Google Earth Engine for the Detection of Soiling on Photovoltaic Solar Panels in Arid Environments

remote sensing

Article Google Earth Engine for the Detection of Soiling on Photovoltaic Solar Panels in Arid Environments

1, 1,2, , 3 4 5 6 Hitesh Supe †, Ram Avtar * † , Deepak Singh , Ankita Gupta , Ali P. Yunus , Jie Dou , Ankit A. Ravankar 7 , Geetha Mohan 8, Saroj Kumar Chapagain 8, Vivek Sharma 9, Chander Kumar Singh 10, Olga Tutubalina 11 and Ali Kharrazi 12,13,14

1 Graduate School of Environmental Earth Science, Hokkaido University, Sapporo 060-0810, Japan; [email protected] 2 Faculty of Environmental Earth Science, Hokkaido University, Sapporo 060-0810, Japan 3 Department of Geography and Resource Management, The Chinese University of Hong Kong, Sha Tin, New Territories, Hong Kong, China; [email protected] 4 Graduate School of Information Science and Technology, Hokkaido University, Sapporo 060-0814, Japan; [email protected] 5 State Key Laboratory of Geohazard Prevention and Geoenvironment Protection, Chengdu University of Technology, Chengdu 610059, China; [email protected] 6 Department of Civil and Environmental Engineering, Nagaoka University of Technology, Nagaoka, Niigata 940-2188, Japan; [email protected] 7 Division of Human Mechanical Systems and Design, Faculty of Engineering, Hokkaido University, Sapporo, Hokkaido 060-8628, Japan; [email protected] 8 Institute for the Advanced Study of Sustainability, United Nations University (UNU-IAS), Tokyo 150-8925, Japan; [email protected] (G.M.); [email protected] (S.K.C.) 9 Rajasthan Renewable Energy Corporation Limited (RRECL), Jaipur 302001, Rajasthan, India; [email protected] 10 Department of Energy and Environment, TERI School of Advanced Studies, New Delhi 110070, India; [email protected] 11 Faculty of Geography, Moscow State University, Leninskiye Gory, 119991 Moscow, Russia; [email protected] 12 Advanced Systems Analysis Group, International Institute for Applied Systems Analysis, Schlossplatz 1, A-2361 Laxenburg, Austria; [email protected] 13 CMCC Foundation—Euro-Mediterranean Center on Climate Change and Ca’ Foscari University of Venice, 30175 Venice, Italy 14 Faculty of International Liberal Arts Global Studies Program, Akita International University, Yuwa City, Akita 010-1292, Japan * Correspondence: [email protected]; Tel.: +81-011-706-2261 These authors contributed equally to this work. † Received: 13 April 2020; Accepted: 30 April 2020; Published: 5 May 2020

Abstract: The soiling of solar panels from dry deposition affects the overall efficiency of power output from solar power plants. This study focuses on the detection and monitoring of sand deposition (wind-blown dust) on photovoltaic (PV) solar panels in arid regions using multitemporal remote sensing data. The study area is located in Bhadla solar park of Rajasthan, India which receives numerous sandstorms every year, carried by westerly and north-westerly winds. This study aims to use Google Earth Engine (GEE) in monitoring the soiling phenomenon on PV panels. Optical imageries archived in the GEE platform were processed for the generation of various sand indices such as the normalized differential sand index (NDSI), the ratio normalized differential soil index (RNDSI), and the dry bare soil index (DBSI). Land surface temperature (LST) derived from Landsat 8 thermal bands were also used to correlate with sand indices and to observe the pattern of sand accumulation in the target region. Additionally, high-resolution PlanetScope images were used to quantitatively validate the sand indices. Our study suggests that the use of freely available satellite data with semiautomated processing on GEE can be a useful alternative to manual methods. The developed method can provide

Remote Sens. 2020, 12, 1466; doi:10.3390/rs12091466 www.mdpi.com/journal/remotesensing Remote Sens. 2020, 12, 1466 2 of 26

