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

Learn More →

Inter-calibration of DMSP-OLS and SNPP-VIIRS-DNB annual nighttime light composites using machine learning

Inter-calibration of DMSP-OLS and SNPP-VIIRS-DNB annual nighttime light composites using machine... GISCIENCE & REMOTE SENSING 2020, VOL. 57, NO. 8, 1144–1165 https://doi.org/10.1080/15481603.2020.1848323 Inter-calibration of DMSP-OLS and SNPP-VIIRS-DNB annual nighttime light composites using machine learning Sumana Sahoo , Prasun Kumar Gupta and S. K. Srivastav Indian Institute of Remote Sensing, Dehradun, India ABSTRACT ARTICLE HISTORY Received 22 April 2020 The satellite-based nighttime lights (NTL) data from the Defense Meteorological Satellite Program’s Accepted 3 November 2020 Operational Linescan System (DMSP-OLS), available in the public domain from 1992 to 2013, are extensively used for socio-economic studies. The improved NTL products from the Visible Infrared KEYWORDS Imaging Radiometer Suite’s Day/Night Band (VIIRS-DNB), on-board the Suomi National Polar-Orbiting VIIRS-DNB; DMSP-OLS; inter- Partnership spacecraft and National Oceanic and Atmospheric Administration – 20 (NOAA-20) space- calibration; machine craft’s, are now available since April 2012. This study investigates the potential of machine-learning learning; multi-layer algorithms for inter-calibrating them (i.e., DMSP-OLS and VIIRS-DNB) to produce time-series annual perceptron; random forest VIIRS-DNB-like NTL datasets for the time when VIIRS-DNB data did not exist, for long-term studies. Uttar Pradesh, one of the most populous and largest States of India, is selected as the study area. Two machine-learning algorithms are utilized: (1) Multi-Layer Perceptron (MLP), having deep neural networks (DNN) architecture, and (2) Random Forest (RF), a widely used method. The DMSP-OLS and VIIRS-DNB data of 2013 (common year of data availability) and ancillary data pertaining to land cover, topography, and road network are used to train the models. The qualitative and quantitative analysis of annual VIIRS-DNB-like NTL images simulated from annual DMSP-OLS composites of 2004–2012 indicates that RF captures better spatial details at the local-scale and is able to efficiently handle the saturation problem at urban centers; while MLP is found to be superior at regional-scale. Both MLP and RF models significantly reduce the blooming effect around settlements, a common problem observed in DMSP-OLS data. It is inferred that depending on the research objectives, both RF and MLP algorithms can be appropriately utilized for producing VIIRS-DNB-like NTL images from DMSP-OLS annual NTL composites. The research can be further expanded by using other DNN architecture-based algorithms and improved spatio-temporal ancillary datasets over areas with different socio-economic, physiographic, and climatic settings. 1. Introduction Administration – 20 (NOAA-20) spacecraft’s, over- For almost two decades, various research works have comes some of the limitations of the DMSP-OLS, involved the use of the Defense Meteorological mainly by providing calibrated NTL data at improved Satellite Program’s Operational Linescan System spatial and radiometric resolutions (Table 1). The (DMSP-OLS) nighttime lights (NTL) time-series data SNPP-VIIRS-DNB (hereafter referred to as VIIRS-DNB) to track and monitor the growth of electrified settle- data made available from April 2012 onwards by the ments (Elvidge et al. 1997; Kiran Chand et al. 2009; Earth Observations Group (EOG) at NOAA’s National Townsend and Bruce 2010; Levin and Duke 2012; Ma Centers for Environmental Information (NOAA-NCEI) et al. 2012; Zhao, Ghosh, and Samson 2012; Min et al. in the form of monthly NTL composites are being 2013; Min and Gaba 2014; Cao et al. 2019). Although widely used by the researchers for different applica- these DMSP-OLS annual NTL composites have helped tions such as, socio-economic dynamics (Shi et al. study social, demographical, and economic temporal 2015; Bennett and Smith 2017; Zhao et al. 2017), dynamics, the data consists of certain technical limita- urban dynamics (Chen et al. 2015; Guo et al. 2018; tions such as lack of calibration, saturation, and Yu et al. 2018), light pollution (Duriscoe, Luginbuhl, blooming effect. The Visible Infrared Imaging and Elvidge 2013; Falchi et al. 2016), military conflicts Radiometer Suite’s Day/Night Band (VIIRS-DNB), on- (Li et al. 2015; Witmer 2015; Levin, Ali, and Crandall board the Suomi National Polar-Orbiting Partnership 2018), etc. The characteristic differences between (SNPP), and National Oceanic and Atmospheric DMSP-OLS and VIIRS-DNB products are shown in CONTACT Prasun Kumar Gupta prasun@iirs.gov.in © 2020 Informa UK Limited, trading as Taylor & Francis Group GISCIENCE & REMOTE SENSING 1145 Table 1. Characteristic differences between DMSP-OLS and Inter-calibration of multi-satellite data is crucial for VIIRS-DNB data. detecting and quantifying the changes in the Earth’s DMSP-OLS VIIRS-DNB environment, for predicting weather conditions, for Spatial resolution 2.7 km 742 m understanding various climatic processes, and monitor- Radiometric quantization 6 bit 14 bit Detection range Limited Large ing land cover changes (Chander et al. 2013). Similarly, On-board calibration Not available Available Saturation problem Present at urban centers Eliminated the inter-calibration between DMSP-OLS and VIIRS-DNB Blooming effect Present Eliminated NTL datasets can help in producing consistent time- series products, which will be useful for long-term ana- lysis of various socio-economic factors. Efforts were Table 1. Research (Bennett and Smith 2017; Zheng, made to calibrate a single-day DMSP-OLS image with Weng, and Wang 2019) has provided a detailed dis- VIIRS-DNB image using Dome C in the Antarctic as the cussion on the inconsistencies amongst DMSP-OLS calibration site (Shao et al. 2014). Another method for satellites; and between DMSP-OLS and VIIRS-DNB inter-calibrating the composites was developed to eval- datasets. Given the fact that DMSP-OLS has served uate city light dynamics during the Syrian civil war which as an important tool for socio-economic studies, the started in 2011 (Li et al. 2017a); however, due to the two datasets (i.e., DMSP-OLS and VIIRS-DNB) can be unavailability of DMSP-OLS data after 2013, an attempt inter-calibrated to create a consistent time-series was made to simulate DMSP-OLS data from VIIRS-DNB dataset for long-term studies. data using the power function and the Gaussian low Several algorithms have been developed to over- pass filter. Li et al. (2020) have used a sigmoid-based come the shortcomings of the DMSP-OLS data in a way regression approach with 2013 as the common year to that can achieve better correlation with Gross Domestic convert VIIRS-DNB datasets to DMSP-like datasets. Product (GDP), Electric Power Consumption (EPC), and Recently, several other attempts have been made by other socio-economic factors (Li and Zhou 2017b; Zhao, researchers to generate DMSP-like products (Ma et al. Zhou, and Samson 2015; Zhao et al. 2017). Radiometric 2020; Zheng, Weng, and Wang 2019). inter-calibration of DMSP-OLS composites is an impor- Several techniques used to perform inter-calibration tant issue, since there was no onboard radiometric cali- of multi-satellite data, are: (1) Regression models (e.g., bration on DMSP satellites (Mukherjee et al. 2017). Elvidge et al. 1997); (2) Statistical inter-calibration (e.g., Several methods have been proposed by researchers Zhang, Pandey, and Seto 2016); (3) Vicarious ground- to address this issue. One of the methods performs based calibration (e.g., Odongo, Hamm, and Milton radiometric inter-calibration based on the pseudo- 2014); (4) Pseudo-invariant calibration sites (e.g., invariant region method, which assumes that there are Mukherjee et al. 2017); and (5) Machine learning (e.g., invariant pixels in multi-temporal NTL images (i.e., pixels Brown et al. 2008). There is a growing need for creating for which the lights have changed very little over time). more sophisticated and robust methods of inter- These invariant pixels are used as training samples to calibration using various machine-learning algorithms generate an inter-calibration function (Elvidge et al. instead of the traditional linear or nonlinear regression 2009; Mukherjee et al. 2017). Other methods for radio- analyses applying empirical and semi-empirical meth- metric calibration are the second-order regression and ods. Machine-learning methods require very less optimal threshold method (Liu et al. 2012), the power- a priori knowledge about each sensor’s data distribu- law regression method (Wu et al. 2013), and the ridge- tions, relationships, sensor operations, calibrations, and line sampling regression method (Zhang, Pandey, and algorithms. These methods learn the patterns from given Seto 2016). Algorithms have also been developed to datasets and work as regularity detectors that discover address the saturation problem. One of the approaches statistically salient properties of investigated data used was inter-calibrating multi-satellite data, wherein (Rodriguez-Galiano et al. 2015; Gumma et al. 2020). Inter- NTL data were combined with Moderate Resolution calibration of multiple satellite and ancillary data for the Imaging Spectroradiometer (MODIS) normalized differ - construction of consistent time-series data using techni- ence vegetation index (NDVI) data to produce vegetation ques like Support Vector Machine (SVM) and Artificial corrected datasets which reduces the saturation effect Neural Networks (ANN) based machine-learning algo- over the urban centers (Zhang, Schaaf, and Seto 2013). rithms attempted earlier (Kwiatkowska and Fargion 1146 S. SAHOO ET AL. 2003; Brown et al. 2008) has proven to be useful. 2.1. Nighttime lights data Likewise, the Random Forest (RF) algorithm is claimed The DMSP-OLS and VIIRS-DNB NTL products available as one of the best machine-learning algorithms for sev- in the public domain (Table 2) are used as primary eral applications (Cracknell and Reading 2014; Kühnlein datasets in this study. The DMSP-OLS annual cloud- et al. 2014; Rodriguez-Galiano et al. 2015). free NTL composites (Version 4), known as stable In the present study, we investigate the potential lights products, produced in 30 arc-second grids are of machine-learning algorithms to inter-calibrate the available from 1992 to 2013 through the website of DMSP-OLS and VIIRS-DNB data with an aim to pro- NOAA-NCEI (NOAA-NCEI 2018). These products show duce VIIRS-DNB-like products for the duration lights from cities, towns, and other sites with persis- 2004–2012, such that these new products can be tent lighting (including gas flares). The ephemeral used in conjunction with the available VIIRS-DNB lights, such as fires, are discarded and the background composites (2013–2020) for long-term socio- noise is replaced with zero pixel value. Pixel values economic studies. The long-term NTL datasets hence range from 0 to 63. The annual NTL composites start- produced have the data range and statistical distribu- ing from the year 2004 to 2013 are used in this study tion close to VIIRS-DNB data. This serves as an impor- (Table 2). There are a few years in which two satellites tant and useful remotely sensed NTL record for have been collecting data. In that case, the annual maintaining uniformity across time (Bennett and composites with maximum Sum of Lights (SOL), Smith 2017). The Multi-Layer Perceptron (MLP) with shown in the gray shade in Table 3, are considered deep neural networks (DNN) architecture and the (Elvidge et al. 2009) for further analysis. widely used Random Forest (RF) algorithms have The VIIRS-DNB NTL datasets (Version 1) are avail- been used to train and test the models, and produce able in the form of monthly composites since the VIIRS-DNB-like annual NTL products from the April 2012. These datasets are produced in 15 arc- DMSP-OLS and ancillary data. second grids and pixel values represent radiance in −2 −1 nW cm sr . In this study, the VIIRS-DNB monthly 2. Study area and data used composites of 2013 are used (Table 2). The present study is carried out for the State of Uttar Pradesh in India (Figure 1), whose capital city is 2.2. Ancillary datasets Lucknow. Lying between 77.1°N and 84.6°N latitudes Human settlements are mainly controlled by physio- and 23.9°E & 30.4°E longitudes, Uttar Pradesh is geo- graphy and socio-economic development of an area graphically the fourth largest state in the country 2 (Fu et al. 2019; Jawarneh, Julian, and Lookingbill 2015; covering an area of 243,290 km . With a population 2 Siddiqui et al. 2018). In this study, DEM and its deri- density of 828 persons per km , it is India’s most vatives are used as proxies for physiography, and populous state. Currently, it is the fourth largest road, land cover, and their derivatives are used as State economy of the country with a gross state proxies for socio-economic development. domestic product of US$ 240 billion (Uttar Pradesh Government 2011). As per reports provided by “Pradhan Mantri Sahaj Bijli Har Ghar Yojana (Prime 2.2.1. Land cover Minister scheme for ease of access to electricity for The Terra/Aqua MODIS Land Cover (Yearly L3 Global each household),” an ongoing electrification scheme, 500 m SIN Grid) products for the years 2004 to 2013 are Uttar Pradesh is among the lowest electrified states in used in this study. Among the five different land cover India (REC Limited 2018). For a State with such a high classification schemes provided [i.e., (i) International population density and poor electrification rate, it is Geosphere Biosphere Programme (IGBP) global vege- beneficial to study the long-term growth patterns of tation classification scheme; (ii) University of Maryland light to get a better understanding of the State’s scheme; (iii) MODIS-derived LAI/fPAR scheme; (iv) electrification performance. MODIS-derived Net Primary Production scheme; (v) The datasets used to carry out the present study Plant Functional Type scheme], IGBP global vegetation are given in Table 2, and a brief description is pro- classification scheme (i.e., Land Cover Type 1) is used vided below. here. The IGBP scheme includes 17 land cover classes: GISCIENCE & REMOTE SENSING 1147 Figure 1. (a) Location and annual DMSP-OLS NTL composite; and (b) annual VIIRS-DNB NTL composite of Uttar Pradesh for the year 2013. The annual composite of VIIRS-DNB is prepared using the monthly NTL composites and PCA method (Sahoo, Gupta, and Srivastav 2019b). 1148 S. SAHOO ET AL. Table 2. Datasets used in the present study. Grid size/Level of Data Year Data Model/Type detail Source DMSP-OLS annual NTL 2004 to 2013 Raster/Continuous 30 arc seconds NOAA-NCEI composites VIIRS-DNB monthly 2013 Raster/Continuous 15 arc seconds NOAA-NCEI composites Land cover (annual) 2004 to 2013 Raster/Categorical 500 m MODIS DEM 2000 Raster/Continuous 30 m SRTM Roads Data available as on 25 October 2018 (data contributed Vector/Categorical Scale independent OpenStreetMap during 2008–2018) Gross domestic product 2004 to 2013 Single value for the Single value for the IndiaStat (annual) State State https://ngdc.noaa.gov/eog/dmsp/downloadV4composites.html https://ngdc.noaa.gov/eog/viirs/download_dnb_composites.html https://e4ftl01.cr.usgs.gov/MOTA/MCD12Q1.006 https://earthexplorer.usgs.gov https://download.geofabrik.de/asia/india.html https://www.indiastat.com Table 3. The annual DMSP-OLS NTL datasets (version 4) are made available by NOAA-NCEI. The datasets used in the current study are highlighted in gray shade, selected based on the maximum SOL value. Satellite Year F15 F16 F18 2004 F152004 F162004 2005 F152005 F162005 2006 F152006 F162006 2007 F152007 F162007 2008 F162008 2009 F162009 2010 F182010 2011 F182011 2012 F182012 2013 F182013 11 natural vegetation classes, 3 human-induced land OpenStreetMap (Figure 2(d)). The road data used classes, and 3 non-vegetated classes. Out of these 17 were contributed to OpenStreetMap between classes, 14 classes fall in the study area (Figure 2(a)). 2008 and 2018, with nearly 70% of the data con- tributed by 2011. There are 26 types of road, 2.2.2. Digital elevation model which can be broadly classified into five major The Shuttle Radar Topography Mission (SRTM) digital classes as follows: elevation model (DEM) having a spatial resolution of 30 m is used here (Figure 2(b)). The elevation in the Class 1: Roads – Roads that consist of roads of region varies from 42 m to 888 m above mean sea higher importance such as National Highway level, with <1% area having elevation >400 m. The (NH), State Highway (SH), major city roads, and high elevation areas are limited to the northwest residential roads. (Himalayan foothills) and southeast (Bundelkhand Class 2: Link Roads – This category consists of roads plateau). The area is mainly characterized by that serve as a link between the Class 1 roads. a gentle eastward slope; 99.5% area has a <10° slope Class 3: Special Roads – Service roads, pedestrian (Figure 2(c)). The most populous cities are on the roads, tracks, and living streets fall under this banks of the rivers, namely, Ganga, Yamuna, Gomti, category, which is made to serve the public to Ghaghara, Betwa, and Son (Figure 2(b)). access a particular facility like business parks, shopping plaza, and residential areas. 