that RAPID can simulate images areas; Huang and Lian Forest Ecosystems (2015) 2:20 Page 5 of 13 Fig. 4 Simulation framework to generate time series of optical images 2) Tree crowns were ellipsoid or cone shaped, similar content) while fixing others, seasonal leaf reflectance with geometric optic models (Schaaf and Strahlerl and transmittance could be simulated (Barry et al. 2009). 1994; Chen et al. 2012); Previous studies have shown that the amount of leaf 3) Individual tree LAI (LAI ) was predicted from tree chlorophyll is correlated with NDVI (Wu et al. 2008; tree height (Xiao et al. 2006); Rulinda et al. 2011; Croft et al. 2013; Feng and Niu 2014). 4) We accepted a spherical leaf angle distribution Therefore, we used a linear relationship between MODIS (LAD ) for all trees due to missing measurements; NDVI products (0.1 to 1.0) and the amount of leaf chloro- tree −2 5) Reflectance of non-vegetation objects, such as walls, phyll (10 to 100 μg∙cm ). water bodies and roads remained constant, following precedents set in existing literature (Wang et al. 2008) Background reflectance or in the ENVI spectral library. Seasonal variation in background reflectance was com- plex. However, soil moisture played a major role (Muller With these assumptions, there were two types of input and Décamps 2001; Weidong et al. 2002; Whiting et al. parameters: fixed or dynamic. Fixed parameters were 2004), which was then derived from TVDI (temperature DEM, land cover map, individual tree map (coordinates, vegetation dryness index) inferred from temperature and DBH, height, crown radius, crown length). DEM and NDVI (Sandholt et al. 2002; Liang et al. 2014). TVDI is land cover map were re-sampled to a resolution of 1 m. highly correlated with soil moisture (Holzman et al. 2014). Land cover maps were generated using a decision tree Therefore, we estimated the soil reflectance as the weighted method with six classes: bare soil, road, birch forest, average of dry soil and wet soil reflectance, where the larch forest, water surface and buildings. Decision rules weights were TVDI and (1-TVDI) respectively. The back- were largely based on the Ratio Vegetation Index (RVI), ground was defined as soil covered by a homogeneous the Normalized Difference Water Index (NDWI) and shrub layer with a LAI of 0.5. Shrub leaves were assumed CHM. Dynamic parameters determined the seasonal to have the same optical parameters as birch leaves. change of reflectance, such as component reflectances, LAI and sun position, obtained mainly from time series Growing season of MODIS products, including NDVI, LAI and land sur- From the Landsat classification map, pure birch and larch face temperature (LST). pixels were selected to determine the beginning and final day (DOY) of the annual growing season, using the follow- Leaf reflectance ing phenology analysis. Leaf reflectance and transmittance were measured only First, MODIS time series NDVI data were fitted using once, which was not sufficient to represent optical features a harmonic analysis (Jonsson and Eklundh 2004) to re- for the entire growing season. Thanks to the PROSPECT move random noises. We referred to the maximum and model and changing the most sensitive input (leaf chlorophyll minimum values of NDVI as NDVI and NDVI . max min Huang and Lian Forest Ecosystems (2015) 2:20 Page 6 of 13 The starting date was defined as the DOY when NDVI equation. Both species showed very similar phenology in exceeded 20 % of (NDVI – NDVI ) between DOY the spring without a difference on the average (Delbart max min 1 and DOY 180. Similarly, the final date was the DOY et al. 2005), so we used the MODIS LAI to calibrate the when NDVI exceeded 20 % of (NDVI – NDVI ) g value for both species: max min between DOY 180 and DOY 365. gðÞ DOY f  H  πR ¼ LAI  Area i MODIS MODIS i i Temporal leaf area index of individual trees LAI  Area MODIS MODIS It was difficult to calculate LAI precisely. Instead, it ⇒gðÞ DOY¼ tree f  H  πR was possible to allocate the MODIS LAI into individual trees. Based on assumption (3), LAI is linearly