Remote Sens. 2020, 11, x FOR PEER REVIEW 2 of 26 near real-time monitoring of soiling on PV panels cost-effectively. This study concludes that the DBSI effectively. This study concludes that the DBSI method has a comparatively higher potential (89.6% method has a comparatively higher potential (89.6% Accuracy, 0.77 Kappa) in the detection of sand Accuracy, 0.77 Kappa) in the detection of sand deposition on PV panels as compared to other deposition on PV panels as compared to other indices. The findings of this study can be useful to solar indices. The findings of this study can be useful to solar energy companies in the development of energy companies in the development of an operational plan for the cleaning of PV panels regularly. an operational plan for the cleaning of PV panels regularly. Keywords: land surface temperature; normalized differential sand index; soiling of solar panels Keywords: land surface temperature; normalized differential sand index; soiling of solar panels

1. Introduction 1. Introduction India holds a sixth of the world’s population and its economic engine places it as the third-largest global carbonIndia emitterholds a [sixth1,2]. Indianof the world’s policymakers population have and developed its economic the ‘Intended engine places Nationally it as the Determined third-largest Contributionsglobal carbon (INDC)’ emitter mechanism [1,2]. to Indian reduce policymakers the country’s carbon have developed footprint. To the this ‘Intended end, the National Nationally SolarDetermined Mission (NSM) Contributions is one such (INDC)’ plan tomechanism fulﬁll India’s to reduce INDC the commitments. country’s carbon Under footprint. the NSM, To Indiathis end, hasthe a mammoth National targetSolar ofMission installing (NSM) 100 is GW one of such solar plan photovoltaic to fulfill (SPV)India’s power INDC [ 1commitments.,3], i.e., 60 Gigawatts Under the (GW)NSM, in solar India parks has a and mammoth 40 GW target in solar of rooftopinstalling system 100 GW (SRS) of s byolar the photovoltaic year 2022. The(SPV) central power public [1,3], i.e., sector60 enterpriseGigawatts (PSE),(GW) Solarin solar Energy parks Corporation and 40 GW of in India solar Limited rooftop (SECI) system and (SRS) state by renewable the yearenergy 2022. The developmentcentral public organizations sector enterprise are responsible (PSE), forSolar the Energy implementation Corporation of NSM of India targets Limited at central (SECI) and and state state levelsrenewable as respective energy nodal development agencies. organizations Both SECI and are stateresponsible renewable for the energy implementation development of agenciesNSM targets coordinateat central to achieve and state the levels NSM targets. as respective SECI also nodal lays agencies. down conventions, Both SECI frameworks and state renewable and standard energy operatingdevelopment procedures agencies for all coordinate kinds of solar to achieve projects the in NSM India. targets. As of December SECI also 2019, lays the down country conventions, has reachedframeworks around 34 and GW standard of solar installations. operating procedures Figure1 illustrates for all kinds the rapid of solar increase projects in PV in installation India. As of in IndiaDecember from 20102019, tothe 2019 country [4]. Lookinghas reached at the around overall 34 numbers,GW of solar the installations. global solar Figure panel 1 installation illustrates the witnessedrapid increase exponential in PV growth, installation with cumulative in India from additions 2010 to of 2019 more [4] than. Looking 400 GW at ofthe SPV overall based numbers, capacity the fromglobal 2009 tosolar 2019 panel [4,5 ].installation In comparison witnessed to leading exponential global SPVgrowth, programs, with cumulative Germany leadsadditions the raceof more with than more400 than GW 40 of GW SPV of based installed capacity solar from power 2009 capacity, to 2019 which [4,5]. contributes In comparison to half to leading of the country’s global SPV electricity programs, consumptionGermany [leads6,7]. the race with more than 40 GW of installed solar power capacity, which contributes to half of the country’s electricity consumption [6,7].