2.2.3. Roads Class 4: Paths – Designated footpaths, steps, and The different types of roads in Uttar Pradesh were paths for walkers or cyclists fall under this extracted from the Road map made available by category. GISCIENCE & REMOTE SENSING 1149 Figure 2. (a) Land cover map for the year 2013; (b) SRTM DEM; (c) slope map; (d) road map (source: Geofabrik GmbH and OpenStreetMap 2018); (e) road density map; and (f) distance from roads map, of Uttar Pradesh. Class 5: Others – Other types of roads like cycle- products from DMSP-OLS annual composites. Apart way or unpaved tracks, which are of least impor- from the DMSP-OLS annual composite, three ancillary tance, fall under this category. data products (Land cover, DEM, and Road map) of Uttar Pradesh have been used as independent vari- Two derivative maps, road density (Figure 2(e) and ables for the modeling. All the datasets are resampled distance from road (Figure 2(f)) are prepared using to 500 m spatial resolution using the nearest neighbor the road map (Figure 2(d)) and used in the analysis. method and further processed for training two types of models, viz. MLP and RF. The modeled (or simulated) VIIRS-DNB annual NTL composites (also referred to as 3. Methodology VIIRS-DNB-like composites) are evaluated both qualita- The objective of this paper is to explore the viability of tively and quantitatively. The qualitative evaluation is a deep learning algorithm as a tool to produce long- carried out based on visual analysis of actual DMSP- term NTL datasets by creating VIIRS-DNB-like annual OLS, actual VIIRS-DNB, and modeled VIIRS-DNB annual 1150 S. SAHOO ET AL. NTL composites. For quantitative evaluation, scatter 3.1. Pre-processing of DMSP-OLS and VIIRS-DNB plots, dynamic range, and SOL of actual and modeled datasets products are analyzed. In addition, the statistical rela- tion between SOL and GDP of the study area is studied To correct for lack of on-board calibration and varying for actual DMSP-OLS and modeled VIIRS-DNB annual instrumental gain levels in the DMSP-OLS dataset, 2004 NTL composites for the period before 2013 (i.e., before to 2013 annual composites are inter-calibrated among the launch of VIIRS). Figure 3 shows the data and steps themselves using Ridgeline Sampling Regression (RSR) used for training and evaluating the models and pre- method, taking 2013 as reference year (Zhang, Pandey, paring the calibrated time-series VIIRS-DNB-like NTL and Seto 2016; Sahoo 2019a). The VIIRS-DNB monthly datasets from DMSP-OLS datasets. composites of 2013 (from January to December) are Figure 3. Methodology flowchart showing types of input data used and steps performed for training the model using machine- learning algorithms, model evaluation and preparation of calibrated time-series VIIRS-DNB-like NTL datasets from DMSP-OLS annual composites. GISCIENCE & REMOTE SENSING 1151 processed to prepare the annual composite (Figure 1 3.3. Selection of independent model variables (b)) using the principal component analysis (PCA) using Variable Inflation Factor method (Sahoo 2019a). The inter-calibrated DMSP- To model the relationship with the VIIRS-DNB NTL OLS annual composites of the 2004–2013 period and image (dependent variable), the selection of indepen- the VIIRS-DNB annual composite of 2013 are then used dent variables is made based on the Variable Inflation for further analysis. Factor (VIF). Variance is the squared deviation of a random variable from its mean. When multi- collinearity exists, the variance of the predictor vari- 3.2. Pre-processing of ancillary datasets able inflates. VIF is the ratio of variance in a model with multiple variables divided by the variance of the 3.2.1. Types of input and encoding of categorical model with one variable alone. It is recommended variables that if the VIF of any predictor variable is greater The input data (primary and derivative) used here than 5, and then it is highly collinear with the other essentially belong to two broad categories: predictor variables and while performing prediction ● large standard errors will be encountered Categorical Variables: Land cover; Roads ● (Akinwande, Dikko, and Samson 2015). Continuous Variables: DMSP-OLS annual NTL The variables on which the spatial growth of elec- composites; VIIRS-DNB annual NTL composites; tricity depends are varied. Some are quantifiable like DEM; Slope (derived from DEM); Distance from accessibility, slope, and stability of the land, land cover Road (derived from road map); Road Density type, resident population, distance to nearest electri- (derived from road map); Distance from each city distribution hub, etc., but some are intangible and land cover type (derived from land cover map). cannot be quantified easily such as local social, politi- cal, and religious constraints. In this study, the ancillary To use the categorical inputs, it is first converted into datasets discussed in Section 2.2 have been explored model-understandable numerical data using the label to understand the impact of each variable on NTL in encoder. The label encoder normalizes labels such that the study region. It is seen that (as discussed in Section they contain only values between 0 and n 1, where n 4.1) more than the land cover, the distance to land represents the number of classes. It also transforms the cover type (e.g., distance to built-up, distance to crop- non-numerical labels into numerical labels. After label land, etc.) is a more influencing variable. Therefore, encoding, one hot encoder is used. In one hot encoding, Euclidean distance from each of the land cover classes the column with categorical values is split into n col- has been taken as independent variables. Similarly, the umns, and the numbers are replaced by 0's and type of road is not considered as an independent vari- 1's depending upon the class. In this case, the 14 IGBP able; instead, the distance from road and road density label-encoded classes are converted into 14-bit binary has been taken as independent variables, based on the numbers. exploratory analysis of these variables (discussed in Section 4.1). VIF is then run on these selected variables 3.2.2. Data normalization to shortlist the predictor variables for model training. In general, learning algorithms benefit from a normalization of datasets as it changes the values to a uniform scale, without distorting differences in 3.4. Machine-learning algorithms used for the ranges of values. In the present study, all the inter-calibration variables are normalized between 0 and 1. The trans- Two machine-learning algorithms, viz. (1) Multi-Layer formation is given by Equation (1) below: Perceptron (MLP), a DNN architecture, and (2) Random ðx x Þ Forest (RF) are used for inter-calibration of DMSP-OLS min x ¼ (1) norm ðx x Þ and VIIRS-DNB annual NTL composites. Based on these max min two algorithms, various models are developed and are where x and x are maximum and minimum max min trained with continuous variables as well as with the values of the variable, respectively. combination of categorical and continuous variables. 1152 S. SAHOO ET AL. Variables such as distance from the road and different prediction is taken as the average of the output land cover types are used as inputs, which had higher value of all the trees. These decision trees are built weights away from the feature, and 0 at wherever the using samples drawn with replacement from the feature existed. The reverse weights of these variables training dataset. During the construction of these are also used as inputs to check whether the model is trees, the best split among a random subset of the affected by reversing the weights. Therefore, four features is selected. This randomness marginally rises types of models were built to train: (1) Categorical the bias, but due to averaging, the variance decreases, and Continuous variables, (2) Categorical and hence yielding an overall better model. Thus, the Reversed Continuous variables. (3) Continuous vari- resulting prediction is taken as an average of the ables, and (4) Reversed Continuous variables. The output value of all the trees (Breiman 2001). DMSP-OLS and land cover data are available annually. N � Hence, for predicting VIIRS-DNB annual composites, θ ðxÞ ¼ þ t ðxÞ (3) n¼1 these two datasets are changed for each year, whereas DEM and roads (primary and derivative) were taken as where θ is a random vector, N is the total number of constant. The DMSP-OLS dataset from 2004 to 2012 trees, n is the sample drawn with replacement (also was inter-calibrated among themselves keeping known as bootstrap sample), x is the input, and t is DMSP-OLS 2013 dataset as the base year (Sahoo, the individual decision tree as given in Equation (4): Gupta, and Srivastav 2019b), using Ridgeline � � � t ðxÞ ¼ t ðx; z ; . . . . . . . . . . . . . . . ; z Þ (4) n n1 nk Sampling Regression (RSR) method (Zhang, Pandey, th and Seto 2016), before using them as inputs in trained where, z (k = 1 . . . K) is the k training sample with nk machine-learning models for preparing VIIRS-DNB-like pairs of values for the target variable and covariates annual NTL composites. (x): z = (x ; y ). ni k k A brief description of both the machine-learning algorithms used in the present study is given below. 3.4.3. K-Fold cross-validation method for accuracy evaluation during model training 3.4.1. Multi-layer perceptron algorithm The general approach of calculating accuracy is to Multi-Layer Perceptron (MLP), a feed-forward neural spatially split the entire dataset into training and network, is implemented for mapping DMSP-OLS to testing (validation) sets. Then, based on error metrics, VIIRS-DNB data. The objective is to discover an accuracy is determined from the testing set, and the unknown function f as shown in Equation (2), which result entirely depends on the random choice for the relates the input vectors in X to the output vectors in pair of training and validation sets. In such cases, the Y (Gardner and Dorling 1998): accuracy may change with a change in the test set. To overcome this problem, we have used the K-Fold Y ¼ fðXÞ (2) cross-validation technique for accuracy evaluation where, X ¼ ½n� k�, Y ¼ ½n� j�, n represents the num- during model training. In the K-Fold cross-validation, ber of training patterns, k the number of input vari- the entire dataset is divided into K folds or K spatial ables, and j the number of output variables. The subsets; K number of experiments are conducted, and neural net is trained with DMSP-OLS annual compo- in each experiment (K – 1) spatial sub-section is used site and other continuous/categorical data as input for training and the left-out section is used for testing variables and VIIRS-DNB annual composite as an out- (Figure 4). This way, the model is tested to work for put variable, for the overlapping year of 2013. the entire study region. Subsequently, the resulting weighting functions are applied to the DMSP-OLS data before 2013 using the ancillary data. The functions enable the production of 4. Results and discussion annual VIIRS-DNB-like NTL datasets. 4.1. Selection of independent variables and their relation with nighttime light data 3.4.2. Random Forest algorithm Random Forest (RF) is an ensemble of decision trees The distribution of electricity in a region or the growth (Equation (3)). In RF-based regression, the final of artificial lights due to anthropogenic activities is GISCIENCE & REMOTE SENSING 1153 Figure 4. Schematic diagram showing K-Fold cross-validation method. dependent on several factors, such as land cover type, composites, the variation in SOL in DMSP-OLS data for terrain, road network, etc. Such factors been have these variables has been analyzed. utilized for inter-calibrating DMSP-OLS and VIIRS- Figure 5 shows the variation in SOL for different DNB datasets. As the purpose of the present study is land cover types. It is clear from the figure that the to create a consistent VIIRS-DNB-like annual time- maximum amount of NTL is contributed by the Urban series NTL datasets using the DMSP-OLS annual NTL and Built-up land class, followed by Open Shrublands, Figure 5. Variation in SOL with respect to land cover types. 1154 S. SAHOO ET AL. Cropland/Natural Vegetation Mosaics, and so on. It is density of road is located in major cities. Further, the obvious that as the distance from the urban center areas consisting of the least density of roads coincide increases and one moves toward croplands and rural with croplands or other natural land cover types. The areas, the SOL decreases due to less population den- variation in SOL to road density is shown in Figure 6(c). sity, few settlements, or lack of urbanization. The dis- It can be observed that the denser the road network tance from these land cover types do influence the more intense is the light. The variation in SOL with lighting mechanism; hence, the distance from each respect to distance from the road is shown in Figure 6 land cover type is taken into consideration for training (d). The SOL can be seen decreasing gradually as the the model. For this, the Euclidean distance of every distance from the road increases. cell in the raster to the nearest source is calculated. Difficult terrain often hinders the development of The variation in SOL for different types of roads is human settlements and, thus, the artificial lights. The analyzed by creating a buffer of 2 km around the roads. topography of the study area is dominantly plain, with The analysis is done on five major classes and their limited hilly terrain in the north and plateau in the south. subclasses as shown in Figure 6(a, b), respectively. It is Figure 7 represents the variation in SOL with respect to observed that roads that fall within the urban areas like the slope. As obvious, the majority of artificial lights are footways, steps, pedestrian paths, service roads contri- concentrated in very gently sloping plains. bute more toward the SOL despite having a smaller Further analysis (at finer bins for Figures 6(c, d) and length as they are surrounded by city lights and are 7) reveals a non-linear relation between SOL and road well equipped with street lighting facilities. On the con- density, distance to road, and road density. As trary, the major types of roads that consist of NH, SH and machine-learning algorithms make predictions based have lengths much higher than paths do not contribute on learning from training data without human inter- much to SOL as the main means of lighting are street- vention (Kubat 2017), such non-linear relation between lights. Moreover, as the road stretches away from cities the SOL and independent variables is largely supposed the amount of light per unit length of road decreases. to be taken care of during the model training and As no proper correlation was observed between SOL model prediction. After analyzing the relation of the and road classes, it was felt necessary to understand variables individually with artificial night-lights, the the relation between SOL and the road network as final selection of independent variables to be used for a whole. Hence, using the road map of Uttar Pradesh, training the models is made based on VIF. A total of 19 road density (Figure 2(e)) and distance from road variables were initially selected for evaluation based on (Figure 2(f)) maps were prepared for analysis. From VIF (Table 4). The iterations performed for VIF calcula- Figure 2(e), it can be observed that the maximum tion are shown in Table 4. In each iteration, the variable Figure 6. Variation in SOL with respect to (a) major road types; (b) subclasses of each major road type (the color represents the major road type as shown in Figure 6(a)); (c) road density; and (d) distance from road. GISCIENCE & REMOTE SENSING 1155 Figure 7. Variation in SOL with respect to slope. having the highest VIF value among all the variables continuous and categorical variables using the two was removed. Finally, after eight iterations, 12 variables machine-learning algorithms, MLP and RF. In the case with VIF <5 (shown in the gray shade in Table 4) were of MLP, the model parameters such as the number of found, which were then used to train the models. The hidden layers, number of nodes in each hidden layer, variables selected for model training include distance number of iterations were changed