related ð3Þ tree to tree height (H ) for each species. Therefore, LAI tree tree is a function of both species and DOY (see Equation 1). LAI ¼ fðÞ species gðÞ DOY H ð1Þ tree tree Evaluation method Since the TreeVaW (Popescu and Wynne 2004) had not where f is a coefficient relating H to LAI , which is tree tree been tested in our study site, we manually segmented a constant for each species and g is a temporal correction few tree crowns in nine sub-plots with different tree factor. Plot LAI and individual tree height in field plots densities in order to evaluate the accuracy of the ex- were used to calibrate f values for both birch or larch: tracted number of trees, height, location and crown f  H  πR ¼ LAI  Area radius. i plot plot i i LAI  Area We carried out four types of evaluations: (a) CCD plot plot ð2Þ ⇒f ¼ image was used to check the pattern of simulated half- H  πR meter images; (b) Landsat images were used to check For birch trees, we calibrated f as 0.25 and 0.20 for reflectance values of nadir images at the same date; (c) larch. From previous studies (Li et al. 2009; Liu and Jin MODIS BRF products were used to compare simulated 2013), we determined that the LAI of birch and larch BRFs and(d) finally, weusedfourdates of Landsatim- varied with DOY and could be fitted with a polynomial ages to evaluate temporal simulations. Fig. 5 Comparisons of tree segmentation between manual operation and TreeVaW in a sparse subplot (a-b) and a dense subplot (c-d); (a) and (c) are manual results; (b) and (d) are TreeVaW results Huang and Lian Forest Ecosystems (2015) 2:20 Page 7 of 13 LðÞ birch¼ 0:5475H−0:0118 R ¼ 0:87 ð6Þ LðÞ larch ¼ 0:6551H−0:0731 R ¼ 0:55 ð7Þ Growing season Figure 6 shows the smoothed NDVI curves for birch and larch-dominated forests. The starting date, final date and length of the growing season were estimated as DOY 140 (May 20), 273 (Sep 3) and 130 days. During the growing season, the birch forest had higher NDVI values than larch forests, and the larch forests in the flat wetland area had significant lower NDVI values than those in mountain areas. Fig. 6 Smoothed MODIS 16-day 250 m NDVI products in 2013 Land cover classification The Landsat 8 image on May 24, 2014 was used to pro- Results duce a 15 m classification map, given a suitable growing Tree structure season and good image quality to distinguish birch and Compared to manual segmentation, TreeVaW detected larch (Fig. 7a). Compared to the old forest map (Fig. 7b), 88 % of the number of trees in sparse plots (Fig. 5a, b), the southern regions (1 and 2) visually matched much but only 74 % in dense plots (Fig. 5c, d). Crown radii ob- better than the northern regions (3 and 4). Fortunately, the tained from TreeVaW ranged from 0.59 to 0.71 m, lower major study area was located in regions 1 and 2, where the than those from manual segmentation. The mean tree accuracy (around 75 %) was calculated from random sam- height error and location bias of detected trees was 0.88 pling points. Major rules of the decision tree were the fol- and 0.91 m. lowing: (1) forest vegetations = (RVI > 0 and NDWI > 0 Based on regression analysis of plot data, both tree and CHM > 2 m); (2) shrubs or grasses = (RVI > 0 and DBH and crown length (L) were well predicted from tree NDWI > 0 and CHM ≤ 2 m); (3) birch = ((1) and RVI > height (H) with coefficients of determination larger than 7.0); (4) larch = ((1) and RVI ≤ 5.0); (5) mixed forests = ((1) 0.80: and RVI > 5.0 and RVI ≤ 7.0). Forest understory in the Genhe Reserve was complex 1:5652 2 DBHðÞ birch¼ 0:2466H R ¼ 0:89 ð4Þ but of considerable value in identifying forest types (see Table 1). Some shrubs were evergreen, while grasses shed 1:8704 2 DBHðÞ larch ¼ 0:1639H R ¼ 0:82 ð5Þ leaves. Therefore, the vegetation detected in SPOT-6 AB Fig. 7 Vegetation map: (a) classified image with the lightest greenness representing birch forests; (b) forest map with white color representing birch forests Huang and Lian Forest Ecosystems (2015) 2:20 Page 8 of 13 Fig. 8 Determining vegetation as evergreen bush in winter season: (a) evergreen understory on Spot-6 image (red color); (b) CCD image; (c) CHM image image (1 m, October) was used to define shrubs as Comparisons of BRF evergreen vegetation because only evergreens had green Five pixels around the central study area showed variation leaves at that time of the year (Fig. 8). in the BRF curves, used as a reference to evaluate the RAPID BRF results (Fig. 11). Generally, the simulated BRF matches the shape of MODIS BRF in spite of absolute Comparisons of nadir images biases in a few view directions. First, the simulated red Simulated nadir images (0.5 m resolution) were compared BRF is higher than all MODIS BRFs when the view zenith to the CCD image in Fig. 9. The spatial texture and land angle (VZA) is between −50° and 40°. Second, in both red cover difference are consistent, but the simulated forests and NIR bands, the backscattering BRF when VZA larger look sparser. than 50°, is lower than the MODIS BRF. The spectral results were compared with Landsat 8 re- flectance images on May 24, 2014 (Fig. 10). Both simulation and Landsat images showed typical vegetation reflectance Temporal results spectra (low red reflectance and high near infrared (NIR) Four Landsat images were used to check the simulation reflectance). Simulation results are significantly lower in ability of temporal variations; the dynamic parameters of blue bands (0.02 to 0.06). birch and larch trees are shown in Table 2. Fig. 9 Comparing nadir image (0.5 m) with CCD: (a) simulated image (R = Near infrared (NIR), G = red, B = green); (b) airborne CCD mosaic image from multiple days Huang and Lian Forest Ecosystems (2015) 2:20 Page 9 of 13 Fig. 10 Comparison of nadir top of canopy (TOC) reflectance image with a Landsat 8 image using linear stretch (0 to 0.3): (a) simulated image (0.5 m, R = NIR, G = red, B = green); (b) re-sampled 15 m image from (a); (c) Landsat 8 (15 m) on May 24, 2014 (R = NIR, G = red, B = green); (d) Spectral curves of dense and sparse canopies Figure 12 compares the results between simulated Discussion and real images in stripes of birch and larch forest Our main objective was to create and test how to couple stands (600 m by 600 m). The resolutions were 15 m lidar data and temporal optical data MODIS in order to except forthe LandsatTMimage (30m)onSeptember simulate high-resolution optical satellite images. A frame- 5, 2013. The birch bands (marked as A) showed signifi- work was built and tested at the Genhe Forest Farm. In cant variation in reflectance from brown color (bare spite of some biases or errors, the approach successfully soil), red color (green canopy), pink color (dense can- produced temporal images with high spatial, spectral and opy) to mixed color (discoloring canopy), reconstructed angular resolutions, which confirmed the possibility to from simulated images in spite of slightly different fuse lidar and MODIS data. colors. In the lower part of the Landsat ETM+ image, a black no-data area showed up, due to a sensor error Major contributors on simulation (SLC-OFF). The results on Sept 5, 2013 showed larger The framework included four main data sources: lidar, discrepancies. Landsat, MODIS and field data. To drive a 3D model, Huang and Lian Forest Ecosystems (2015) 2:20 Page 10 of 13 Fig. 11 Comparisons between MODIS BRF product and RAPID simulations: (a) red band (0.620–0.670 nm); (b) NIR band (0.841–0.876 nm) the most important inputs were 3D scenes and the inside Major errors reflectance and transmittance of 3D objects. Lidar was the Despite the fact that three types of evaluation on re- first contributor to providing 3D structures of individual flectance, i.e., spatial texture, BRF and Landsat simula- trees and background. Lidar-derived 3D structures were tion demonstrated the capability to simulate temporal normally static, but 3D scenic objects, especially their LAI, images, quantitative validation was still missing due to were dynamic. Therefore, we used an allocation method lots of uncertainties in the entire workflow. We tried to to downscale MODIS LAI into each tree, a