35 Yearly Cumulative Capacity of India (GW)

5 Accomulative PV installed capacity (GW) capacity installedPV Accomulative

0 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019

Figure 1. Time series plot of installed photovoltaic capacity of India from 2010 to 2019. Figure 1. Time series plot of installed photovoltaic capacity of India from 2010 to 2019. The global operational and value chain of the SPV industry is helping to resolve the challenges of decarbonizationThe global byoperational diﬀusing and solar value technology. chain of However,the SPV industry the solar is panelhelping system’s to resolve power the outputchallenges eﬃciencyof decarbonization generally depends by diffusing on the design,solar technology. environmental, However, and climaticthe solar factors panel system’s such as orientation, power output efficiency generally depends on the design, environmental, and climatic factors such as orientation, sloping angle, shading, and weather conditions [8]. Two distinct challenges i.e., intermittence and soiling are dynamic. Intermittence is handled by forecasting and subsequent conventional

Remote Sens. 2020, 12, 1466 3 of 26 sloping angle, shading, and weather conditions [8]. Two distinct challenges i.e., intermittence and soiling are dynamic. Intermittence is handled by forecasting and subsequent conventional generation-scheduling through automation at the macro and micro levels. Soiling on solar panels, specifically dust and sand, have been one of the most underestimated factors that significantly affect the performance of solar panels [8]. The deposition of sand particles on solar panels leads to temperature fluctuations, resulting in a slump in power generation [9,10]. It has been noticed that a soiled solar farm can significantly decrease revenue [11]. Several installed solar parks in India have already started facing this issue of solar panel soiling [12]. These challenges are accentuated in arid and dry regions, which are prone to dust storms and harsh weather conditions. To solve this issue, several entrepreneurial companies with a specialization in robotics, automation, and geospatial data have started working with different solar parks in India [12,13]. The challenge of soiled solar panels opens a new research frontier to explore the timing and intensity for the cleaning of solar panels. Looking at the technical operations, Saidan and colleagues [14] revealed that due to soiling in desert areas, the efficiency of SPV can drop up to 18%. In one case study, researchers observed that the power output of SPV in the Riyadh region of Saudi Arabia was drastically reduced by up to 32% in about eight months [15]. This is because the high and continuous wind in the plain desert field leads to the formation of a sand layer on the PV panel surfaces [11]. Eventually, this drops the intensity of incident sunlight and blocks the electromagnetic waves of some specific wavelengths. As a result, both electric current (I) and voltage (V) drops, subsequently reducing power generation [16]. A study of sand deposition and its impact on the solar panel performance by Jiang et al. [17] revealed that only 22gm/m2 of dust deposition can reduce the SPV panel efficiency by 26%. A study conducted by Yap et al. [8] near Casuarina, Northern Territory, concluded that during the dry season, dust accumulation is one of the primary factors behind the system’s poor efficiency—leading to a reduction of 19.6% and 9.2% of the maximum energy output [8]. The output of an SPV system can be increased by removing the dust layer via conventional cleaning methods such as water-based cleaning, which is the most extensively used method [18]. A study conducted by India’s Council of Energy, Environment and Water (CEEW), has estimated the water requirements for weekly cleaning cycles of solar plants in the country ranges between 7000 and 20,000 liters per MW per wash [19]. Furthermore, cleaning with water is both labor-intensive and carries the risk of physical damages due to human error. The scarcity of water in arid and desert regions adds another challenging dimension and for a water-stressed country like India, this overlaps with the sustainability challenges of livelihoods and agricultural priorities [20]. As of March 2020, Bhadla is one of the world’s largest solar parks with 2245 MW capacity. It is located in the Thar Desert of Rajasthan, India, and brings some operational challenges vis-a-vis the terrain and the landscape. The gusty winds with soil erosion at the rate of 7.67 g/m2/day can pose a serious threat to the installed solar farms [18]. Three days of soiling effects are capable of dropping the solar farm’s efficiency by one-fourth of its original capacity [19,20]. This makes the cleaning of solar panels the inevitable option for the plant operators. Water cleaning is the most common option, but the shortage of water resources in Rajasthan state handicaps the conventional cleaning methods. Furthermore, the water diversion canals built for the cleaning of solar panels in Bhadla have drawn criticism from the farmers and other environmental activists and groups. Due to the severe water scarcity issues in this region of India, the huge quantum of water cannot be guaranteed for cleaning purposes. Furthermore, any future water supplies will have to adhere to the dynamics of a tariff structure to avoid ‘the tragedy of the commons’ [21]. This can lead to tariff precariousness in some parts of India, where water tariff rates have been hiked 100 times for industries [22]. A similar escalation in water tariffs in Bhadla can drastically influence the economics of solar plant operations. The limits of water as a cleaning agent for solar panels call for the search of a water-efficient or water-free cleaning mechanism aided by image-based surveillance tools collected from drones and satellites. This will require the identification of dust and soil spots on panels with precision and accuracy. It can help in both scheduling and optimization of panel cleaning inventories across the vast Remote Sens. 2020, 12, 1466 4 of 26 tracts of solar parks. In this context, the use of drones helps collect images of inaccessible areas with high-resolution data. A few studies have used photographic imagery to detect sand deposition on panels by using drones [8]. In the absence of drone technology, the satellite remote sensing techniques have been widely used in a variety of fields like the monitoring of soil types, geomorphology, land use/land cover, and the monitoring of rice crops [23,24]. Recently, with improved radiometric performance and higher spatial and temporal resolutions, satellite images have become indispensable tools in a variety of fields. Several studies discuss the use of satellite images in detecting and extracting soil information [25–27]. Such methods can be further applied to solar farms to detect the deposition of sediments on the panels. However, there are no specific satellite-based indices to monitor soiling on solar PV. Robotic panel cleaning can reduce the manual and water costs. However, the excessive use of robots can affect the panel’s lifespan. It can potentially add to more secondary capital costs (with the replacement of damaged panels) and deplete the profit margins. This requires robotics arm cleaning to be supervised with the exact location of dust and dirt markups on the panels. It can further help in efficiently scheduling the cleanup of the panels. The purpose of this study is to introduce the application of the Google Earth Engine (GEE) in monitoring the soiling of solar panels. In this study, we explored various soil and sand-based indices using various satellite data to monitor soiling in PV panels. The GEE is an open-source platform with great potential in expanding the research frontiers. The GEE has also been used in various applications such as monitoring of rice extent, cropping patterns, and growth stages [28]. Archived Landsat 8 and Sentinel-2 images from September 2017 to February 2019 are used in the GEE for the analysis. The approach used in this study is a combination of data retrieval, image processing, outcome analysis, and visual interpretation. The periods with a high amount of sand deposition can be determined through satellite image analysis, which can further help to optimize the use of water and self-cleaning robots.