to achieve the to several land cover classes, distance to road, road highest accuracy and lowest error. The other para- density, slope, and DMSP-OLS. meters, like activation function, optimizer, regulariza- tion term, and learning rate were kept the same in all the cases. The “rectified linear unit function” was used 4.2. Machine-learning-based models for as an activation function. Stochastic gradient-based inter-calibrating DMSP-OLS and VIIRS-DNB annual optimizer “adam” was used as an optimizer. The reg- nighttime light composites ularization term and learning rate were kept as As discussed earlier, four types of models were devel- 0.00001 and 0.0001, respectively. The K-fold valida- oped and tested with different combinations of tion method was implemented. The DMSP-OLS and Table 4. VIF values of different variables from eight iterations. The final 12 variables selected for modeling are highlighted in gray shade. For brevity “Distance from” has been abbreviated as “Dist.” Iteration Number → 1 2 3 4 5 6 7 8 Classes ↓ VIF Dist. woody savannas 54.57 - - - - - - - Dist. evergreen broadleaf forest 34.80 34.30 - - - - - - Dist. mixed forest 15.81 15.81 13.48 - - - - - Elevation 25.95 25.67 11.01 10.35 - - - - Dist. evergreen needleleaf forest 14.50 12.20 9.00 8.28 7.55 - - - Dist. deciduous forest 46.20 12.72 12.42 7.74 7.53 7.14 - - Dist. barren land 6.61 6.59 6.58 6.58 6.57 6.54 6.53 - Dist. savannas 5.93 5.65 5.62 5.61 5.53 5.37 4.86 4.86 Dist. open shrublands 11.53 10.67 7.75 7.66 5.57 5.53 5.09 4.73 Dist. crop/vegetation mosaic 7.53 7.43 6.47 5.71 5.31 4.78 4.74 4.66 Dist. permanent wetlands 8.03 7.41 6.49 5.88 5.83 5.50 4.90 4.58 Dist. water bodies 9.85 9.25 8.99 8.72 7.35 6.71 3.74 3.02 Dist. grasslands 4.92 4.77 4.56 4.39 4.29 4.26 4.26 2.88 Dist. urban & built-up 3.55 3.50 3.33 3.33 3.28 2.80 2.63 2.62 Road density 2.76 2.65 2.60 2.58 2.57 2.55 2.41 2.28 Dist. road 2.29 2.28 2.27 2.27 2.27 2.23 2.20 2.20 DMSP-OLS 1.68 1.67 1.66 1.66 1.66 1.64 1.64 1.63 Slope 1.36 1.36 1.35 1.35 1.30 1.28 1.28 1.27 Dist. croplands 1.27 1.27 1.23 1.23 1.20 1.20 1.20 1.19 1156 S. SAHOO ET AL. VIIRS-DNB annual NTL composites for the All the four model configurations were tested common year (i.e., 2013) were used for developing using the RF algorithm, with different numbers of the models. The modeled (predicted) and actual trees ranging from 10 to 500. With the increase in VIIRS-DNB annual NTL composites of 2013 were com- the number of trees, minor changes occurred in R pared with each other using coefficient of determina- and RMSE, but the time for training and prediction tion (R ), and root-mean-square error (RMSE). Using increased. Finally, all the models were trained by a simple grid-based approach, the best-achieved keeping the number of trees as 200. Other para- parameter configuration for all the four types of mod- meters, viz. maximum features to select and the max- els, and the results achieved are shown in Table 5 and imum depth of trees were kept the same in all the Figure 8, respectively. cases. The results achieved are shown in Figure 9. Table 5. MLP architecture used for each model. Model Inputs Hidden Layer Configuration No. of Iterations R RMSE Categorical and Continuous variables 256, 128, 64, 32 1000 0.75 4.77 Categorical and reversed Continuous variables 256, 128, 64, 32 1000 0.78 5.20 Continuous variables 512, 256, 128, 64, 32 1000 0.86 3.78 Reversed continuous variables 512, 256, 128, 64,32 1000 0.77 4.89 Figure 8. Density scatter plots between actual and predicted VIIRS-DNB annual NTL composites for the year 2013 using MLP model. The different configurations of MLP model are (a) categorical and continuous inputs, (b) categorical and reversed continuous inputs, −2 −1 (c) continuous inputs and (d) reversed continuous inputs. All quantities in the graphs are in nW cm sr . GISCIENCE & REMOTE SENSING 1157 Figure 9. Density scatter plots between actual and predicted VIIRS-DNB annual NTL composites for the year 2013 using RF model. The different configurations of RF model are (a) categorical and continuous inputs, (b) categorical and reversed continuous inputs, (c) −2 −1 continuous inputs and (d) reversed continuous inputs. All quantities in the graphs are in nW cm sr . The results obtained through MLP and RF-based strengthened by visually analyzing the areas around models are evaluated both qualitatively and quantita- select cities. tively and are elaborated in Section 4.2.1. Figure 10 shows the actual DMSP-OLS, actual VIIRS- DNB, the RF, and MLP model predicted VIIRS-DNB annual NTL composites around Lucknow, Agra, and 4.2.1. Evaluation of predicted VIIRS-DNB annual Prayagraj cities [locations highlighted in Figure 11 a(i)] nighttime light composites for the year 2013. The following observations are The scatter plots of actual versus predicted VIIRS-DNB made based on a detailed visual analysis of the mod- NTL radiance values (for the year 2013) obtained eled products. First, the predicted VIIRS-DNB NTL using the MLP-based models (Figure 8) show images show more spatial granularity than the actual a larger spread of values across the 1:1 line, lower R DMSP-OLS data, thereby signifying that better spatial values (0.75–0.86), and higher RMSE values (3.75–- night-light details can be derived by using advanced 5.20) as compared to RF-based models (Figure 9; machine-learning algorithms and ancillary data. R = 0.87–0.89; RMSE = 2.12–2.32). Further, the MLP Second, both the RF and MLP models successfully and RF models with continuous variables as inputs remove the blooming effect present in the actual performed better than the remaining models. The DMSP-OLS data. Third, the MLP outputs (Figure 10, pixel-wise spatial comparison shows that the RF- column d) are comparatively smoother, which leads based models yield better results as compared to to loss of spatial details. Fourth, the RF outputs are MLP-based models. This observation is further 1158 S. SAHOO ET AL. Figure 10. Actual and predicted annual NTL composites around select cities in the study area for the year 2013: (a) actual DMSP-OLS, (b) actual VIIRS-DNB, (c) predicted VIIRS-DNB (using continuous input RF model), and (d) predicted VIIRS-DNB (using continuous input MLP model) around the cities of (i) Lucknow, (ii) Agra and (iii) Prayagraj. All images are at same spatial scale, as shown in the bottom right image. relatively brighter (higher radiances) in regions of low improvement in the RF and MLP model performances lighting. The roads and smaller towns with low light- by changing the model parameters. Therefore, using ing are particularly apparent on the RF outputs, often the best model achieved in both the cases, the VIIRS- not visible on the MLP outputs. Fifth, the saturated DNB annual NTL composites are generated for the pixels in the DMSP-OLS data in the center of major years 2004 to 2012. The details are provided in the cities [e.g., Lucknow in Figure 10d(i) and Kanpur (not next section. shown)] have relatively smaller radiance values in the MLP-based models, which is not the case in the RF- 4.2.2. Generation of time-series VIIRS-DNB-like NTL based model outputs. In other words, the MLP-based datasets models are not able to adequately capture the urban Using the trained MLP and RF models, the VIIRS-DNB- centers of the major cities where DMSP-OLS data are like NTL images are generated for the years 2004–2012. saturated. Sixth, a significant reduction in radiance In doing so, the DMSP-OLS and land cover (i.e., distance values with increasing distance from the urban cen- to individual land cover types) inputs are varied corre- ters is observed in the RF model-derived products, spondingly for each year, while keeping the other whereas the MLP products show a gradual decrease independent variables (elevation and its derivatives, in radiance values away from the urban centers. Given road derivatives) unchanged with time. The represen- these observations, the RF model (Figure 10, column tative outputs of the study area for 3 years (2004, 2008, c) performs better as it emulates the VIIRS-DNB data and 2012) based on RF and MLP models are shown in (Figure 10, column b) more closely. Figures 11 and 12, respectively. The gradual intensifica - As mentioned earlier, the RF and MLP models with tion of light and lighting of new towns with time are continuous variables as inputs performed better. the visible characteristics of the time-series data gen- Moreover, it is observed that there was no significant erated. The RF model-based outputs indicate that the GISCIENCE & REMOTE SENSING 1159 Figure 11. Predicted VIIRS-DNB-like annual NTL composites for select cities, using continuous input RF model. Columns show different years (a) 2004, (b) 2008 and (c) 2012. Rows show (i) entire study area (Indian state of Uttar Pradesh), (ii) Lucknow, (iii) Agra, and (iv) Prayagraj cities. The SOL for each of the cities is calculated by taking a buffer of 0.4° x 0.4° around each city. Units of SOL and image −2 −1 legends are in nW cm sr . All city images are at same spatial scale, as shown in (a-iv). SOL values increased by 7.2% and 9.5% during 2004 to composites of 2013 are also relevant for the 2008 and 2008 to 2012, respectively, whereas MLP 2004–2012 composites. Particularly, it is evident that outputs indicate the increase in SOL values by 4.9% the RF model can better handle the problem of satu- and 30.3% for the corresponding period. It is important rated urban cores, which the input dataset DMSP-OLS to note that in Figures 11 and 12, the same stretch is plagued with. The MLP model, on the other hand, is values have been used for each row, i.e., row (i) Uttar not able to capture the saturated urban cores. Pradesh is stretched from 0 to 25, row (ii) Lucknow is The accuracy of predicted VIIRS-DNB annual NTL stretched from 0 to 200, and so on. This has been done composites of 2004–2012 was also tested on for easy visual comparison of the two models. The a regional scale by studying the statistical relation observations made in Section 4.2.1 based on visual between SOL and GDP of the State (Figure 13). 2 2 comparison of actual and modeled VIIRS-DNB NTL Notably, the R values of MLP (R = 0.93) and RF 1160 S. SAHOO ET AL. Figure 12. Predicted VIIRS-DNB-like annual NTL composites for select cities, using continuous input MLP model. Columns show different years (a) 2004, (b) 2008 and (c) 2012. Rows show (i) entire study area (Indian state of Uttar Pradesh), (ii) Lucknow, (iii) Agra, and (iv) Prayagraj cities. The SOL for each of the cities is calculated by taking a buffer of 0.4° x 0.4° around each city. Units of SOL and −2 −1 image legends are in nW cm sr . All city images are at same spatial scale, as shown in (a-iv). (R = 0.66) model outputs are higher than that for the study area are significantly higher in the case of RF. actual DMSP-OLS products (R = 0.60). It is, thus, evi- Moreover, RF model outputs show a larger variability in dent that machine-learning algorithms can be signifi - SOL with time as compared to MLP outputs, which cantly utilized for inter-calibrating DMSP-OLS and caused lower R value between RF-derived SOL VIIRS-DNB datasets to create consistent annual time- and GDP. series NTL products for longer periods. Further, the The modeled VIIRS-DNB-like NTL datasets for 2004 MLP model performs better at a regional scale, as to 2012 using both the RF and MLP models are made compared to RF. Detailed analysis indicates that while available for download by the researchers from the the maximum radiance values in the modeled products following link (https://doi.org/10.5281/zenodo. from both the models are similar, the SOL values in the 3967104). GISCIENCE & REMOTE SENSING 1161 Figure 13. Relation between GDP and SOL derived from (a) predicted VIIRS-DNB annual NTL composites using RF algorithm, (b) predicted VIIRS-DNB annual NTL composites using MLP algorithm, and (c) for actual DMSP-OLS annual NTL composites, for the period 2004–2012. 4.3. Discussion a prediction. On the other hand, the MLP model The qualitative and quantitative analysis of the RF and takes time to train data, but the prediction is very MLP model outputs have shown particular merits and fast. Therefore, holistically, both the models have limitations of each model. The R and RMSE values of their strong areas of applications: MLP is recom- the model predicted outputs for the reference year mended for regional-scale studies, while RF is recom- 2013 and the visual analysis of predicted temporal mended for local-scale studies. VIIRS-DNB-like NTL images indicated that RF model The DMSP-OLS products have been used for more performance is better as far as capturing the spatial than two decades as a proxy for monitoring socio- details at the local-scale is concerned. Both models economic activity and human well-being at National are able to significantly reduce the blooming effect, and Global levels (Chen and Nordhaus 2011; Ghosh seen in the actual DMSP-OLS data. On the other hand, et al. 2013). In this regard, to accurately study the the temporal VIIRS-DNB-like annual NTL products spatio-temporal trend of artificial night-lights from generated from the past DMSP-OLS data using the the time when VIIRS-DNB did not exist, machine learn- MLP model show a better correlation with economic ing can be used to simulate datasets and study histor- development. This indicates that the MLP provides ical aspects at a much better resolution compared to stable outputs in terms of SOL of the study area as then available DMSP-OLS data. a whole, but it fails to handle the saturated city cores. The proposed approach of inter-calibration using Further, the RF model is easy to implement and machine-learning architectures will not only help to train when the number of trees is kept small. For perform the long-term analysis of NTL at a regional better predictive power, the number of trees needs scale through the combined study of DMSP-OLS and to be increased, which increases the time for recently available VIIRS-DNB datasets but will also 1162 S. SAHOO ET AL. provide improved results and understanding of socio- environment like an increase in air pollution due to economic development that would have been industrialization and urbanization. obtained using standalone DMSP-OLS and VIIRS-DNB datasets. In the present study, among the ancillary 5. Conclusions spatial datasets, only land cover is used as a time- variant variable along with DMSP-OLS datasets. The availability of NTL composites from the VIIRS-DNB Incorporation of the temporal information on the opened the scope for inter-calibrating them with DMSP- roads and other infrastructure in the model is likely OLS and create consistent annual time-series NTL data- to improve the inter-calibration results further. The sets for the long-term studies. The potential of two proposed methods for inter-calibration can also be machine-learning algorithms, MLP (having a DNN archi- tested using other global DEMs to study the improve- tecture) and RF, is evaluated in this study for inter- ment in results if any. The proposed framework was calibration of the DMSP-OLS and VIIRS-DNB annual found to work well over a large State of Uttar Pradesh, NTL composites. The annual NTL composites of DMSP- which is primarily a flat, tropical, agrarian, and densely OLS and VIIRS-DNB for the common year of 2013 along populated area. Researchers can explore the frame- with ancillary datasets on land cover, road network, and work further over other sites with different topogra- topography are used to train the MLP and RF models. phy, climate, and socio-economic setting. The best model configuration is then used to simulate The annual availability of both DMSP-OLS and VIIRS- VIIRS-DNB-like images from DMSP-OLS annual NTL com- DNB is limited to only 1 year (i.e., 2013). This poses posites using both MLP and RF algorithms. Detailed a unique challenge in inter-calibrating the two datasets. analysis of the outputs revealed that: (1) Both MLP and Recent studies (Shao et al. 2014; Li et al. 2017a; Zheng, RF models significantly reduce the blooming effect Weng, and Wang 2019; Li et al. 2020; Ma et al. 2020) around settlements, a common problem observed in have attempted to overcome this issue and generated DMSP-OLS data; (2) visual appearance and spatial gran- DMSP-OLS-like products after 2013, whereas the present ularity of the model predicted outputs are close to the study leverages the capabilities of machine learning to actual VIIRS-DNB images; (3) RF captures better spatial generate VIIRS-DNB-like products before 2013. It is also details at local-scale and is able to efficiently handle the important to note that the local time of over-pass of the saturation problem at urban centers, which is again two satellites is different (DMSP-OLS: 1900–2200 hours a common issue with DMSP-OLS data; (4) MLP, on the (Elvidge et al. 2009); VIIRS-DNB: ~0130 hours) and, there- other hand, is found to be superior at regional-scale as it fore, its impact on the inter-calibration of the two data- provided better statistical relation between model- sets needs to be further explored. Earlier studies (Li et al. derived SOL of and GDP of the study region. It is, thus, 2017a, 2020; Jeswani et al. 2019; Ma et al. 2020; Zheng, concluded that both the machine-learning algorithms Weng, and Wang 2019) have also noted this as could be appropriately utilized for producing VIIRS-DNB- a limitation and proceeded for further analysis. The end- like NTL images from DMSP-OLS annual NTL compo- of-life sensor degradation of DMSP-OLS and the lead sites, depending on the purpose of the study. The mod- time required for calibration and testing of the VIIRS- eled VIIRS-DNB-like annual NTL composites for 9 years DNB datasets are the other important factors for the (2004 to 2012) generated using trained RF and MLP inter-calibration of the two datasets, which may also models are made available for download by the be explored by the researchers in the future. researchers from the following link (https://doi.org/10. The potential of various other DNN architectures 5281/zenodo.3967104). such as Recurrent Neural Network, Convolutional The utilization of RF and MLP is, thus, proposed for Neural network, etc. can also be investigated in the inter-calibration of DMSP-OLS and VIIRS-DNB annual inter-calibration of NTL datasets. The derived annual NTL datasets at local and regional scales, respectively, time-series NTL data can be used to study various for long-term studies. The long-term NTL datasets are factors that hinder or accelerate the growth of artifi - envisaged to have critical applications in monitoring cial lights in a region. The datasets can also be used to socio-economic development, disaster impact, and study the impact of the growth of lights on the land cover change. The study can be further expanded GISCIENCE & REMOTE SENSING 1163 by using other machine-learning algorithms and Brown, M. E., D. J. Lary, A. Vrieling, D. Stathakis, and H. Mussa. 