technique not address major error sources and assess their uncertainty: found in previous studies. Landsat images were used to classify birch and larch, supporting the generation of 1) 3D structure errors: 3D trees. The optical parameters of 3D objects were collected in It has to be admitted that suppressed trees and irregular the field or obtained from existing references; these were tree crowns are hard to detect from CHM. A previous also dynamic. Therefore, MODIS NDVI data were used to study has shown that TreeVaW method can identify more calibrate leaf chlorophyll for the PROSPECT model, which than 95 % of the trees in planted forests but only 70 % in then simulated dynamic leaf reflectance and transmittance. natural forests (Antonarakis et al. 2008). Although other Background soil reflectance varied over time and was diffi- detection algorithms may help improve the accuracy, the cult to obtain. An alternative is to use TVDI to adjust soil inter-comparisons between detection methods found that reflectance, which is a more recent idea and needs to be the correct percentage of the number of trees was gener- evaluated in any future research. ally between 50 to 90 % (Kaartinen et al. 2012). In our Table 2 Dynamic parameters of birch or larch forests −2 2 −2 2 −2 Date DOY Chlorophyll content (μg∙cm ) LAI (m ∙m ) LAI (m ∙m ) TVDI larch birch May 5, 2013 125 41 0.16 × H 0.13 × H 0.61 tree tree May 24, 2014 144 52 0.17 × H 0.18 × H 0.68 tree tree June 27, 2012 179 85 0.20 × H 0.23 × H 0.91 tree tree Sept 5, 2013 248 72 0.19 × H 0.11 × H 0.62 tree tree Huang and Lian Forest Ecosystems (2015) 2:20 Page 11 of 13 Fig. 12 Comparison between simulated and Landsat images with false color composition (RGB = [NIR, RED, GREEN]); A and B represent birch and larch trees, respectively study, the percentage of the correct number was between season will be low. Continuous field observations are 74 to 90 %, which is consistent with results above. The strongly suggested. high level of missed detection leads to a higher clumping effect and sparser forests (Figs. 7 and 8), which then re- 4) MODIS data uncertainty sulted in higher reflectance biases induced by background uncertainties. The most recent MODIS LAI product is Collection 5 (this a version code), which has uncertainties around +/−1.0 2) Unknown background: for relatively pure pixels (Fang et al. 2012). However, considering the low resolution of MODIS pixels, the Although we classified evergreen bush, its background uncertainties of inversed LAIs are even larger for mixed type and dynamic reflectance were almost unknown. pixels. The image matching between MODIS (1 km) Therefore, a very rude LAI of 0.5 was assumed for all and CHM (0.5 m) sounds tricky. However, as the only understories. In fact, it is possible to retrieve forest available product, it was used in our simulation frame- background reflectance from satellite data (Canisius and work. In any future study, we will use Landsat images Chen 2007; Pisek and Chen 2009; Pisek et al. 2010; to bridge the gap of higher resolution of LAI products Tuanmu et al. 2010; Rautiainen et al. 2011; Pisek et al. (Gao et al. 2014). The BRF biases between simulation 2012). We will try these methods to inverse background and MODIS can be partially attributed to the limitations reflectance in later studies. We were able to validate the of MODIS BRF in reconstructing higher and narrower TVDI-adjusted soil reflectance, although it should have hotspots (Huang et al. 2013b). directional effects. Actually, we used isotropic soil reflect- ance, which may explain the BRF biases with large angles 5) Landsat data uncertainty in backward view directions. Landsat images were used to compare simulated nadir 3) Leaf discoloring reflectance and image textures. Figure 10 shows sig- nificant differences in the blue band, which can be In September, the leaves of both birch and larch chan- largely explained by an atmospheric correction error ged color. However, these changes varied even for trees because blue band reflectance should be lower after a of the