2. Study Area The state of Rajasthan, India has a vast unused, barren, and aﬀordable land that has solar irradiation of 5.72kWh/m2/day, the highest in the country, and thus making it a very suitable site for solar park development [29]. Rajasthan has about 20 million ha (208,110 km2) of desert land, which is 60% of its total geographical area of 342,300 km2 [30,31]. Moreover, this area receives scanty rainfall, hence good sunshine is available throughout the year [30]. The study area focuses on Bhadla solar park in the Jodhpur district of Rajasthan, India, as shown in Figure2. This solar park is a 2.25 GW project located at 27.5015 N latitude and 71.9358 E longitude, approximately 220 km away from Jodhpur headquarters on the Bap-Bhadla road, as displayed in Figure2. For the development and operations of Bhadla Solar Park, the central PSE-SECI is coordinating with its state-level counterpart—Rajasthan Renewable Energy Corporation Limited (RRECL). The project construction was started in July 2015, spanning a total area of 5,783 ha [29]. About 10,000 ha of government-owned land has been allocated for this solar park development. Figure2 highlights four sections of the Bhadla solar park (P1, P2, P3, and P4 marked with red boundaries) which are considered for this study. Diﬀerent sections and phases of these plants are owned by various public and private sector companies acquired through bidding and auctions. SECI, RRECL, Power Grid Corporation of India Ltd. (PGCIL), state power generation, transmission and distribution companies along with the park developers are the key stakeholders in the Bhadla Solar Project. SECI through RRECL coordinates with all the stakeholders for the development and operations of this solar park [32]. RemoteRemote Sens.Sens.2020 2020,,12 11,, 1466 x FOR PEER REVIEW 55 ofof 2626