2008. “Neural Networks as a Tool for Constructing improved spatio-temporal ancillary datasets over Continuous NDVI Time Series from AVHRR and MODIS.” areas with different socio-economic, physiographic, International Journal of Remote Sensing 29 (24): 7141–7158. and climate setting. doi:10.1080/01431160802238435. Cao, S., D. Hu, W. Zhao, Y. Mo, Y. Chen, and Y. Zhang. 2019. “Monitoring Changes in the Impervious Surfaces of Urban Acknowledgements Functional Zones Using Multisource Remote Sensing Data: A Case Study of Tianjin, China.” GIScience & Remote Sensing This study is a part of M.Tech. Dissertation carried out by the 56 (7): 967–987. doi:10.1080/15481603.2019.1600110.s. first author. Thanks to the Director, Indian Institute of Remote Chander, G., T. J. Hewison, N. Fox, X. Wu, X. Xiong, and Sensing, Indian Space Research Organisation (ISRO), Dehradun, W. J. Blackwell. 2013. “Overview of Intercalibration of Satellite India for his support and encouragement. Sincere thanks to Instruments.” IEEE Transactions on Geoscience and Remote National Oceanic and Atmospheric Administration’s National Sensing 51 (3): 1056–1080. doi:10.1109/TGRS.2012.2228654. Centres (NOAA) for providing the DMSP-OLS and VIIRS-NNB Chen, X., and W. D. Nordhaus. 2011. “Using Luminosity Data as nighttime lights data. We thank the anonymous reviewers for a Proxy for Economic Statistics.” Proceedings of the National their useful and critical comments that helped to improve the Academy of Sciences 108 (21): 8589–8594. doi:10.1073/ manuscript to a large extent. pnas.1017031108. Chen, Z., B. Yu, Y. Hu, C. Huang, K. Shi, and J. Wu. 2015. “Estimating House Vacancy Rate in Metropolitan Areas Disclosure statement Using NPP-VIIRS Nighttime Light Composite Data.” IEEE No potential conflict of interest was reported by the authors. Journal of Selected Topics in Applied Earth Observations and Remote Sensing 8 (5): 2188–2197. doi:10.1109/ JSTARS.2015.2418201. ORCID Cracknell, M. J., and A. M. Reading. 2014. “Geological Mapping Using Remote Sensing Data: A Comparison of Five Machine Sumana Sahoo http://orcid.org/0000-0003-3655-9415 Learning Algorithms, Their Response to Variations in the Prasun Kumar Gupta http://orcid.org/0000-0003-0689-4167 Spatial Distribution of Training Data and the Use of Explicit Spatial Information.” Computers & Geosciences 63: 22–33. doi:10.1016/j.cageo.2013.10.008. Author responsibilities Duriscoe, D. M., C. B. Luginbuhl, and C. D. Elvidge. 2013. “The Relation of Outdoor Lighting Characteristics to Sky Glow Sumana Sahoo was responsible for performing the experi- from Distant Cities.” Lighting Research and Technology 46 ments, writing computer programs, data analysis, and prepar- (1): 35–49. doi:10.1177/1477153513506729. ing the first draft of the manuscript. Elvidge, C. D., K. E. Baugh, E. A. Kihn, H. W. Kroehl, E. R. Davis, Prasun Kumar Gupta conceived and designed the experi- and C. W. Davis. 1997. “Relation between Satellite Observed ments, and was responsible for supervising the computer pro- Visible-near Infrared Emissions, Population, Economic gramming, checking of data analysis results, and revising the Activity and Electric Power Consumption.” International manuscript. Journal of Remote Sensing 18 (6): 1373–1379. doi:10.1080/ Dr. S. K. Srivastav was responsible for conceiving the pro- blem, supervising and improving the study, and revising the Elvidge, C. D., D. Ziskin, K. E. Baugh, B. T. Tuttle, T. Ghosh, D. W. Pack, manuscript. E. H. Erwin, and M. Zhizhin. 2009. “A Fifteen Year Record of Global Natural Gas Flaring Derived from Satellite Data.” Energies 2 (3): 595–622. doi:10.3390/en20300595. References Falchi, F., P. Cinzano, D. Duriscoe, C. C. M. Kyba, C. D. Elvidge, Akinwande, M. O., H. G. Dikko, and A. Samson. 2015. “Variance K. Baugh, B. A. Portnov, et al. 2016. “The New World Atlas of Inflation Factor: As a Condition for the Inclusion of Artificial Night Sky Brightness.” Science Advances 2 (6): Suppressor Variable(s) in Regression Analysis.” Open e1600377. doi:10.1126/sciadv.1600377. Journal of Statistics 5 (7): 754–767. doi:10.4236/ Fu, Y., J. Li, Q. Weng, Q. Zheng, L. Li, S. Dai, and B. Guo. 2019. ojs.2015.57075. “Characterizing the Spatial Pattern of Annual Urban Growth by Bennett, M. M., and L. C. Smith. 2017. “Advances in Using Using Time Series Landsat Imagery.” Science of the Total Multitemporal Night-time Lights Satellite Imagery to Environment 666: 274–284. doi:10.1016/j.scitotenv.2019.02.178. Detect, Estimate, and Monitor Socioeconomic Dynamics.” Gardner, M. W., and S. R. Dorling. 1998. “Artificial Neural Remote Sensing of Environment 192: 176–197. doi:10.1016/j. Networks (The Multilayer Perceptron) - a Review of rse.2017.01.005. Applications in the Atmospheric Sciences.” Atmospheric Breiman, L. 2001. “Random Forests.” Machine Learning 45 (1): Environment 32 (14/15): 2627–2636. doi:10.1016/S1352- 5–32. doi:10.1023/A:1010933404324. 2310(97)00447-0. 1164 S. SAHOO ET AL. Geofabrik GmbH and OpenStreetMap. 2018. “India”. Accessed Syrian Civil War.” International Journal of Remote Sensing 38 26 October 2018. https://download.geofabrik.de/asia/india. (21): 5934–5951. doi:10.1080/01431161.2017.1331476. html Li, X., R. Zhang, C. Huang, and D. Li. 2015. “Detecting 2014 Ghosh, T., S. J. Anderson, C. D. Elvidge, and P. C. Sutton. 2013. Northern Iraq Insurgency Using Night-time Light Imagery.” “Using Nighttime Satellite Imagery as a Proxy Measure of International Journal of Remote Sensing 36 (13): 3446–3458. Human Well-being.” Sustainability 5 (12): 4988–5019. doi:10.1080/01431161.2015.1059968. doi:10.3390/su5124988. Li, X., and Y. Zhou. 2017b. “A Stepwise Calibration of Global Gumma, M. K., P. S. Thenkabail, P. G. Teluguntla, A. Oliphant, DMSP/OLS Stable Nighttime Light Data (1992-2013).” J. Xiong, C. Giri, V. Pyla, S. Dixit, and A. M. Whitbread. 2020. Remote Sensing 9 (6): 637. doi:10.3390/rs9060637. “Agricultural Cropland Extent and Areas of South Asia Li, X., Y. Zhou, M. Zhao, and X. Zhao. 2020. “A Harmonized Derived Using Landsat Satellite 30-m Time-series Big-data Global Nighttime Light Dataset 1992-2018.” Scientific Data 7 Using Random Forest Machine Learning Algorithms on the (1): 1–9. doi:10.1038/s41597-020-0510-y. Google Earth Engine Cloud.” GIScience & Remote Sensing 57 Liu, Z., C. He, Q. Zhang, Q. Huang, and Y. Yang. 2012. (3): 302–322. doi:10.1080/15481603.2019.1690780. “Extracting the Dynamics of Urban Expansion in China Guo, W., G. Li, W. Ni, Y. Zhang, and D. Li. 2018. “Exploring Using DMSP-OLS Nighttime Light Data from 1992 to 2008.” Improvement of Impervious Surface Estimation at National Landscape and Urban Planning 106 (1): 62–72. doi:10.1016/j. Scale through Integration of Nighttime Light and Proba-V landurbplan.2012.02.013. Data.” GIScience & Remote Sensing 55 (5): 699–717. Ma, J., J. Guo, S. Ahmad, Z. Li, and J. Hong. 2020. “Constructing doi:10.1080/15481603.2018.1436425. a New Inter-Calibration Method for DMSP-OLS and Jawarneh, R. N., J. P. Julian, and T. R. Lookingbill. 2015. “The NPP-VIIRS Nighttime Light.” Remote Sensing 12 (6): 937. Influence of Physiography on Historical and Future Land doi:10.3390/rs12060937. Development Changes: A Case Study of Central Arkansas Ma, T., C. Zhou, T. Pei, S. Haynie, and J. Fan. 2012. “Quantitative (USA), 1857–2030.” Landscape and Urban Planning 143: Estimation of Urbanization Dynamics Using Time Series of 76–89. doi:10.1016/j.landurbplan.2015.06.010. DMSP/OLS Nighttime Light Data: A Comparative Case Study Jeswani, R., A. Kulshrestha, P. K. Gupta, and S. K. Srivastav. 2019. from China’s Cities.” Remote Sensing of Environment 124: “Evaluation of the Consistency of DMSP-OLS and SNPP-VIIRS 99–107. doi:10.1016/j.rse.2012.04.018. Night-time Light Datasets.” Journal of Geomatics 13 (1): Min, B., and K. M. Gaba. 2014. “Tracking Electrification in 98–105. Vietnam Using Nighttime Lights.” Remote Sensing 6 (10): Kiran Chand, T. R., K. V. S. Badarinath, C. D. Elvidge, and B. T. Tuttle. 9511–9529. doi:10.3390/rs6109511. 2009. “Spatial Characterization of Electrical Power Min, B., K. M. Gaba, O. F. Sarr, and A. Agalassou. 2013. Consumption Patterns over India Using Temporal DMSP-OLS “Detection of Rural Electrification in Africa Using Night-time Satellite Data.” International Journal of Remote DMSP-OLS Night Lights Imagery.” International Journal of Sensing 30 (3): 647–661. doi:10.1080/01431160802345685. Remote Sensing 34 (22): 8118–8141. doi:10.1080/ Kubat, M. 2017. An Introduction to Machine Learning. Springer 01431161.2013.833358. International Publishing AG. doi:10.1007/978-3-319-63913-0 Mukherjee, S., S. K. Srivastav, P. K. Gupta, N. A. S. Hamm, and Kühnlein, M., T. Appelhans, B. Thies, and T. Nauss. 2014. “Improving V. A. Tolpekin. 2017. “An Algorithm for Inter-calibration of the Accuracy of Rainfall Rates from Optical Satellite Sensors Time-series DMSP/OLS Night-time Light Images.” with Machine Learning - a Random Forests-based Approach Proceedings of the National Academy of Sciences India Applied to MSG SEVIRI.” Remote Sensing of Environment 141: Section A - Physical Sciences 87 (4): 721–731. doi:10.1007/ 129–143. doi:10.1016/j.rse.2013.10.026. s40010-017-0444-8. Kwiatkowska, E. J., and G. S. Fargion. 2003. “Application of NOAA-NCEI. 2018. “DMSP & VIIRS Data Download.” Accessed 15 Machine-learning Techniques toward the Creation of July 2018. https://ngdc.noaa.gov/eog/download.html a Consistent and Calibrated Global Chlorophyll Concentration Odongo, V. O., N. A. S. Hamm, and E. J. Milton. 2014. “Spatio- Baseline Dataset Using Remotely Sensed Ocean Color Data.” temporal Assessment of Tuz Gölü, Turkey as a Potential IEEE Transactions on Geoscience and Remote Sensing 41 (12): Radiometric Vicarious Calibration Site.” Remote Sensing 6 2844–2860. doi:10.1109/TGRS.2003.818016. (3): 2494–2513. doi:10.3390/rs6032494. Levin, N., S. Ali, and D. Crandall. 2018. “Utilizing Remote REC Limited. 2018. “Rural Electrification Corporation.” Accessed Sensing and Big Data to Quantify Conflict Intensity: The 1 January 2019. https://www.recindia.nic.in/ Arab Spring as a Case Study.” Applied Geography 94: 1–17. Rodriguez-Galiano, V., M. Sanchez-Castillo, M. Chica-Olmo, and doi:10.1016/j.apgeog.2018.03.001. M. Chica-Rivas. 2015. “Machine Learning Predictive Models Levin, N., and Y. Duke. 2012. “High Spatial Resolution Night-time for Mineral Prospectivity: An Evaluation of Neural Networks, Light Images for Demographic and Socio-economic Studies.” Random Forest, Regression Trees and Support Vector Remote Sensing of Environment 119: 1–10. doi:10.1016/j. Machines.” Ore Geology Reviews 71: 804–818. doi:10.1016/j. rse.2011.12.005. oregeorev.2015.01.001. Li, X., D. Li, H. Xu, and C. Wu. 2017a. “Intercalibration between Sahoo, S. 2019a. “Evaluating the Potential of Night-time Light DMSP/OLS and VIIRS Night-time Light Images to Evaluate City Data to Study Electrification in Uttar Pradesh.” M.Tech. the- Light Dynamics of Syria’s Major Human Settlement during sis, IIRS/Andhra University, Viskhapatnam, India GISCIENCE & REMOTE SENSING 1165 Sahoo, S., P. K. Gupta, and S. K. Srivastav. 2019b. “Comparative Yu, B., M. Tang, Q. Wu, C. Yang, K. Shi, C. Peng, J. Wu, and Analysis between VIIRS-DNB and DMSP-OLS Night-time Light Z. Chen. 2018. “Urban Built-Up Area Extraction From Data to Estimate Electric Power Consumption in Uttar Pradesh, Log-Transformed NPP-VIIRS Nighttime Light Composite India.” International Journal of Remote Sensing 41 (7): Data.” IEEE Transactions on Geoscience and Remote 2565–2580. doi:10.1080/01431161.2019.1693077. Sensing Letters 15 (8): 1279–1283. doi:10.1109/ Shao, X., C. Cao, B. Zhang, S. Qiu, C. Elvidge, and M. V. Hendy. 2014. LGRS.2018.2830797. “Radiometric Calibration of DMSP-OLS Sensor Using VIIRS Day/ Zhang, Q., B. Pandey, and K. C. Seto. 2016. “A Robust Method to night Band.” In Proceedings of SPIE Vol. 9264, Earth Observing Generate A Consistent Time Series from DMSP/OLS Missions and Sensors: Development, Implementation, and Nighttime Light Data.” IEEE Transactions on Geoscience and Characterization III. doi: 10.1117/12.2068999. Remote Sensing 54 (10): 5821–5831. doi:10.1109/ Shi, K., B. Yu, Y. Hu, C. Huang, Y. Chen, Y. Huang, Z. Chen, and TGRS.2016.2572724. J. Wu. 2015. “Modeling and Mapping Total Freight Traffic in Zhang, Q., C. Schaaf, and K. C. Seto. 2013. “The Vegetation China Using NPP-VIIRS Nighttime Light Composite Data.” Adjusted NTL Urban Index: A New Approach to Reduce GIScience & Remote Sensing52 (3): 274–289. doi:10.1080/ Saturation and Increase Variation in Nighttime Luminosity.” 15481603.2015.1022420. Remote Sensing of Environment 129: 32–41. doi:10.1016/j. Siddiqui, A., A. Siddiqui, S. Maithani, A. K. Jha, P. Kumar, and rse.2012.10.022. S. K. Srivastav. 2018. “Urban Growth Dynamics of an Indian Zhao, N., T. Ghosh, and E. L. Samson. 2012. “Mapping Metropolitan Using CA Markov and Logistic Regression.” The Spatio-temporal Changes of Chinese Electric Power Egyptian Journal of Remote Sensing and Space Science 21 (3): Consumption Using Night-time Imagery.” International 229–236. doi:10.1016/j.ejrs.2017.11.006. Journal of Remote Sensing 33 (20): 6304–6320. doi:10.1080/ Townsend, A. C., and D. A. Bruce. 2010. “The Use of Night-time 01431161.2012.684076. Lights Satellite Imagery as a Measure of Australia’s Regional Zhao, N., F.-C. Hsu, G. Cao, and E. L. Samson. 2017. “Improving Electricity Consumption and Population Distribution.” Accuracy of Economic Estimations with VIIRS DNB Image International Journal of Remote Sensing 31 (16): 4459–4480. Products.” International Journal of Remote Sensing 38 (21): doi:10.1080/01431160903261005. 5899–5918. doi:10.1080/01431161.2017.1331060. Uttar Pradesh Government. 2011. “Social Demography of Uttar Zhao, N., Y. Zhou, and E. L. Samson. 2015. “Correcting Pradesh.” Accessed 21 June 2019. http://up.gov.in/Social- Incompatible DN Values and Geometric Errors in Nighttime Demography.pdf Lights Time-series Images.” IEEE Transactions on Geoscience Witmer, F. D. W. 2015. “Remote Sensing of Violent Conflict: Eyes and Remote Sensing 53 (4): 2039–2049. doi:10.1109/ from Above.” International Journal of Remote Sensing 36 (9): TGRS.2014.2352598. 2326–2352. doi:10.1080/01431161.2015.1035412. Zheng, Q., Q. Weng, and K. Wang. 2019. “Developing a New Wu, J., S. He, J. Peng, W. Li, and X. Zhong. 2013. “Intercalibration Cross-sensor Calibration Model for DMSP-OLS and of DMSP-OLS Night-time Light Data by the Invariant Region Suomi-NPP VIIRS Night-light Imageries.” ISPRS Journal of Method.” International Journal of Remote Sensing 34 (20): Photogrammetry and Remote Sensing 153: 36–47. 7356–7368. doi:10.1080/01431161.2013.820365. doi:10.1016/j.isprsjprs.2019.04.019. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png GIScience & Remote Sensing Taylor & Francis