same species, probably an effect of age, elevation correct removal of aerosol scatter. This atmospheric or density, making it difficult to identify individual trees. correction was carried out by using the FLAASH mod- Therefore, the accuracy in discoloration during the growing ule of the ENVI 5.1 software, where the aerosol optical Huang and Lian Forest Ecosystems (2015) 2:20 Page 12 of 13 depth and water content were only estimated from scale optical image dataset will be useful to support the images. understanding of scale problems. Conclusions Efficiency problems We presented a simulation framework which links lidar RAPID is relatively fast with 3D models, but running with optical images to produce series of temporal im- one case (1 km) still needs four to six hours at individual ages. The study provides a proof-of-concept approach to tree scale on a workstation (using 10 CPU cores). We link lidar data in the parameterization of a RAPID model are of the opinion that it is not feasible for the gener- for temporal image reconstruction in forest dominated ation of operational products. However, for some scien- areas. Demonstrations were applied at the Genhe Forest tific use, focused on local areas, it may be worthwhile to Farm, a remote forest reserve in China. Evaluations on obtain images with very high resolutions (spatial, spec- nadir reflectance, spatial textures and BRF confirmed tral and angular) for research within an acceptable time that 3D simulation provides an insight look into how im- frame and cost structure. Because RAPID can run at ages vary over time. Many uncertainties were identified, scalable resolutions, 3D scenes of dense forests can be which can be expected to be reduced in any future study. up-scaled to regular grids with a medium resolution (e.g. Strategies to improve efficiency are possible and discussed. 5 to 10 m), which significantly improves calculation effi- ciency (less than 30 min) without much loss in accuracy. Competing interests Furthermore, we can create a reference table of 3D scenes, The authors declare that they have no competing interests. classifying a study area into fewer categories with possible combinations of DEM, understory, tree locations, tree Authors’ contributions HH is the major and contact author for most work. JL processed a small part heights and tree LAI. The corresponding reflectance of the data. Both authors read and approved the final manuscript. images will be simulated and stored as an image database. Once the database is created, a quick search method can be Acknowledgements used to pick up desired images based on input parameters The authors gratefully acknowledge the Chinese National Basic Research Program (2013CB733401) and the Chinese Natural Science Foundation such as DEM, understory, tree distribution and LAI. In the Project (41171278). current framework, we only dealt with the capability of coupling simulation. Improvements will be presented Received: 28 February 2015 Accepted: 18 June 2015 in our next study. References Adams T, Beets P, Parrish C (2012) Extracting more data from LiDAR in forested Scale issues areas by analyzing waveform shape. Remote Sens 4(3):682–702. Scale effect and scaling have been big issues in remote doi:10.3390/rs4030682 sensing community. When models or algorithms at small Ahl DE, Gower ST, Burrows SN, Shabanov NV, Myneni RB, Knyazikhin Y (2006) Monitoring spring canopy phenology of a deciduous broadleaf forest using scales are used at large scales, they may produce certain MODIS. Remote Sens Environ 104(1):88–95. doi:10.1016/j.rse.2006.05.003 errors, especially for non-linear models (Tao et al. 2009). Antonarakis AS, Richards KS, Brasington J, Bithell M, Muller E (2008) Retrieval of Scale issues constrain the accuracy of retrieval and limit vegetative fluid resistance terms for rigid stems using airborne lidar. J Geophys Res 113(G2):G02S07. doi:10.1029/2007JG000543 the development of remote sensing applications. In this Arno J, Escola A, Valles JM, Llorens J, Sanz R, Masip J, Palacin J, Rosell-Polo JR study, MODIS