Figure 2. Location of Bhadla solar park in the state of Rajasthan in India and location of four photovoltaic Figure 2. Location of Bhadla solar park in the state of Rajasthan in India and location of four (PV) parks (P1, P2, P3 and P4). Red square in lower left panel indicates the location of the study area in photovoltaic (PV) parks (P1, P2, P3 and P4). Red square in lower left panel indicates the location of Rajasthan State. the study area in Rajasthan State. The ultra-mega scale Bhadla solar park covering 37 plants spanning into four different phases, offersThe an excellentultra-mega opportunity scale Bhadla to solar study park the soilingcovering phenomenon 37 plants spanning using remote into four sensing different technology. phases, Theoffers monthly an excellent mean climaticopportunity conditions to study in the the Bhadla soiling region phenomenon of Rajasthan using is shownremote in sensing Figure 3technology.. Figure3a illustratesThe monthly the mean plot of climatic the monthly conditions average in temperaturethe Bhadla region and irradiance. of Rajasthan Due is toshown the cloudy in Figure atmosphere, 3. Figure the3a monthsillustrate ofs Junethe plot and Julyof the show monthly the lowest average irradiance. temperat Thisure has and been irradiance. confirmed byDue visual to the inspection cloudy ofatmosphere, satellite images the m availableonths of on June the Unitedand July States show Geological the lowest Survey irradiance. (USGS) This Earth has Explorer. been confirmed The average by monthlyvisual inspection mean precipitation of satellite throughout images available the year on is aboutthe United 21 mm. States As the Geological rainfall is Survey low and (USGS inconsistent) Earth throughoutExplorer. The the a year,verage the monthly monthly mean average precipitation precipitation, throughout therefore, the ranges year from is about 2 mm 21 to mm 84 mm. As and the peaksrainfall around is low Julyand andinconsistent August [throughout33]. Throughout the year the, majoritythe monthly of the average year, thisprecipitation area is subjected, therefore, to westerlyranges from and 2 north-westerly mm to 84 mm winds.and peak Figures around3b shows July and the August monthly [33] average. Throughout wind speed the m whichajority varies of the fromyear, 2.2 this m / areas to 4.6 is m subjected/s [34]. The to highestwesterly wind and speed north is-westerly observed wind in thes. monthFigure of 3b July. shows The fine the particlesmonthly ofaverage sand carried wind speed by the which wind varies are deposited from 2.2on m/s the to surfaces4.6 m/s [34] exposing. The highest the solar wind panel speed farm is throughoutobserved in thethe season.month Furthermore, of July. The there fine particles are no man-made of sand carriedor natural by windbreakersthe wind are such deposited as buildings, on the trees, surfaces and shrubsexposing to the block solar or panel alter the farm movement throughout of the season winds.. Furthermore Without the, cover there ofare vegetation no man-made and or high-rise natural buildingswindbreaker in thes such face ofas winds,buildings there, trees is a, continuous and shrubs process to block of duneor alter deposition. the movement The regular of the exposure winds. toWithout winds makesthe cover the of panels vegetation vulnerable and tohigh soiling.-rise buildings Figure4 shows in the the face field of winds photographs, there is of a solar continuous power panelsprocess collected of dune during deposition. the field The visit regular to the exposure study area. to winds Figure 4makesa,b illustrate the panels clean vulnera PV solarble panels to soiling. and thoseFigure soiled 4 shows by sandy the field deposition, photographs respectively. of solar power panels collected during the field visit to the study area. Figure 4a,b illustrate clean PV solar panels and those soiled by sandy deposition, respectively.

RemoteRemote Sens. Sens. 20202020,, 1112,, x 1466 FOR PEER REVIEW 66 of of 26 26 Remote Sens. 2020, 11, x FOR PEER REVIEW 6 of 26

(a) 50 (a) 10 50 10 45 Temp. (°C) 45 Temp. (°C) 9 40 9 40 Average Direct Normal Irradiance (kWh/m2/day)

C) Average Direct Normal Irradiance (kWh/m2/day) 8 ° 35 C) 8 ° 35 30 30 7 7 25 25 6 20 6 20 15 5 15 5 10 10 irradiance average Monthly 4 irradiance average Monthly Monthly average Temp( average Monthly 5 4 Monthly average Temp( average Monthly 5 0 3 0 3 J F M A M J J A S O N D J F M A M J J A S O N D