Inter-calibration of DMSP-OLS and SNPP-VIIRS-DNB annual nighttime light composites using machine learning

Loading next page...
 
/lp/taylor-francis/inter-calibration-of-dmsp-ols-and-snpp-viirs-dnb-annual-nighttime-8LZm6z0MuK

References (53)

Publisher
Taylor & Francis
Copyright
© 2020 Informa UK Limited, trading as Taylor & Francis Group
ISSN
1943-7226
eISSN
1548-1603
DOI
10.1080/15481603.2020.1848323
Publisher site
See Article on Publisher Site

Abstract

GISCIENCE & REMOTE SENSING 2020, VOL. 57, NO. 8, 1144–1165 https://doi.org/10.1080/15481603.2020.1848323 Inter-calibration of DMSP-OLS and SNPP-VIIRS-DNB annual nighttime light composites using machine learning Sumana Sahoo , Prasun Kumar Gupta and S. K. Srivastav Indian Institute of Remote Sensing, Dehradun, India ABSTRACT ARTICLE HISTORY Received 22 April 2020 The satellite-based nighttime lights (NTL) data from the Defense Meteorological Satellite Program’s Accepted 3 November 2020 Operational Linescan System (DMSP-OLS), available in the public domain from 1992 to 2013, are extensively used for socio-economic studies. The improved NTL products from the Visible Infrared KEYWORDS Imaging Radiometer Suite’s Day/Night Band (VIIRS-DNB), on-board the Suomi National Polar-Orbiting VIIRS-DNB; DMSP-OLS; inter- Partnership spacecraft and National Oceanic and Atmospheric Administration – 20 (NOAA-20) space- calibration; machine craft’s, are now available since April 2012. This study investigates the potential of machine-learning learning; multi-layer algorithms for inter-calibrating them (i.e., DMSP-OLS and VIIRS-DNB) to produce time-series annual perceptron; random forest VIIRS-DNB-like NTL datasets for the time when VIIRS-DNB data did not exist, for long-term studies. Uttar Pradesh, one of the most populous and largest States of India, is selected as the study area. Two machine-learning algorithms are utilized: (1) Multi-Layer Perceptron (MLP), having deep neural networks (DNN) architecture, and (2) Random Forest (RF), a widely used method. The DMSP-OLS and VIIRS-DNB data of 2013 (common year of data availability) and ancillary data pertaining to land cover, topography, and road network are used to train the models. The qualitative and quantitative analysis of annual VIIRS-DNB-like NTL images simulated from annual DMSP-OLS composites of 2004–2012 indicates that RF captures better spatial details at the local-scale and is able to efficiently handle the saturation problem at urban centers; while MLP is found to be superior at regional-scale. Both MLP and RF models significantly reduce the blooming effect around settlements, a common problem observed in DMSP-OLS data. It is inferred that depending on the research objectives, both RF and MLP algorithms can be appropriately utilized for producing VIIRS-DNB-like NTL images from DMSP-OLS annual NTL composites. The research can be further expanded by using other DNN architecture-based algorithms and improved spatio-temporal ancillary datasets over areas with different socio-economic, physiographic, and climatic settings. 1. Introduction Administration – 20 (NOAA-20) spacecraft’s, over- For almost two decades, various research works have comes some of the limitations of the DMSP-OLS, involved the use of the Defense Meteorological mainly by providing calibrated NTL data at improved Satellite Program’s Operational Linescan System spatial and radiometric resolutions (Table 1). The (DMSP-OLS) nighttime lights (NTL) time-series data SNPP-VIIRS-DNB (hereafter referred to as VIIRS-DNB) to track and monitor the growth of electrified settle- data made available from April 2012 onwards by the ments (Elvidge et al. 1997; Kiran Chand et al. 2009; Earth Observations Group (EOG) at NOAA’s National Townsend and Bruce 2010; Levin and Duke 2012; Ma Centers for Environmental Information (NOAA-NCEI) et al. 2012; Zhao, Ghosh, and Samson 2012; Min et al. in the form of monthly NTL composites are being 2013; Min and Gaba 2014; Cao et al. 2019). Although widely used by the researchers for different applica- these DMSP-OLS annual NTL composites have helped tions such as, socio-economic dynamics (Shi et al. study social, demographical, and economic temporal 2015; Bennett and Smith 2017; Zhao et al. 2017), dynamics, the data consists of certain technical limita- urban dynamics (Chen et al. 2015; Guo et al. 2018; tions such as lack of calibration, saturation, and Yu et al. 2018), light pollution (Duriscoe, Luginbuhl, blooming effect. The Visible Infrared Imaging and Elvidge 2013; Falchi et al. 2016), military conflicts Radiometer Suite’s Day/Night Band (VIIRS-DNB), on- (Li et al. 2015; Witmer 2015; Levin, Ali, and Crandall board the Suomi National Polar-Orbiting Partnership 2018), etc. The characteristic differences between (SNPP), and National Oceanic and Atmospheric DMSP-OLS and VIIRS-DNB products are shown in CONTACT Prasun Kumar Gupta prasun@iirs.gov.in © 2020 Informa UK Limited, trading as Taylor & Francis Group GISCIENCE & REMOTE SENSING 1145 Table 1. Characteristic differences between DMSP-OLS and Inter-calibration of multi-satellite data is crucial for VIIRS-DNB data. detecting and quantifying the changes in the Earth’s DMSP-OLS VIIRS-DNB environment, for predicting weather conditions, for Spatial resolution 2.7 km 742 m understanding various climatic processes, and monitor- Radiometric quantization 6 bit 14 bit Detection range Limited Large ing land cover changes (Chander et al. 2013). Similarly, On-board calibration Not available Available Saturation problem Present at urban centers Eliminated the inter-calibration between DMSP-OLS and VIIRS-DNB Blooming effect Present Eliminated NTL datasets can help in producing consistent time- series products, which will be useful for long-term ana- lysis of various socio-economic factors. Efforts were Table 1. Research (Bennett and Smith 2017; Zheng, made to calibrate a single-day DMSP-OLS image with Weng, and Wang 2019) has provided a detailed dis- VIIRS-DNB image using Dome C in the Antarctic as the cussion on the inconsistencies amongst DMSP-OLS calibration site (Shao et al. 2014). Another method for satellites; and between DMSP-OLS and VIIRS-DNB inter-calibrating the composites was developed to eval- datasets. Given the fact that DMSP-OLS has served uate city light dynamics during the Syrian civil war which as an important tool for socio-economic studies, the started in 2011 (Li et al. 2017a); however, due to the two datasets (i.e., DMSP-OLS and VIIRS-DNB) can be unavailability of DMSP-OLS data after 2013, an attempt inter-calibrated to create a consistent time-series was made to simulate DMSP-OLS data from VIIRS-DNB dataset for long-term studies. data using the power function and the Gaussian low Several algorithms have been developed to over- pass filter. Li et al. (2020) have used a sigmoid-based come the shortcomings of the DMSP-OLS data in a way regression approach with 2013 as the common year to that can achieve better correlation with Gross Domestic convert VIIRS-DNB datasets to DMSP-like datasets. Product (GDP), Electric Power Consumption (EPC), and Recently, several other attempts have been made by other socio-economic factors (Li and Zhou 2017b; Zhao, researchers to generate DMSP-like products (Ma et al. Zhou, and Samson 2015; Zhao et al. 2017). Radiometric 2020; Zheng, Weng, and Wang 2019). inter-calibration of DMSP-OLS composites is an impor- Several techniques used to perform inter-calibration tant issue, since there was no onboard radiometric cali- of multi-satellite data, are: (1) Regression models (e.g., bration on DMSP satellites (Mukherjee et al. 2017). Elvidge et al. 1997); (2) Statistical inter-calibration (e.g., Several methods have been proposed by researchers Zhang, Pandey, and Seto 2016); (3) Vicarious ground- to address this issue. One of the methods performs based calibration (e.g., Odongo, Hamm, and Milton radiometric inter-calibration based on the pseudo- 2014); (4) Pseudo-invariant calibration sites (e.g., invariant region method, which assumes that there are Mukherjee et al. 2017); and (5) Machine learning (e.g., invariant pixels in multi-temporal NTL images (i.e., pixels Brown et al. 2008). There is a growing need for creating for which the lights have changed very little over time). more sophisticated and robust methods of inter- These invariant pixels are used as training samples to calibration using various machine-learning algorithms generate an inter-calibration function (Elvidge et al. instead of the traditional linear or nonlinear regression 2009; Mukherjee et al. 2017). Other methods for radio- analyses applying empirical and semi-empirical meth- metric calibration are the second-order regression and ods. Machine-learning methods require very less optimal threshold method (Liu et al. 2012), the power- a priori knowledge about each sensor’s data distribu- law regression method (Wu et al. 2013), and the ridge- tions, relationships, sensor operations, calibrations, and line sampling regression method (Zhang, Pandey, and algorithms. These methods learn the patterns from given Seto 2016). Algorithms have also been developed to datasets and work as regularity detectors that discover address the saturation problem. One of the approaches statistically salient properties of investigated data used was inter-calibrating multi-satellite data, wherein (Rodriguez-Galiano et al. 2015; Gumma et al. 2020). Inter- NTL data were combined with Moderate Resolution calibration of multiple satellite and ancillary data for the Imaging Spectroradiometer (MODIS) normalized differ - construction of consistent time-series data using techni- ence vegetation index (NDVI) data to produce vegetation ques like Support Vector Machine (SVM) and Artificial corrected datasets which reduces the saturation effect Neural Networks (ANN) based machine-learning algo- over the urban centers (Zhang, Schaaf, and Seto 2013). rithms attempted earlier (Kwiatkowska and Fargion 1146 S. SAHOO ET AL. 2003; Brown et al. 2008) has proven to be useful. 2.1. Nighttime lights data Likewise, the Random Forest (RF) algorithm is claimed The DMSP-OLS and VIIRS-DNB NTL products available as one of the best machine-learning algorithms for sev- in the public domain (Table 2) are used as primary eral applications (Cracknell and Reading 2014; Kühnlein datasets in this study. The DMSP-OLS annual cloud- et al. 2014; Rodriguez-Galiano et al. 2015). free NTL composites (Version 4), known as stable In the present study, we investigate the potential lights products, produced in 30 arc-second grids are of machine-learning algorithms to inter-calibrate the available from 1992 to 2013 through the website of DMSP-OLS and VIIRS-DNB data with an aim to pro- NOAA-NCEI (NOAA-NCEI 2018). These products show duce VIIRS-DNB-like products for the duration lights from cities, towns, and other sites with persis- 2004–2012, such that these new products can be tent lighting (including gas flares). The ephemeral used in conjunction with the available VIIRS-DNB lights, such as fires, are discarded and the background composites (2013–2020) for long-term socio- noise is replaced with zero pixel value. Pixel values economic studies. The long-term NTL datasets hence range from 0 to 63. The annual NTL composites start- produced have the data range and statistical distribu- ing from the year 2004 to 2013 are used in this study tion close to VIIRS-DNB data. This serves as an impor- (Table 2). There are a few years in which two satellites tant and useful remotely sensed NTL record for have been collecting data. In that case, the annual maintaining uniformity across time (Bennett and composites with maximum Sum of Lights (SOL), Smith 2017). The Multi-Layer Perceptron (MLP) with shown in the gray shade in Table 3, are considered deep neural networks (DNN) architecture and the (Elvidge et al. 2009) for further analysis. widely used Random Forest (RF) algorithms have The VIIRS-DNB NTL datasets (Version 1) are avail- been used to train and test the models, and produce able in the form of monthly composites since the VIIRS-DNB-like annual NTL products from the April 2012. These datasets are produced in 15 arc- DMSP-OLS and ancillary data. second grids and pixel values represent radiance in −2 −1 nW cm sr . In this study, the VIIRS-DNB monthly 2. Study area and data used composites of 2013 are used (Table 2). The present study is carried out for the State of Uttar Pradesh in India (Figure 1), whose capital city is 2.2. Ancillary datasets Lucknow. Lying between 77.1°N and 84.6°N latitudes Human settlements are mainly controlled by physio- and 23.9°E & 30.4°E longitudes, Uttar Pradesh is geo- graphy and socio-economic development of an area graphically the fourth largest state in the country 2 (Fu et al. 2019; Jawarneh, Julian, and Lookingbill 2015; covering an area of 243,290 km . With a population 2 Siddiqui et al. 2018). In this study, DEM and its deri- density of 828 persons per km , it is India’s most vatives are used as proxies for physiography, and populous state. Currently, it is the fourth largest road, land cover, and their derivatives are used as State economy of the country with a gross state proxies for socio-economic development. domestic product of US$ 240 billion (Uttar Pradesh Government 2011). As per reports provided by “Pradhan Mantri Sahaj Bijli Har Ghar Yojana (Prime 2.2.1. Land cover Minister scheme for ease of access to electricity for The Terra/Aqua MODIS Land Cover (Yearly L3 Global each household),” an ongoing electrification scheme, 500 m SIN Grid) products for the years 2004 to 2013 are Uttar Pradesh is among the lowest electrified states in used in this study. Among the five different land cover India (REC Limited 2018). For a State with such a high classification schemes provided [i.e., (i) International population density and poor electrification rate, it is Geosphere Biosphere Programme (IGBP) global vege- beneficial to study the long-term growth patterns of tation classification scheme; (ii) University of Maryland light to get a better understanding of the State’s scheme; (iii) MODIS-derived LAI/fPAR scheme; (iv) electrification performance. MODIS-derived Net Primary Production scheme; (v) The datasets used to carry out the present study Plant Functional Type scheme], IGBP global vegetation are given in Table 2, and a brief description is pro- classification scheme (i.e., Land Cover Type 1) is used vided below. here. The IGBP scheme includes 17 land cover classes: GISCIENCE & REMOTE SENSING 1147 Figure 1. (a) Location and annual DMSP-OLS NTL composite; and (b) annual VIIRS-DNB NTL composite of Uttar Pradesh for the year 2013. The annual composite of VIIRS-DNB is prepared using the monthly NTL composites and PCA method (Sahoo, Gupta, and Srivastav 2019b). 1148 S. SAHOO ET AL. Table 2. Datasets used in the present study. Grid size/Level of Data Year Data Model/Type detail Source DMSP-OLS annual NTL 2004 to 2013 Raster/Continuous 30 arc seconds NOAA-NCEI composites VIIRS-DNB monthly 2013 Raster/Continuous 15 arc seconds NOAA-NCEI composites Land cover (annual) 2004 to 2013 Raster/Categorical 500 m MODIS DEM 2000 Raster/Continuous 30 m SRTM Roads Data available as on 25 October 2018 (data contributed Vector/Categorical Scale independent OpenStreetMap during 2008–2018) Gross domestic product 2004 to 2013 Single value for the Single value for the IndiaStat (annual) State State https://ngdc.noaa.gov/eog/dmsp/downloadV4composites.html https://ngdc.noaa.gov/eog/viirs/download_dnb_composites.html https://e4ftl01.cr.usgs.gov/MOTA/MCD12Q1.006 https://earthexplorer.usgs.gov https://download.geofabrik.de/asia/india.html https://www.indiastat.com Table 3. The annual DMSP-OLS NTL datasets (version 4) are made available by NOAA-NCEI. The datasets used in the current study are highlighted in gray shade, selected based on the maximum SOL value. Satellite Year F15 F16 F18 2004 F152004 F162004 2005 F152005 F162005 2006 F152006 F162006 2007 F152007 F162007 2008 F162008 2009 F162009 2010 F182010 2011 F182011 2012 F182012 2013 F182013 11 natural vegetation classes, 3 human-induced land OpenStreetMap (Figure 2(d)). The road data used classes, and 3 non-vegetated classes. Out of these 17 were contributed to OpenStreetMap between classes, 14 classes fall in the study area (Figure 2(a)). 2008 and 2018, with nearly 70% of the data con- tributed by 2011. There are 26 types of road, 2.2.2. Digital elevation model which can be broadly classified into five major The Shuttle Radar Topography Mission (SRTM) digital classes as follows: elevation model (DEM) having a spatial resolution of 30 m is used here (Figure 2(b)). The elevation in the Class 1: Roads – Roads that consist of roads of region varies from 42 m to 888 m above mean sea higher importance such as National Highway level, with <1% area having elevation >400 m. The (NH), State Highway (SH), major city roads, and high elevation areas are limited to the northwest residential roads. (Himalayan foothills) and southeast (Bundelkhand Class 2: Link Roads – This category consists of roads plateau). The area is mainly characterized by that serve as a link between the Class 1 roads. a gentle eastward slope; 99.5% area has a <10° slope Class 3: Special Roads – Service roads, pedestrian (Figure 2(c)). The most populous cities are on the roads, tracks, and living streets fall under this banks of the rivers, namely, Ganga, Yamuna, Gomti, category, which is made to serve the public to Ghaghara, Betwa, and Son (Figure 2(b)). access a particular facility like business parks, shopping plaza, and residential areas. 