LAI products at a 1 km scale were down- (2013) Leaf area index estimation in vineyards using a ground-based LiDAR scaled to the level of LAI of individual trees, using lidar at scanner. Precision Agric 14(3):290–306. doi:10.1007/s11119-012-9295-0 Barbier N, Proisy C, Vega C, Sabatier D, Couteron P (2011) Bidirectional texture a scale of only a few meters, used to coincide with the function of high resolution optical images of tropical forest: An approach RAPID model scale. Assuming that MODIS LAI products using LiDAR hillshade simulations. Remote Sens Environ 115(1):167–179. are scale-corrected, this scaling does not change total leaf doi:10.1016/j.rse.2010.08.015 Barry KM, Newnham GJ, Stone C (2009) Estimation of chlorophyll content in area. Since LAI was found to be correlated with tree tree Eucalyptus globulus foliage with the leaf reflectance model PROSPECT. Agr Forest height, the down-scaling model should also be linear with- Meteorol 149(6–7):1209–1213, http://dx.doi.org/10.1016/j.agrformet.2009.01.005 out model scale issues. In comparison with satellite images, Canisius F, Chen JM (2007) Retrieving forest background reflectance in a boreal region from Multi-anglo Imaging SpectroRadiometer (MISR) data. Remote the simulated RAPID images should also be re-sampled to Sens Environ 107(1–2):312–321. doi:10.1016/j.rse.2006.07.023 coincide with the satellite scale, including the MODIS and Chen G, Wulder MA, White JC, Hilker T, Coops NC (2012) Lidar calibration and Landsat scales. Fortunately, reflectance or radiance images validation for geometric-optical modeling with Landsat imagery. Remote Sens Environ 124(0):384–393, http://dx.doi.org/10.1016/j.rse.2012.05.026 are scale independent. Croft H, Chen JM, Zhang Y, Simic A (2013) Modelling leaf chlorophyll content in Using MODIS and lidar data to reconstruct 3D scenes broadleaf and needle leaf canopies from ground, CASI, Landsat TM 5 at an individual tree scale, the RAPID model is capable and MERIS reflectance data. Remote Sens Environ 133(0):128–140, http://dx.doi.org/10.1016/j.rse.2013.02.006 of simulating time series of images at spatial scales from Delbart N, Kergoat L, Le Toan T, Lhermitte J, Picard G (2005) Determination of 0.5 m to 1 km and temporal scales of a few days. Although phenological dates in boreal regions using normalized difference water the 3D scene estimation is not perfect, this type of multi- index. Remote Sens Environ 97(1):26–38. doi:10.1016/j.rse.2005.03.011 Huang and Lian Forest Ecosystems (2015) 2:20 Page 13 of 13 Fang H, Wei S, Liang S (2012) Validation of MODIS and CYCLOPES LAI products Muller E, Décamps H (2001) Modeling soil moisture–reflectance. Remote Sens using global field measurement data. Remote Sens Environ 119:43–54. Environ 76(2):173–180, http://dx.doi.org/10.1016/S0034-4257(00)00198-X doi:10.1016/j.rse.2011.12.006 Nitze I, Barrett B, Cawkwell F (2015) Temporal optimisation of image acquisition Feng M, Niu Z (2014) Chlorophyll content retrieve of vegetation using Hyperion for land cover classification with Random Forest and MODIS time-series. Int J data based on empirical models. Remote Sens Land Resour 26(1):71–77. Appl Earth Observ Geoinformation 34:136–146. doi:10.1016/j.jag.2014.08.001 doi:10.6046/gtzyyg.2014.01.13 Pisek J, Chen JM (2009) Mapping forest background reflectivity over North Gao F, Anderson MC, Kustas WP, Houborg R (2014) Retrieving leaf area index America with Multi-angle Imaging SpectroRadiometer (MISR) data. Remote from landsat using MODIS LAI products and field measurements. IEEE Geosci Sens Environ 113(11):2412–2423. doi:10.1016/j.rse.2009.07.003 Remote Sens Lett 11(4):773–777. doi:10.1109/lgrs.2013.2278782 Pisek J, Chen JM, Miller JR, Freemantle JR, Peltoniemi JI, Simic A (2010) Mapping Gao F, Masek J, Schwaller M, Hall F (2006) On the blending of the Landsat and forest background reflectance in a boreal region using multiangle compact MODIS surface reflectance: predicting daily Landsat surface reflectance. IEEE airborne spectrographic imager data. IEEE Trans Geosci Remote Sens Trans Geosci