Figure 3. Monthly average climatology bar graph for Bhadla region of Rajasthan. (a) monthly average Figure 3. MMonthlyonthly average climatology bar graph for Bhadla region of Rajasthan. (a) monthly average temperature (bar graph) and irradiance (thick line); and (b) graph showing monthly average wind temperaturetemperature (bar graph)graph) and irradianceirradiance (thick(thick line);line); and ((bb)) graphgraph showingshowing monthlymonthly averageaverage windwind speed (dashed line) and rainfall values (bar graph). Data source: [33,34]. speed (dashed(dashed line)line) andand rainfallrainfall valuesvalues (bar(bar graph).graph). DataData source: source: [ [3333,3,344]]..

Figure 4. (a) Photographs showing the photovoltaic solar panels during the clean condition and (b) (a) Clean PV solar panels (b) Unclean PV solar panels during the(a) sandy Clean deposition PV solar stagepanels in the Bhadla Solar Park, Jodhpur(b) Unclean District, RajasthanPV solar panels State, India. Figure 4. (a) Photographs showing the photovoltaic solar panels during the clean condition and (b) Figure 4. (a) Photographs showing the photovoltaic solar panels during the clean condition and (b) during the sandy deposition stage in the Bhadla Solar Park, Jodhpur District, Rajasthan State, India. during the sandy deposition stage in the Bhadla Solar Park, Jodhpur District, Rajasthan State, India.

Remote Sens. 2020, 12, 1466 7 of 26

3. Materials and Methods This study attempts to provide an alternate model for scheduling the cleaning of solar panels. The field survey was conducted to collect the primary data on the soiling of solar panels, understand the cleaning approaches, and accumulate other firsthand information from the ground. To understand the plant operation, we conducted a total of eight interviews with the project managers and other operating staffs in four different solar plants at Bhadla. To estimate the extent of soiling on the panels between September 2017 and February 2019, we performed a time-series analysis of various soil indices using remote sensing technology (see following subsections). We aim to analyze and compare the multitemporal images to correlate the climatic effects with the soiling phenomenon.

3.1. Satellite Data Readily available Landsat 8 and Sentinel-2 satellite data in the GEE platform were used for monitoring with a time frame from September 2017 to February 2019. The Landsat 8 launched in February 2013, is a joint National Aeronautics and Space Administration (NASA)/USGS program. Similarly, Sentinel-2 is an Earth observation mission of the European Union Copernicus Program that systematically acquires optical imagery at high spatial resolution. In this study, monthly median time series data were obtained from the Landsat 8 and Sentinel-2 image collection. The list of images used in this study and acquisition details are provided in Table1. To automate the satellite image analysis processes, we used the ‘Google Earth Engine’ (GEE) of Google Inc., an open cloud-based geospatial processing platform, designed mainly for planetary-scale environmental data analysis. The GEE platform combines a multi-petabyte catalog of satellite imagery and geospatial datasets, which allow users to visualize, manipulate, edit and create spatial data in an easy and fast way [35]. GEE incorporates a wide range of spatial manipulation tools which allows scientists, researchers, and developers to detect changes, map trends, and quantify diﬀerences on the Earth’s surface [36]. The GEE can directly import Landsat 8 and Sentinel-2 images from its respective service provider and can be used for real-time processing.

Table 1. List of satellite data products used in this study and their date of acquisition. Sentinel-2 data includes monthly median values of the images.

Date of Acquisition Data Source (YYYY.MM.DD)/ Number of Images (YYYY.MM) 2017.09.09; 2017.09.25; 2017.10.11; 2017.10.27; 2017.11.12; 2017.11.28; 2017.12.30; 2018.01.15; 2018.01.31; 2018.02.16; Landsat 8 (30 meters) 2018.03.20; 2018.04.05; 2018.04.21; 2018.05.07; 2018.05.23; 25 2018.06.08; 2018.06.24; 2018.07.10; 2018.08.11; 2018.09.12; 2018.10.14; 2018.11.15; 2018.12.17; 2019.01.18; 2019.02.03 2017.09; 2017.10; 2017.11; 2017.12; 2018.01; 2018.02; 2018.03; Sentinel-2 (10 meters) 2018.04; 2018.05; 2018.06; 2018.07; 2018.08; 2018.09; 2018.10; 18 2018.11; 2018.12; 2019.01; 2019.02; PlanetScope (3 meters) 2018.01.31 1