2.2.3. Roads Class 4: Paths – Designated footpaths, steps, and The different types of roads in Uttar Pradesh were paths for walkers or cyclists fall under this extracted from the Road map made available by category. GISCIENCE & REMOTE SENSING 1149 Figure 2. (a) Land cover map for the year 2013; (b) SRTM DEM; (c) slope map; (d) road map (source: Geofabrik GmbH and OpenStreetMap 2018); (e) road density map; and (f) distance from roads map, of Uttar Pradesh. Class 5: Others – Other types of roads like cycle- products from DMSP-OLS annual composites. Apart way or unpaved tracks, which are of least impor- from the DMSP-OLS annual composite, three ancillary tance, fall under this category. data products (Land cover, DEM, and Road map) of Uttar Pradesh have been used as independent vari- Two derivative maps, road density (Figure 2(e) and ables for the modeling. All the datasets are resampled distance from road (Figure 2(f)) are prepared using to 500 m spatial resolution using the nearest neighbor the road map (Figure 2(d)) and used in the analysis. method and further processed for training two types of models, viz. MLP and RF. The modeled (or simulated) VIIRS-DNB annual NTL composites (also referred to as 3. Methodology VIIRS-DNB-like composites) are evaluated both qualita- The objective of this paper is to explore the viability of tively and quantitatively. The qualitative evaluation is a deep learning algorithm as a tool to produce long- carried out based on visual analysis of actual DMSP- term NTL datasets by creating VIIRS-DNB-like annual OLS, actual VIIRS-DNB, and modeled VIIRS-DNB annual 1150 S. SAHOO ET AL. NTL composites. For quantitative evaluation, scatter 3.1. Pre-processing of DMSP-OLS and VIIRS-DNB plots, dynamic range, and SOL of actual and modeled datasets products are analyzed. In addition, the statistical rela- tion between SOL and GDP of the study area is studied To correct for lack of on-board calibration and varying for actual DMSP-OLS and modeled VIIRS-DNB annual instrumental gain levels in the DMSP-OLS dataset, 2004 NTL composites for the period before 2013 (i.e., before to 2013 annual composites are inter-calibrated among the launch of VIIRS). Figure 3 shows the data and steps themselves using Ridgeline Sampling Regression (RSR) used for training and evaluating the models and pre- method, taking 2013 as reference year (Zhang, Pandey, paring the calibrated time-series VIIRS-DNB-like NTL and Seto 2016; Sahoo 2019a). The VIIRS-DNB monthly datasets from DMSP-OLS datasets. composites of 2013 (from January to December) are Figure 3. Methodology flowchart showing types of input data used and steps performed for training the model using machine- learning algorithms, model evaluation and preparation of calibrated time-series VIIRS-DNB-like NTL datasets from DMSP-OLS annual composites. GISCIENCE & REMOTE SENSING 1151 processed to prepare the annual composite (Figure 1 3.3. Selection of independent model variables (b)) using the principal component analysis (PCA) using Variable Inflation Factor method (Sahoo 2019a). The inter-calibrated DMSP- To model the relationship with the VIIRS-DNB NTL OLS annual composites of the 2004–2013 period and image (dependent variable), the selection of indepen- the VIIRS-DNB annual composite of 2013 are then used dent variables is made based on the Variable Inflation for further analysis. Factor (VIF). Variance is the squared deviation of a random variable from its mean. When multi- collinearity exists, the variance of the predictor vari- 3.2. Pre-processing of ancillary datasets able inflates. VIF is the ratio of variance in a model with multiple variables divided by the variance of the 3.2.1. Types of input and encoding of categorical model with one variable alone. It is recommended variables that if the VIF of any predictor variable is greater The input data (primary and derivative) used here than 5, and then it is highly collinear with the other essentially belong to two broad categories: predictor variables and while performing prediction ● large standard errors will be encountered Categorical Variables: Land cover; Roads ● (Akinwande, Dikko, and Samson 2015). Continuous Variables: DMSP-OLS annual NTL The variables on which the spatial growth of elec- composites; VIIRS-DNB annual NTL composites; tricity depends are varied. Some are quantifiable like DEM; Slope (derived from DEM); Distance from accessibility, slope, and stability of the land, land cover Road (derived from road map); Road Density type, resident population, distance to nearest electri- (derived from road map); Distance from each city distribution hub, etc., but some are intangible and land cover type (derived from land cover map). cannot be quantified easily such as local social, politi- cal, and religious constraints. In this study, the ancillary To use the categorical inputs, it is first converted into datasets discussed in Section 2.2 have been explored model-understandable numerical data using the label to understand the impact of each variable on NTL in encoder. The label encoder normalizes labels such that the study region. It is seen that (as discussed in Section they contain only values between 0 and n 1, where n 4.1) more than the land cover, the distance to land represents the number of classes. It also transforms the cover type (e.g., distance to built-up, distance to crop- non-numerical labels into numerical labels. After label land, etc.) is a more influencing variable. Therefore, encoding, one hot encoder is used. In one hot encoding, Euclidean distance from each of the land cover classes the column with categorical values is split into n col- has been taken as independent variables. Similarly, the umns, and the numbers are replaced by 0's and type of road is not considered as an independent vari- 1's depending upon the class. In this case, the 14 IGBP able; instead, the distance from road and road density label-encoded classes are converted into 14-bit binary has been taken as independent variables, based on the numbers. exploratory analysis of these variables (discussed in Section 4.1). VIF is then run on these selected variables 3.2.2. Data normalization to shortlist the predictor variables for model training. In general, learning algorithms benefit from a normalization of datasets as it changes the values to a uniform scale, without distorting differences in 3.4. Machine-learning algorithms used for the ranges of values. In the present study, all the inter-calibration variables are normalized between 0 and 1. The trans- Two machine-learning algorithms, viz. (1) Multi-Layer formation is given by Equation (1) below: Perceptron (MLP), a DNN architecture, and (2) Random ðx x Þ Forest (RF) are used for inter-calibration of DMSP-OLS min x ¼ (1) norm ðx x Þ and VIIRS-DNB annual NTL composites. Based on these max min two algorithms, various models are developed and are where x and x are maximum and minimum max min trained with continuous variables as well as with the values of the variable, respectively. combination of categorical and continuous variables. 1152 S. SAHOO ET AL. Variables such as distance from the road and different prediction is taken as the average of the output land cover types are used as inputs, which had higher value of all the trees. These decision trees are built weights away from the feature, and 0 at wherever the using samples drawn with replacement from the feature existed. The reverse weights of these variables training dataset. During the construction of these are also used as inputs to check whether the model is trees, the best split among a random subset of the affected by reversing the weights. Therefore, four features is selected. This randomness marginally rises types of models were built to train: (1) Categorical the bias, but due to averaging, the variance decreases, and Continuous variables, (2) Categorical and hence yielding an overall better model. Thus, the Reversed Continuous variables. (3) Continuous vari- resulting prediction is taken as an average of the ables, and (4) Reversed Continuous variables. The output value of all the trees (Breiman 2001). DMSP-OLS and land cover data are available annually. N � Hence, for predicting VIIRS-DNB annual composites, θ ðxÞ ¼ þ t ðxÞ (3) n¼1 these two datasets are changed for each year, whereas DEM and roads (primary and derivative) were taken as where θ is a random vector, N is the total number of constant. The DMSP-OLS dataset from 2004 to 2012 trees, n is the sample drawn with replacement (also was inter-calibrated among themselves keeping known as bootstrap sample), x is the input, and t is DMSP-OLS 2013 dataset as the base year (Sahoo, the individual decision tree as given in Equation (4): Gupta, and Srivastav 2019b), using Ridgeline � � � t ðxÞ ¼ t ðx; z ; . . . . . . . . . . . . . . . ; z Þ (4) n n1 nk Sampling Regression (RSR) method (Zhang, Pandey, th and Seto 2016), before using them as inputs in trained where, z (k = 1 . . . K) is the k training sample with nk machine-learning models for preparing VIIRS-DNB-like pairs of values for the target variable and covariates annual NTL composites. (x): z = (x ; y ). ni k k A brief description of both the machine-learning algorithms used in the present study is given below. 3.4.3. K-Fold cross-validation method for accuracy evaluation during model training 3.4.1. Multi-layer perceptron algorithm The general approach of calculating accuracy is to Multi-Layer Perceptron (MLP), a feed-forward neural spatially split the entire dataset into training and network, is implemented for mapping DMSP-OLS to testing (validation) sets. Then, based on error metrics, VIIRS-DNB data. The objective is to discover an accuracy is determined from the testing set, and the unknown function f as shown in Equation (2), which result entirely depends on the random choice for the relates the input vectors in X to the output vectors in pair of training and validation sets. In such cases, the Y (Gardner and Dorling 1998): accuracy may change with a change in the test set. To overcome this problem, we have used the K-Fold Y ¼ fðXÞ (2) cross-validation technique for accuracy evaluation where, X ¼ ½n� k�, Y ¼ ½n� j�, n represents the num- during model training. In the K-Fold cross-validation, ber of training patterns, k the number of input vari- the entire dataset is divided into K folds or K spatial ables, and j the number of output variables. The subsets; K number of experiments are conducted, and neural net is trained with DMSP-OLS annual compo- in each experiment (K – 1) spatial sub-section is used site and other continuous/categorical data as input for training and the left-out section is used for testing variables and VIIRS-DNB annual composite as an out- (Figure 4). This way, the model is tested to work for put variable, for the overlapping year of 2013. the entire study region. Subsequently, the resulting weighting functions are applied to the DMSP-OLS data before 2013 using the ancillary data. The functions enable the production of 4. Results and discussion annual VIIRS-DNB-like NTL datasets. 4.1. Selection of independent variables and their relation with nighttime light data 3.4.2. Random Forest algorithm Random Forest (RF) is an ensemble of decision trees The distribution of electricity in a region or the growth (Equation (3)). In RF-based regression, the final of artificial lights due to anthropogenic activities is GISCIENCE & REMOTE SENSING 1153 Figure 4. Schematic diagram showing K-Fold cross-validation method. dependent on several factors, such as land cover type, composites, the variation in SOL in DMSP-OLS data for terrain, road network, etc. Such factors been have these variables has been analyzed. utilized for inter-calibrating DMSP-OLS and VIIRS- Figure 5 shows the variation in SOL for different DNB datasets. As the purpose of the present study is land cover types. It is clear from the figure that the to create a consistent VIIRS-DNB-like annual time- maximum amount of NTL is contributed by the Urban series NTL datasets using the DMSP-OLS annual NTL and Built-up land class, followed by Open Shrublands, Figure 5. Variation in SOL with respect to land cover types. 1154 S. SAHOO ET AL. Cropland/Natural Vegetation Mosaics, and so on. It is density of road is located in major cities. Further, the obvious that as the distance from the urban center areas consisting of the least density of roads coincide increases and one moves toward croplands and rural with croplands or other natural land cover types. The areas, the SOL decreases due to less population den- variation in SOL to road density is shown in Figure 6(c). sity, few settlements, or lack of urbanization. The dis- It can be observed that the denser the road network tance from these land cover types do influence the more intense is the light. The variation in SOL with lighting mechanism; hence, the distance from each respect to distance from the road is shown in Figure 6 land cover type is taken into consideration for training (d). The SOL can be seen decreasing gradually as the the model. For this, the Euclidean distance of every distance from the road increases. cell in the raster to the nearest source is calculated. Difficult terrain often hinders the development of The variation in SOL for different types of roads is human settlements and, thus, the artificial lights. The analyzed by creating a buffer of 2 km around the roads. topography of the study area is dominantly plain, with The analysis is done on five major classes and their limited hilly terrain in the north and plateau in the south. subclasses as shown in Figure 6(a, b), respectively. It is Figure 7 represents the variation in SOL with respect to observed that roads that fall within the urban areas like the slope. As obvious, the majority of artificial lights are footways, steps, pedestrian paths, service roads contri- concentrated in very gently sloping plains. bute more toward the SOL despite having a smaller Further analysis (at finer bins for Figures 6(c, d) and length as they are surrounded by city lights and are 7) reveals a non-linear relation between SOL and road well equipped with street lighting facilities. On the con- density, distance to road, and road density. As trary, the major types of roads that consist of NH, SH and machine-learning algorithms make predictions based have lengths much higher than paths do not contribute on learning from training data without human inter- much to SOL as the main means of lighting are street- vention (Kubat 2017), such non-linear relation between lights. Moreover, as the road stretches away from cities the SOL and independent variables is largely supposed the amount of light per unit length of road decreases. to be taken care of during the model training and As no proper correlation was observed between SOL model prediction. After analyzing the relation of the and road classes, it was felt necessary to understand variables individually with artificial night-lights, the the relation between SOL and the road network as final selection of independent variables to be used for a whole. Hence, using the road map of Uttar Pradesh, training the models is made based on VIF. A total of 19 road density (Figure 2(e)) and distance from road variables were initially selected for evaluation based on (Figure 2(f)) maps were prepared for analysis. From VIF (Table 4). The iterations performed for VIF calcula- Figure 2(e), it can be observed that the maximum tion are shown in Table 4. In each iteration, the variable Figure 6. Variation in SOL with respect to (a) major road types; (b) subclasses of each major road type (the color represents the major road type as shown in Figure 6(a)); (c) road density; and (d) distance from road. GISCIENCE & REMOTE SENSING 1155 Figure 7. Variation in SOL with respect to slope. having the highest VIF value among all the variables continuous and categorical variables using the two was removed. Finally, after eight iterations, 12 variables machine-learning algorithms, MLP and RF. In the case with VIF <5 (shown in the gray shade in Table 4) were of MLP, the model parameters such as the number of found, which were then used to train the models. The hidden layers, number of nodes in each hidden layer, variables selected for model training include distance