Remote Sens 44(8):2207–2218. doi:10.1109/tgrs.2006.872081 48(1):499–510. doi:10.1109/tgrs.2009.2024756 Pisek J, Rautiainen M, Heiskanen J, Mottus M (2012) Retrieval of seasonal García Millán VE, Sánchez-Azofeifa A, Málvarez García GC, Rivard B (2014) Quantifying dynamics of forest understory reflectance in a Northern European boreal tropical dry forest succession in the Americas using CHRIS/PROBA. Remote Sens forest from MOD'S BRDF data. Remote Sens Environ 117:464–468. Environ 144(0):120–136, http://dx.doi.org/10.1016/j.rse.2014.01.010 doi:10.1016/j.rse.2011.09.012 Hilker T, Wulder MA, Coops NC, Seitz N, White JC, Gao F, Masek JG, Stenhouse G Popescu SC, Wynne RH (2004) Seeing the trees in the forest: using lidar and (2009) Generation of dense time series synthetic Landsat data through data multispectral data fusion with local filtering and variable window size for blending with MODIS using a spatial and temporal adaptive reflectance fusion estimating tree height. Photogrammetric Eng Remote Sens 70(5):589–604 model. Remote Sens Environ 113(9):1988–1999. doi:10.1016/j.rse.2009.05.011 Rautiainen M, Lang M, Mõttus M, Kuusk A, Nilson T, Kuusk J, Lükk T (2008) Holzman ME, Rivas R, Bayala M (2014) Subsurface soil moisture estimation by Multi-angular reflectance properties of a hemiboreal forest: An analysis VI-LST method. IEEE Geosci Remote Sens Lett 11(11):1951–1955. using CHRIS PROBA data. Remote Sens Environ 112(5):2627–2642, doi:10.1109/lgrs.2014.2314617 http://dx.doi.org/10.1016/j.rse.2007.12.005 Huang H, Qin W, Liu Q (2013a) RAPID: a radiosity applicable to porous individual Rautiainen M, Mottus M, Heiskanen J, Akujarvi A, Majasalmi T, Stenberg P (2011) objects for directional reflectance over complex vegetated scenes. Remote Seasonal reflectance dynamics of common understory types in a northern Sens Environ 132:221–237. doi:10.1016/j.rse.2013.01.013 European boreal forest. Remote Sens Environ 115(12):3020–3028. Huang X, Jiao Z, Dong Y, Zhang H, Li X (2013b) Analysis of BRDF and albedo retrieved doi:10.1016/j.rse.2011.06.005 by kernel-driven models using field measurements. IEEE J Selected Topics Appl Rulinda CM, Bijker W, Stein A (2011) The chlorophyll variability in Meteosat Earth Observ Remote Sens 6(1):149–161. doi:10.1109/jstars.2012.2208264 derived NDVI in a context of drought monitoring. Procedia Environ Sci Inglada J, Hagolle O, Dedieu G (2011) A framework for the simulation of high 3(0):32–37, http://dx.doi.org/10.1016/j.proenv.2011.02.007 temporal resolution image series. Paper presented at the IEEE International Sandholt I, Rasmussen K, Andersen J (2002) A simple interpretation of the surface Geoscience and Remote Sensing Symposium (IGARSS), 24–29 July 2011 temperature/vegetation index space for assessment of surface moisture Jacquemoud S, Verhoef W, Baret F, Bacour C, Zarco-Tejada PJ, Asner GP, François status. Remote Sens Environ 79(2–3):213–224, http://dx.doi.org/10.1016/ C, Ustin SL (2009) PROSPECT + SAIL models: a review of use for vegetation S0034-4257(01)00274-7 characterization. Remote Sens Environ 113(Supplement 1 (0)):S56–S66, Schaaf CB, Strahlerl AH (1994) Validation of bidirectional and hemispherical http://dx.doi.org/10.1016/j.rse.2008.01.026 reflectances from a geometric-optical model using ASAS imagery and Jonsson P, Eklundh L (2004) TIMESAT - a program for analyzing time-series of satellite pyranometer measurements of a spruce forest. Remote Sens Environ sensor data. Comput Geosci 30(8):833–845. doi:10.1016/j.cageo.2004.05.006 49(2):138–144, http://dx.doi.org/10.1016/0034-4257(94)90050-7 Kaartinen H, Hyyppä J, Yu X, Vastaranta M, Hyyppä H, Kukko A, Holopainen M, Sharma RC, Kajiwara K, Honda Y (2013) Estimation of forest canopy structural Heipke C, Hirschmugl M, Morsdorf F, Næsset E, Pitkänen J, Popescu S, Solberg S, parameters using kernel-driven bi-directional reflectance model based Wolf BM, Wu J-C (2012) An international comparison of individual tree detection multi-angular vegetation indices. ISPRS J Photogrammetry Remote Sens and extraction using airborne laser scanning. Remote