number of iterations were changed to achieve the to several land cover classes, distance to road, road highest accuracy and lowest error. The other para- density, slope, and DMSP-OLS. meters, like activation function, optimizer, regulariza- tion term, and learning rate were kept the same in all the cases. The “rectified linear unit function” was used 4.2. Machine-learning-based models for as an activation function. Stochastic gradient-based inter-calibrating DMSP-OLS and VIIRS-DNB annual optimizer “adam” was used as an optimizer. The reg- nighttime light composites ularization term and learning rate were kept as As discussed earlier, four types of models were devel- 0.00001 and 0.0001, respectively. The K-fold valida- oped and tested with different combinations of tion method was implemented. The DMSP-OLS and Table 4. VIF values of different variables from eight iterations. The final 12 variables selected for modeling are highlighted in gray shade. For brevity “Distance from” has been abbreviated as “Dist.” Iteration Number → 1 2 3 4 5 6 7 8 Classes ↓ VIF Dist. woody savannas 54.57 - - - - - - - Dist. evergreen broadleaf forest 34.80 34.30 - - - - - - Dist. mixed forest 15.81 15.81 13.48 - - - - - Elevation 25.95 25.67 11.01 10.35 - - - - Dist. evergreen needleleaf forest 14.50 12.20 9.00 8.28 7.55 - - - Dist. deciduous forest 46.20 12.72 12.42 7.74 7.53 7.14 - - Dist. barren land 6.61 6.59 6.58 6.58 6.57 6.54 6.53 - Dist. savannas 5.93 5.65 5.62 5.61 5.53 5.37 4.86 4.86 Dist. open shrublands 11.53 10.67 7.75 7.66 5.57 5.53 5.09 4.73 Dist. crop/vegetation mosaic 7.53 7.43 6.47 5.71 5.31 4.78 4.74 4.66 Dist. permanent wetlands 8.03 7.41 6.49 5.88 5.83 5.50 4.90 4.58 Dist. water bodies 9.85 9.25 8.99 8.72 7.35 6.71 3.74 3.02 Dist. grasslands 4.92 4.77 4.56 4.39 4.29 4.26 4.26 2.88 Dist. urban & built-up 3.55 3.50 3.33 3.33 3.28 2.80 2.63 2.62 Road density 2.76 2.65 2.60 2.58 2.57 2.55 2.41 2.28 Dist. road 2.29 2.28 2.27 2.27 2.27 2.23 2.20 2.20 DMSP-OLS 1.68 1.67 1.66 1.66 1.66 1.64 1.64 1.63 Slope 1.36 1.36 1.35 1.35 1.30 1.28 1.28 1.27 Dist. croplands 1.27 1.27 1.23 1.23 1.20 1.20 1.20 1.19 1156 S. SAHOO ET AL. VIIRS-DNB annual NTL composites for the All the four model configurations were tested common year (i.e., 2013) were used for developing using the RF algorithm, with different numbers of the models. The modeled (predicted) and actual trees ranging from 10 to 500. With the increase in VIIRS-DNB annual NTL composites of 2013 were com- the number of trees, minor changes occurred in R pared with each other using coefficient of determina- and RMSE, but the time for training and prediction tion (R ), and root-mean-square error (RMSE). Using increased. Finally, all the models were trained by a simple grid-based approach, the best-achieved keeping the number of trees as 200. Other para- parameter configuration for all the four types of mod- meters, viz. maximum features to select and the max- els, and the results achieved are shown in Table 5 and imum depth of trees were kept the same in all the Figure 8, respectively. cases. The results achieved are shown in Figure 9. Table 5. MLP architecture used for each model. Model Inputs Hidden Layer Configuration No. of Iterations R RMSE Categorical and Continuous variables 256, 128, 64, 32 1000 0.75 4.77 Categorical and reversed Continuous variables 256, 128, 64, 32 1000 0.78 5.20 Continuous variables 512, 256, 128, 64, 32 1000 0.86 3.78 Reversed continuous variables 512, 256, 128, 64,32 1000 0.77 4.89 Figure 8. Density scatter plots between actual and predicted VIIRS-DNB annual NTL composites for the year 2013 using MLP model. The different configurations of MLP model are (a) categorical and continuous inputs, (b) categorical and reversed continuous inputs, −2 −1 (c) continuous inputs and (d) reversed continuous inputs. All quantities in the graphs are in nW cm sr . GISCIENCE & REMOTE SENSING 1157 Figure 9. Density scatter plots between actual and predicted VIIRS-DNB annual NTL composites for the year 2013 using RF model. The different configurations of RF model are (a) categorical and continuous inputs, (b) categorical and reversed continuous inputs, (c) −2 −1 continuous inputs and (d) reversed continuous inputs. All quantities in the graphs are in nW cm sr . The results obtained through MLP and RF-based strengthened by visually analyzing the areas around models are evaluated both qualitatively and quantita- select cities. tively and are elaborated in Section 4.2.1. Figure 10 shows the actual DMSP-OLS, actual VIIRS- DNB, the RF, and MLP model predicted VIIRS-DNB annual NTL composites around Lucknow, Agra, and 4.2.1. Evaluation of predicted VIIRS-DNB annual Prayagraj cities [locations highlighted in Figure 11 a(i)] nighttime light composites for the year 2013. The following observations are The scatter plots of actual versus predicted VIIRS-DNB made based on a detailed visual analysis of the mod- NTL radiance values (for the year 2013) obtained eled products. First, the predicted VIIRS-DNB NTL using the MLP-based models (Figure 8) show images show more spatial granularity than the actual a larger spread of values across the 1:1 line, lower R DMSP-OLS data, thereby signifying that better spatial values (0.75–0.86), and higher RMSE values (3.75–- night-light details can be derived by using advanced 5.20) as compared to RF-based models (Figure 9; machine-learning algorithms and ancillary data. R = 0.87–0.89; RMSE = 2.12–2.32). Further, the MLP Second, both the RF and MLP models successfully and RF models with continuous variables as inputs remove the blooming effect present in the actual performed better than the remaining models. The DMSP-OLS data. Third, the MLP outputs (Figure 10, pixel-wise spatial comparison shows that the RF- column d) are comparatively smoother, which leads based models yield better results as compared to to loss of spatial details. Fourth, the RF outputs are MLP-based models. This observation is further 1158 S. SAHOO ET AL. Figure 10. Actual and predicted annual NTL composites around select cities in the study area for the year 2013: (a) actual DMSP-OLS, (b) actual VIIRS-DNB, (c) predicted VIIRS-DNB (using continuous input RF model), and (d) predicted VIIRS-DNB (using continuous input MLP model) around the cities of (i) Lucknow, (ii) Agra and (iii) Prayagraj. All images are at same spatial scale, as shown in the bottom right image. relatively brighter (higher radiances) in regions of low improvement in the RF and MLP model performances lighting. The roads and smaller towns with low light- by changing the model parameters. Therefore, using ing are particularly apparent on the RF outputs, often the best model achieved in both the cases, the VIIRS- not visible on the MLP outputs. Fifth, the saturated DNB annual NTL composites are generated for the pixels in the DMSP-OLS data in the center of major years 2004 to 2012. The details are provided in the cities [e.g., Lucknow in Figure 10d(i) and Kanpur (not next section. shown)] have relatively smaller radiance values in the MLP-based models, which is not the case in the RF- 4.2.2. Generation of time-series VIIRS-DNB-like NTL based model outputs. In other words, the MLP-based datasets models are not able to adequately capture the urban Using the trained MLP and RF models, the VIIRS-DNB- centers of the major cities where DMSP-OLS data are like NTL images are generated for the years 2004–2012. saturated. Sixth, a significant reduction in radiance In doing so, the DMSP-OLS and land cover (i.e., distance values with increasing distance from the urban cen- to individual land cover types) inputs are varied corre- ters is observed in the RF model-derived products, spondingly for each year, while keeping the other whereas the MLP products show a gradual decrease independent variables (elevation and its derivatives, in radiance values away from the urban centers. Given road derivatives) unchanged with time. The represen- these observations, the RF model (Figure 10, column tative outputs of the study area for 3 years (2004, 2008, c) performs better as it emulates the VIIRS-DNB data and 2012) based on RF and MLP models are shown in (Figure 10, column b) more closely. Figures 11 and 12, respectively. The gradual intensifica - As mentioned earlier, the RF and MLP models with tion of light and lighting of new towns with time are continuous variables as inputs performed better. the visible characteristics of the time-series data gen- Moreover, it is observed that there was no significant erated. The RF model-based outputs indicate that the GISCIENCE & REMOTE SENSING 1159 Figure 11. Predicted VIIRS-DNB-like annual NTL composites for select cities, using continuous input RF model. Columns show different years (a) 2004, (b) 2008 and (c) 2012. Rows show (i) entire study area (Indian state of Uttar Pradesh), (ii) Lucknow, (iii) Agra, and (iv) Prayagraj cities. The SOL for each of the cities is calculated by taking a buffer of 0.4° x 0.4° around each city. Units of SOL and image −2 −1 legends are in nW cm sr . All city images are at same spatial scale, as shown in (a-iv). SOL values increased by 7.2% and 9.5% during 2004 to composites of 2013 are also relevant for the 2008 and 2008 to 2012, respectively, whereas MLP 2004–2012 composites. Particularly, it is evident that outputs indicate the increase in SOL values by 4.9% the RF model can better handle the problem of satu- and 30.3% for the corresponding period. It is important rated urban cores, which the input dataset DMSP-OLS to note that in Figures 11 and 12, the same stretch is plagued with. The MLP model, on the other hand, is values have been used for each row, i.e., row (i) Uttar not able to capture the saturated urban cores. Pradesh is stretched from 0 to 25, row (ii) Lucknow is The accuracy of predicted VIIRS-DNB annual NTL stretched from 0 to 200, and so on. This has been done composites of 2004–2012 was also tested on for easy visual comparison of the two models. The a regional scale by studying the statistical relation observations made in Section 4.2.1 based on visual between SOL and GDP of the State (Figure 13). 2 2 comparison of actual and modeled VIIRS-DNB NTL Notably, the R values of MLP (R = 0.93) and RF 1160 S. SAHOO ET AL. Figure 12. Predicted VIIRS-DNB-like annual NTL composites for select cities, using continuous input MLP model. Columns show different years (a) 2004, (b) 2008 and (c) 2012. Rows show (i) entire study area (Indian state of Uttar Pradesh), (ii) Lucknow, (iii) Agra, and (iv) Prayagraj cities. The SOL for each of the cities is calculated by taking a buffer of 0.4° x 0.4° around each city. Units of SOL and −2 −1 image legends are in nW cm sr . All city images are at same spatial scale, as shown in (a-iv). (R = 0.66) model outputs are higher than that for the study area are significantly higher in the case of RF. actual DMSP-OLS products (R = 0.60). It is, thus, evi- Moreover, RF model outputs show a larger variability in dent that machine-learning algorithms can be signifi - SOL with time as compared to MLP outputs, which cantly utilized for inter-calibrating DMSP-OLS and caused lower R value between RF-derived SOL VIIRS-DNB datasets to create consistent annual time- and GDP. series NTL products for longer periods. Further, the The modeled VIIRS-DNB-like NTL datasets for 2004 MLP model performs better at a regional scale, as to 2012 using both the RF and MLP models are made compared to RF. Detailed analysis indicates that while available for download by the researchers from the the maximum radiance values in the modeled products following link (https://doi.org/10.5281/zenodo. from both the models are similar, the SOL values in the 3967104). GISCIENCE & REMOTE SENSING 1161 Figure 13. Relation between GDP and SOL derived from (a) predicted VIIRS-DNB annual NTL composites using RF algorithm, (b) predicted VIIRS-DNB annual NTL composites using MLP algorithm, and (c) for actual DMSP-OLS annual NTL composites, for the period 2004–2012. 4.3. Discussion a prediction. On the other hand, the MLP model The qualitative and quantitative analysis of the RF and takes time to train data, but the prediction is very MLP model outputs have shown particular merits and fast. Therefore, holistically, both the models have limitations of each model. The R and RMSE values of their strong areas of applications: MLP is recom- the model predicted outputs for the reference year mended for regional-scale studies, while RF is recom- 2013 and the visual analysis of predicted temporal mended for local-scale studies. VIIRS-DNB-like NTL images indicated that RF model The DMSP-OLS products have been used for more performance is better as far as capturing the spatial than two decades as a proxy for monitoring socio- details at the local-scale is concerned. Both models economic activity and human well-being at National are able to significantly reduce the blooming effect, and Global levels (Chen and Nordhaus 2011; Ghosh seen in the actual DMSP-OLS data. On the other hand, et al. 2013). In this regard, to accurately study the the temporal VIIRS-DNB-like annual NTL products spatio-temporal trend of artificial night-lights from generated from the past DMSP-OLS data using the the time when VIIRS-DNB did not exist, machine learn- MLP model show a better correlation with economic ing can be used to simulate datasets and study histor- development. This indicates that the MLP provides ical aspects at a much better resolution compared to stable outputs in terms of SOL of the study area as then available DMSP-OLS data. a whole, but it fails to handle the saturated city cores. The proposed approach of inter-calibration using Further, the RF model is easy to implement and machine-learning architectures will not only help to train when the number of trees is kept small. For perform the long-term analysis of NTL at a regional better predictive power, the number of trees needs scale through the combined study of DMSP-OLS and to be increased, which increases the time for recently available VIIRS-DNB datasets but will also 1162 S. SAHOO ET AL. provide improved results and understanding of socio- environment like an increase in air pollution due to economic development that would have been industrialization and urbanization. obtained using standalone DMSP-OLS and VIIRS-DNB datasets. In the present study, among the ancillary 5. Conclusions spatial datasets, only land cover is used as a time- variant variable along with DMSP-OLS datasets. The availability of NTL composites from the VIIRS-DNB Incorporation of the temporal information on the opened the scope for inter-calibrating them with DMSP- roads and other infrastructure in the model is likely OLS and create consistent annual time-series NTL data- to improve the inter-calibration results further. The sets for the long-term studies. The potential of two proposed methods for inter-calibration can also be machine-learning algorithms, MLP (having a DNN archi- tested using other global DEMs to study the improve- tecture) and RF, is evaluated in this study for inter- ment in results if any. The proposed framework was calibration of the DMSP-OLS and VIIRS-DNB annual found to work well over a large State of Uttar Pradesh, NTL composites. The annual NTL composites of DMSP- which is primarily a flat, tropical, agrarian, and densely OLS and VIIRS-DNB for the common year of 2013 along populated area. Researchers can explore the frame- with ancillary datasets on land cover, road network, and work further over other sites with different topogra- topography are used to train the MLP and RF models. phy, climate, and socio-economic setting. The best model configuration is then used to simulate The annual availability of both DMSP-OLS and VIIRS- VIIRS-DNB-like images from DMSP-OLS annual NTL com- DNB is limited to only 1 year (i.e., 2013). This poses posites using both MLP and RF algorithms. Detailed a unique challenge in inter-calibrating the two datasets. analysis of the outputs revealed that: (1) Both MLP and Recent studies (Shao et al. 2014; Li et al. 2017a; Zheng, RF models significantly reduce the blooming effect Weng, and Wang 2019; Li et al. 2020; Ma et al. 2020) around settlements, a common problem observed in have attempted to overcome this issue and generated DMSP-OLS data; (2) visual appearance and spatial gran- DMSP-OLS-like products after 2013, whereas the present ularity of the model predicted outputs are close to the study leverages the capabilities of machine learning to actual VIIRS-DNB images; (3) RF captures better spatial generate VIIRS-DNB-like products before 2013. It is also details at local-scale and is able to efficiently handle the important to note that the local time of over-pass of the saturation problem at urban centers, which is again two satellites is different (DMSP-OLS: 1900–2200 hours a common issue with DMSP-OLS data; (4) MLP, on the (Elvidge et al. 2009); VIIRS-DNB: ~0130 hours) and, there- other hand, is found to be superior at regional-scale as it fore, its impact on the inter-calibration of the two data- provided better statistical relation between model- sets needs to be further explored. Earlier studies (Li et al. derived SOL of and GDP of the study region. It is, thus, 2017a, 2020; Jeswani et al. 2019; Ma et al. 2020; Zheng, concluded that both the machine-learning algorithms Weng, and Wang 2019) have also noted this as could be appropriately utilized for producing VIIRS-DNB- a limitation and proceeded for further analysis. The end- like NTL images from DMSP-OLS annual NTL compo- of-life sensor degradation of DMSP-OLS and the lead sites, depending on the purpose of the study. The mod- time required for calibration and testing of the VIIRS- eled VIIRS-DNB-like annual NTL composites for 9 years DNB datasets are the other important factors for the (2004 to 2012) generated using trained RF and MLP inter-calibration of the two datasets, which may also models are made available for download by the be explored by the researchers in the future. researchers from the following link (https://doi.org/10. The potential of various other DNN architectures 5281/zenodo.3967104). such as Recurrent Neural Network, Convolutional The utilization of RF and MLP is, thus, proposed for Neural network, etc. can also be investigated in the inter-calibration of DMSP-OLS and VIIRS-DNB annual inter-calibration of NTL datasets. The derived annual NTL datasets at local and regional scales, respectively, time-series NTL data can be used to study various for long-term studies. The long-term NTL datasets are factors that hinder or accelerate the growth of artifi - envisaged to have critical applications in monitoring cial lights in a region. The datasets can also be used to socio-economic development, disaster impact, and study the impact of the growth of lights on the land cover change. The study can be further expanded GISCIENCE & REMOTE SENSING 1163 by using other machine-learning algorithms and Brown, M. E., D. J. Lary, A. Vrieling, D. Stathakis, and H. Mussa. 2008. “Neural Networks as a Tool for Constructing improved spatio-temporal ancillary datasets over Continuous NDVI Time Series from AVHRR and MODIS.” areas with different socio-economic, physiographic, International Journal of Remote Sensing 29 (24): 7141–7158. and climate setting. doi:10.1080/01431160802238435. Cao, S., D. Hu, W. Zhao, Y. Mo, Y. Chen, and Y. Zhang. 2019. “Monitoring Changes in the Impervious Surfaces of Urban Acknowledgements Functional Zones Using Multisource Remote Sensing Data: A Case Study of Tianjin, China.” GIScience & Remote Sensing This study is a part of M.Tech. Dissertation carried out by the 56 (7): 967–987. doi:10.1080/15481603.2019.1600110.s. first author. Thanks to the Director, Indian Institute of Remote Chander, G., T. J. Hewison, N. Fox, X. Wu, X. Xiong, and Sensing, Indian Space Research Organisation (ISRO), Dehradun, W. J. Blackwell. 2013. “Overview of Intercalibration of Satellite India for his support and encouragement. Sincere thanks to Instruments.” IEEE Transactions on Geoscience and Remote National Oceanic and Atmospheric Administration’s National Sensing 51 (3): 1056–1080. doi:10.1109/TGRS.2012.2228654. Centres (NOAA) for providing the DMSP-OLS and VIIRS-NNB Chen, X., and W. D. Nordhaus. 2011. “Using Luminosity Data as nighttime lights data. We thank the anonymous reviewers for a Proxy for Economic Statistics.” Proceedings of the National their useful and critical comments that helped to improve the Academy of Sciences 108 (21): 8589–8594. doi:10.1073/ manuscript to a large extent. pnas.1017031108. Chen, Z., B. Yu, Y. Hu, C. Huang, K. Shi, and J. Wu. 2015. “Estimating House Vacancy Rate in Metropolitan Areas Disclosure statement Using NPP-VIIRS Nighttime Light Composite Data.” IEEE No potential conflict of interest was reported by the authors. Journal of Selected Topics in Applied Earth Observations and Remote Sensing 8 (5): 2188–2197. doi:10.1109/ JSTARS.2015.2418201. ORCID Cracknell, M. J., and A. M. Reading. 2014. “Geological Mapping Using Remote Sensing Data: A Comparison of Five Machine Sumana Sahoo http://orcid.org/0000-0003-3655-9415 Learning Algorithms, Their Response to Variations in the Prasun Kumar Gupta http://orcid.org/0000-0003-0689-4167 Spatial Distribution of Training Data and the Use of Explicit Spatial Information.” Computers & Geosciences 63: 22–33. doi:10.1016/j.cageo.2013.10.008. Author responsibilities Duriscoe, D. M., C. B. Luginbuhl, and C. D. Elvidge. 2013. “The Relation of Outdoor Lighting Characteristics to Sky Glow Sumana Sahoo was responsible for performing the experi- from Distant Cities.” Lighting Research and Technology 46 ments, writing computer programs, data analysis, and prepar- (1): 35–49. doi:10.1177/1477153513506729. ing the first draft of the manuscript. Elvidge, C. D., K. E. Baugh, E. A. Kihn, H. W. Kroehl, E. R. Davis, Prasun Kumar Gupta conceived and designed the experi- and C. W. Davis. 1997. “Relation between Satellite Observed ments, and was responsible for supervising the computer pro- Visible-near Infrared Emissions, Population, Economic gramming, checking of data analysis results, and revising the Activity and Electric Power Consumption.” International manuscript. Journal of Remote Sensing 18 (6): 1373–1379. doi:10.1080/ Dr. S. K. Srivastav was responsible for conceiving the pro- blem, supervising and improving the study, and revising the Elvidge, C. D., D. Ziskin, K. E. Baugh, B. T. Tuttle, T. Ghosh, D. W. Pack, manuscript. E. H. Erwin, and M. Zhizhin. 2009. “A Fifteen Year Record of Global Natural Gas Flaring Derived from Satellite Data.” Energies 2 (3): 595–622. doi:10.3390/en20300595. References Falchi, F., P. Cinzano, D. Duriscoe, C. C. M. Kyba, C. D. Elvidge, Akinwande, M. O., H. G. Dikko, and A. Samson. 2015. “Variance K. Baugh, B. A. Portnov, et al. 2016. “The New World Atlas of Inflation Factor: As a Condition for the Inclusion of Artificial Night Sky Brightness.” Science Advances 2 (6): Suppressor Variable(s) in Regression Analysis.” Open e1600377. doi:10.1126/sciadv.1600377. Journal of Statistics 5 (7): 754–767. doi:10.4236/ Fu, Y., J. Li, Q. Weng, Q. Zheng, L. Li, S. Dai, and B. Guo. 2019. ojs.2015.57075. “Characterizing the Spatial Pattern of Annual Urban Growth by Bennett, M. M., and L. C. Smith. 2017. “Advances in Using Using Time Series Landsat Imagery.” Science of the Total Multitemporal Night-time Lights Satellite Imagery to Environment 666: 274–284. doi:10.1016/j.scitotenv.2019.02.178. Detect, Estimate, and Monitor Socioeconomic Dynamics.” Gardner, M. W., and S. R. Dorling. 1998. “Artificial Neural Remote Sensing of Environment 192: 176–197. doi:10.1016/j. Networks (The Multilayer Perceptron) - a Review of rse.2017.01.005. Applications in the Atmospheric Sciences.” Atmospheric Breiman, L. 2001. “Random Forests.” Machine Learning 45 (1): Environment 32 (14/15): 2627–2636. doi:10.1016/S1352- 5–32. doi:10.1023/A:1010933404324. 2310(97)00447-0. 1164 S. SAHOO ET AL. Geofabrik GmbH and OpenStreetMap. 2018. “India”. Accessed Syrian Civil War.” International Journal of Remote Sensing 38 26 October 2018. https://download.geofabrik.de/asia/india. (21): 5934–5951. doi:10.1080/01431161.2017.1331476. html Li, X., R. Zhang, C. Huang, and D. Li. 2015. “Detecting 2014 Ghosh, T., S. J. Anderson, C. D. Elvidge, and P. C. Sutton. 2013. Northern Iraq Insurgency Using Night-time Light Imagery.” “Using Nighttime Satellite Imagery as a Proxy Measure of International Journal of Remote Sensing 36 (13): 3446–3458. Human Well-being.” Sustainability 5 (12): 4988–5019. doi:10.1080/01431161.2015.1059968. doi:10.3390/su5124988. Li, X., and Y. Zhou. 2017b. “A Stepwise Calibration of Global Gumma, M. K., P. S. Thenkabail, P. G. Teluguntla, A. Oliphant, DMSP/OLS Stable Nighttime Light Data (1992-2013).” J. Xiong, C. Giri, V. Pyla, S. Dixit, and A. M. Whitbread. 2020. Remote Sensing 9 (6): 637. doi:10.3390/rs9060637. “Agricultural Cropland Extent and Areas of South Asia Li, X., Y. Zhou, M. Zhao, and X. Zhao. 2020. “A Harmonized Derived Using Landsat Satellite 30-m Time-series Big-data Global Nighttime Light Dataset 1992-2018.” Scientific Data 7 Using Random Forest Machine Learning Algorithms on the (1): 1–9. doi:10.1038/s41597-020-0510-y. Google Earth Engine Cloud.” GIScience & Remote Sensing 57 Liu, Z., C. He, Q. Zhang, Q. Huang, and Y. Yang. 2012. (3): 302–322. doi:10.1080/15481603.2019.1690780. “Extracting the Dynamics of Urban Expansion in China Guo, W., G. Li, W. Ni, Y. Zhang, and D. Li. 2018. “Exploring Using DMSP-OLS Nighttime Light Data from 1992 to 2008.” Improvement of Impervious Surface Estimation at National Landscape and Urban Planning 106 (1): 62–72. doi:10.1016/j. Scale through Integration of Nighttime Light and Proba-V landurbplan.2012.02.013. Data.” GIScience & Remote Sensing 55 (5): 699–717. Ma, J., J. Guo, S. Ahmad, Z. Li, and J. Hong. 2020. “Constructing doi:10.1080/15481603.2018.1436425. a New Inter-Calibration Method for DMSP-OLS and Jawarneh, R. N., J. P. Julian, and T. R. Lookingbill. 2015. “The NPP-VIIRS Nighttime Light.” Remote Sensing 12 (6): 937. Influence of Physiography on Historical and Future Land doi:10.3390/rs12060937. Development Changes: A Case Study of Central Arkansas Ma, T., C. Zhou, T. Pei, S. Haynie, and J. Fan. 2012. “Quantitative (USA), 1857–2030.” Landscape and Urban Planning 143: Estimation of Urbanization Dynamics Using Time Series of 76–89. doi:10.1016/j.landurbplan.2015.06.010. DMSP/OLS Nighttime Light Data: A Comparative Case Study Jeswani, R., A. Kulshrestha, P. K. Gupta, and S. K. Srivastav. 2019. from China’s Cities.” Remote Sensing of Environment 124: “Evaluation of the Consistency of DMSP-OLS and SNPP-VIIRS 99–107. doi:10.1016/j.rse.2012.04.018. Night-time Light Datasets.” Journal of Geomatics 13 (1): Min, B., and K. M. Gaba. 2014. “Tracking Electrification in 98–105. Vietnam Using Nighttime Lights.” Remote Sensing 6 (10): Kiran Chand, T. R., K. V. S. Badarinath, C. D. Elvidge, and B. T. Tuttle. 9511–9529. doi:10.3390/rs6109511. 2009. “Spatial Characterization of Electrical Power Min, B., K. M. Gaba, O. F. Sarr, and A. Agalassou. 2013. Consumption Patterns over India Using Temporal DMSP-OLS “Detection of Rural Electrification in Africa Using Night-time Satellite Data.” International Journal of Remote DMSP-OLS Night Lights Imagery.” International Journal of Sensing 30 (3): 647–661. doi:10.1080/01431160802345685. Remote Sensing 34 (22): 8118–8141. doi:10.1080/ Kubat, M. 2017. An Introduction to Machine Learning. Springer 01431161.2013.833358. International Publishing AG. doi:10.1007/978-3-319-63913-0 Mukherjee, S., S. K. Srivastav, P. K. Gupta, N. A. S. Hamm, and Kühnlein, M., T. Appelhans, B. Thies, and T. Nauss. 2014. “Improving V. A. Tolpekin. 2017. “An Algorithm for Inter-calibration of the Accuracy of Rainfall Rates from Optical Satellite Sensors Time-series DMSP/OLS Night-time Light Images.” with Machine Learning - a Random Forests-based Approach Proceedings of the National Academy of Sciences India Applied to MSG SEVIRI.” Remote Sensing of Environment 141: Section A - Physical Sciences 87 (4): 721–731. doi:10.1007/ 129–143. doi:10.1016/j.rse.2013.10.026. s40010-017-0444-8. Kwiatkowska, E. J., and G. S. Fargion. 2003. “Application of NOAA-NCEI. 2018. “DMSP & VIIRS Data Download.” Accessed 15 Machine-learning Techniques toward the Creation of July 2018. https://ngdc.noaa.gov/eog/download.html a Consistent and Calibrated Global Chlorophyll Concentration Odongo, V. O., N. A. S. Hamm, and E. J. Milton. 2014. “Spatio- Baseline Dataset Using Remotely Sensed Ocean Color Data.” temporal Assessment of Tuz Gölü, Turkey as a Potential IEEE Transactions on Geoscience and Remote Sensing 41 (12): Radiometric Vicarious Calibration Site.” Remote Sensing 6 2844–2860. doi:10.1109/TGRS.2003.818016. (3): 2494–2513. doi:10.3390/rs6032494. Levin, N., S. Ali, and D. Crandall. 2018. “Utilizing Remote REC Limited. 2018. “Rural Electrification Corporation.” Accessed Sensing and Big Data to Quantify Conflict Intensity: The 1 January 2019. https://www.recindia.nic.in/ Arab Spring as a Case Study.” Applied Geography 94: 1–17. Rodriguez-Galiano, V., M. Sanchez-Castillo, M. Chica-Olmo, and doi:10.1016/j.apgeog.2018.03.001. M. Chica-Rivas. 2015. “Machine Learning Predictive Models Levin, N., and Y. Duke. 2012. “High Spatial Resolution Night-time for Mineral Prospectivity: An Evaluation of Neural Networks, Light Images for Demographic and Socio-economic Studies.” Random Forest, Regression Trees and Support Vector Remote Sensing of Environment 119: 1–10. doi:10.1016/j. Machines.” Ore Geology Reviews 71: 804–818. doi:10.1016/j. rse.2011.12.005. oregeorev.2015.01.001. Li, X., D. Li, H. Xu, and C. Wu. 2017a. “Intercalibration between Sahoo, S. 2019a. “Evaluating the Potential of Night-time Light DMSP/OLS and VIIRS Night-time Light Images to Evaluate City Data to Study Electrification in Uttar Pradesh.” M.Tech. the- Light Dynamics of Syria’s Major Human Settlement during sis, IIRS/Andhra University, Viskhapatnam, India GISCIENCE & REMOTE SENSING 1165 Sahoo, S., P. K. Gupta, and S. K. Srivastav. 2019b. “Comparative Yu, B., M. Tang, Q. Wu, C. Yang, K. Shi, C. Peng, J. Wu, and Analysis between VIIRS-DNB and DMSP-OLS Night-time Light Z. Chen. 2018. “Urban Built-Up Area Extraction From Data to Estimate Electric Power Consumption in Uttar Pradesh, Log-Transformed NPP-VIIRS Nighttime Light Composite India.” International Journal of Remote Sensing 41 (7): Data.” IEEE Transactions on Geoscience and Remote 2565–2580. doi:10.1080/01431161.2019.1693077. Sensing Letters 15 (8): 1279–1283. doi:10.1109/ Shao, X., C. Cao, B. Zhang, S. Qiu, C. Elvidge, and M. V. Hendy. 2014. LGRS.2018.2830797. “Radiometric Calibration of DMSP-OLS Sensor Using VIIRS Day/ Zhang, Q., B. Pandey, and K. C. Seto. 2016. “A Robust Method to night Band.” In Proceedings of SPIE Vol. 9264, Earth Observing Generate A Consistent Time Series from DMSP/OLS Missions and Sensors: Development, Implementation, and Nighttime Light Data.” IEEE Transactions on Geoscience and Characterization III. doi: 10.1117/12.2068999. Remote Sensing 54 (10): 5821–5831. doi:10.1109/ Shi, K., B. Yu, Y. Hu, C. Huang, Y. Chen, Y. Huang, Z. Chen, and TGRS.2016.2572724. J. Wu. 2015. “Modeling and Mapping Total Freight Traffic in Zhang, Q., C. Schaaf, and K. C. Seto. 2013. “The Vegetation China Using NPP-VIIRS Nighttime Light Composite Data.” Adjusted NTL Urban Index: A New Approach to Reduce GIScience & Remote Sensing52 (3): 274–289. doi:10.1080/ Saturation and Increase Variation in Nighttime Luminosity.” 15481603.2015.1022420. Remote Sensing of Environment 129: 32–41. doi:10.1016/j. Siddiqui, A., A. Siddiqui, S. Maithani, A. K. Jha, P. Kumar, and rse.2012.10.022. S. K. Srivastav. 2018. “Urban Growth Dynamics of an Indian Zhao, N., T. Ghosh, and E. L. Samson. 2012. “Mapping Metropolitan Using CA Markov and Logistic Regression.” The Spatio-temporal Changes of Chinese Electric Power Egyptian Journal of Remote Sensing and Space Science 21 (3): Consumption Using Night-time Imagery.” International 229–236. doi:10.1016/j.ejrs.2017.11.006. Journal of Remote Sensing 33 (20): 6304–6320. doi:10.1080/ Townsend, A. C., and D. A. Bruce. 2010. “The Use of Night-time 01431161.2012.684076. Lights Satellite Imagery as a Measure of Australia’s Regional Zhao, N., F.-C. Hsu, G. Cao, and E. L. Samson. 2017. “Improving Electricity Consumption and Population Distribution.” Accuracy of Economic Estimations with VIIRS DNB Image International Journal of Remote Sensing 31 (16): 4459–4480. Products.” International Journal of Remote Sensing 38 (21): doi:10.1080/01431160903261005. 5899–5918. doi:10.1080/01431161.2017.1331060. Uttar Pradesh Government. 2011. “Social Demography of Uttar Zhao, N., Y. Zhou, and E. L. Samson. 2015. “Correcting Pradesh.” Accessed 21 June 2019. http://up.gov.in/Social- Incompatible DN Values and Geometric Errors in Nighttime Demography.pdf Lights Time-series Images.” IEEE Transactions on Geoscience Witmer, F. D. W. 2015. “Remote Sensing of Violent Conflict: Eyes and Remote Sensing 53 (4): 2039–2049. doi:10.1109/ from Above.” International Journal of Remote Sensing 36 (9): TGRS.2014.2352598. 2326–2352. doi:10.1080/01431161.2015.1035412. Zheng, Q., Q. Weng, and K. Wang. 2019. “Developing a New Wu, J., S. He, J. Peng, W. Li, and X. Zhong. 2013. “Intercalibration Cross-sensor Calibration Model for DMSP-OLS and of DMSP-OLS Night-time Light Data by the Invariant Region Suomi-NPP VIIRS Night-light Imageries.” ISPRS Journal of Method.” International Journal of Remote Sensing 34 (20): Photogrammetry and Remote Sensing 153: 36–47. 7356–7368. doi:10.1080/01431161.2013.820365. doi:10.1016/j.isprsjprs.2019.04.019.

Journal

GIScience & Remote SensingTaylor & Francis

Published: Nov 16, 2020

Keywords: VIIRS-DNB; DMSP-OLS; inter-calibration; machine learning; multi-layer perceptron; random forest

There are no references for this article.