Sensing 4(4):950–974 78:50–57. doi:10.1016/j.isprsjprs.2012.12.006 Kobayashi H, Suzuki R, Kobayashi S (2007) Reflectance seasonality and its relation Tao X, Yan B, Wang K, Wu D, Fan W, Xu X, Liang S (2009) Scale transformation to the canopy leaf area index in an eastern Siberian larch forest: Multi-satellite of Leaf Area Index product retrieved from multiresolution remotely sensed data and radiative transfer analyses. Remote Sens Environ 106(2):238–252. data: analysis and case studies. Int J Remote Sens 30(20):5383–5395. doi:10.1016/j.rse.2006.08.011 doi:10.1080/01431160903130978 Li GZ, Wang HX, Zhu JJ (2009) Monthly changes of leaf area index and canopy Tuanmu M-N, Vina A, Bearer S, Xu W, Ouyang Z, Zhang H, Liu J (2010) Mapping openness of Larix olgensis in mountainous regions in east Liaoning province. understory vegetation using phenological characteristics derived from remotely J Northeast Forestry Univ 37(7):13 sensed data. Remote Sens Environ 114(8):1833–1844. doi:10.1016/j.rse.2010.03.008 Liang L, Zhao S-h, Qin Z-h, He K-x, Chen C, Luo Y-x, Zhou X-d (2014) Drought Wang JD, Zhang LX, Liu QH, Zhang B, Yin Q (2008) The spectrum knowledge change trend using MODIS TVDI and its relationship with climate factors database of typical land surface objects in China. Science Press, Beijing in china from 2001 to 2010. J Integr Agric 13(7):1501–1508, Wanner W, Li X, Strahler AH (1995) On the derivation of kernels for kernel-driven http://dx.doi.org/10.1016/S2095-3119(14)60813-3 models of bidirectional reflectance. J Geophys Res 100(D10):21077–21089. Liu ZL, Jin GZ (2013) Estimation of leaf area index of secondary Betula platyphylla doi:10.1029/95JD02371 forest in Xiaoxing' an Mountains. Acta Ecologica Sinica 33(8):9 Weidong L, Baret F, Xingfa G, Qingxi T, Lanfen Z, Bing Z (2002) Relating soil Lucas RM, Lee AC, Williams ML (2006) Enhanced simulation of radar backscatter surface moisture to reflectance. Remote Sens Environ 81(2–3):238–246, from forests using LiDAR and optical data. IEEE Trans Geosci Remote Sens http://dx.doi.org/10.1016/S0034-4257(01)00347-9 44(10):2736–2754. doi:10.1109/tgrs.2006.881802 Whiting ML, Li L, Ustin SL (2004) Predicting water content using Gaussian model on Lunetta RS, Johnson DM, Lyon JG, Crotwell J (2004) Impacts of imagery temporal soil spectra. Remote Sens Environ 89(4):535–552, http://dx.doi.org/10.1016/ frequency on land-cover change detection monitoring. Remote Sens Environ j.rse.2003.11.009 89(4):444–454. doi:10.1016/j.rse.2003.10.022 Wu C, Niu Z, Tang Q, Huang W (2008) Estimating chlorophyll content from Masek JG, Huang C, Wolfe R, Cohen W, Hall F, Kutler J, Nelson P (2008) North hyperspectral vegetation indices: Modeling and validation. Agr Forest Meteorol American forest disturbance mapped from a decadal Landsat record. Remote 148(8–9):1230–1241, http://dx.doi.org/10.1016/j.agrformet.2008.03.005 Sens Environ 112(6):2914–2926. doi:10.1016/j.rse.2008.02.010 Wu M-Q, Wang J, Niu Z, Zhao Y-Q, Wang C-Y (2012) A model for spatial and temporal Montesano P, Cook B, Sun G, Simard M, Nelson R, Ranson K, Zhang Z, Luthcke S data fusion. J Infrared Millimeter Waves 31(1):80–84 (2013) Achieving accuracy requirements for forest biomass mapping: a Xiao C-W, Janssens IA, Curiel Yuste J, Ceulemans R (2006) Variation of specific leaf spaceborne data fusion method for estimating forest biomass and LiDAR area and upscaling to leaf area index in mature Scots pine. Trees 20(3):304–310. sampling error. Remote Sens Environ 130:153–170. doi:10.1016/ doi:10.1007/s00468-005-0039-x j.rse.2012.11.016 Xu G-c, Pang Y, Li Z-y, Zhao K-r, Liu L-x (2013) The changes of forest canopy Mu X, Zhang Q, Liu Q, Pang Y, Hu K (2015) A study on typical forest biomass spectral reflectance with seasons in Xiaoxing'anling. Spectrosc Spectral Anal mapping technology of great khingan using airborne laser scanner data. 33(12):3303–3307 Remote Sens Technol Appl 30(2):220–225

Journal

"Forest Ecosystems" – Springer Journals

Published: Dec 1, 2015

Keywords: Ecology; Ecosystems; Forestry

There are no references for this article.