The Landsat 8 satellite data consists of 12 spectral bands out of which, band 3 (green band), band 4 (red band) and band 7 (short wave infra-red 2 bands) were used to develop normalized difference sand index (NDSI) and ratio normalized difference soil index (RNDSI) with a spatial resolution of 30 m. While in the Sentinel-2 satellite data, band 3 (green band) and band 12 (short wave infra-red 2), having respectively a spatial resolution of 10 meters and 20 meters, were used to calculate the dry bare soil index (DBSI). The DBSI also involved using band 4 (red) and band 8 (vegetation infrared edge) to generate the normalized differential vegetation index (NDVI). The data acquisition and semi-automated processing of the images were completed on the GEE platform and the direct results of the images were extracted from the cloud-based platform. Table2 illustrates the specification and band details of Landsat 8, Sentinel-2 and Planet Scope satellites. Remote Sens. 2020, 12, 1466 8 of 26

Table 2. Band speciﬁcations of satellite sensors used in this study.

Sensors Band Bands Spatial Resolution Band Range Radiometric Revisit Cycle Number (Wavelength) (Meters) (µm) Resolution (bit) (Days) A. Landsat 8 16 16 2 Blue 30 0.45–0.51 3 Green 30 0.53–0.59 4 Red 30 0.64–0.67 5 Near-Infrared 30 0.85–0.88 6 SWIR 1 30 1.57–1.65 7 SWIR 2 30 2.11–2.29 10 TIR Sensor 1 100 10.6–11.19 11 TIR Sensor 2 100 11.5–12.51 B. Sentinel-2 12 5 3 Green 10 0.54–0.57 4 Red 10 0.65–0.68 8 Near-Infrared 10 0.78–0.89 11 SWIR 1 20 1.56–1.65 12 SWIR 2 20 2.10–2.28 C. Planet Scope 12 Daily 1 Blue 3 0.455–0.515 2 Green 3 0.5–0.59 3 Red 3 0.59–0.67 4 Near-Infrared 3 0.78–0.86

The validation of results was performed by adopting two customized techniques. Firstly, the land surface temperature (LST) generated from the Landsat 8 thermal bands were compared with results from the soil indices. Here the change in LST was expected for the region aﬀected by the soiling phenomenon. Secondly, the PlanetScope high-resolution satellite images from “Planet Labs” were observed to view the soiling interfaces. PlanetScope data consists of the composite image of the red, green, blue (RGB) bands, which provides an actual view of the scene. Its spatial resolution is as high as 3 meters and is capable of daily acquisitions. We also performed an accuracy assessment using quantitative information of various indices against LST and PlantScope data.

3.2. Methodology Figure5 depicts the methodology deployed in this paper in the form of a ﬂowchart. The study commenced by using GEE to detect soiling on solar PV panels by utilizing diﬀerent types of indices based on the Landsat 8 and Sentinel-2 satellite data. Remote Sens. 2020, 12, 1466 9 of 26 Remote Sens. 2020, 11, x FOR PEER REVIEW 9 of 26

Figure 5. Flowchart of the study employed for the detection of soiling on PV solar panels. Figure 5. Flowchart of the study employed for the detection of soiling on PV solar panels. 3.2.1. Normalized Difference Sand Index (NDSI) 3.2.1. Normalized Difference Sand Index (NDSI) The normalized differential sand index (NDSI) was previously used for monitoring, mapping, and assessingThe n theormalized sand dune differential encroachment sand i inndex some (NDSI sites) inwas the previously northern centralused for part monitoring, of Iraq over mapping, 21 years fromand assessing 1988 to 2009 the [sand27]. Thisdune index encroachment is helpful inin thesome selection sites in of the areas northern suitable central for control part of measures Iraq over [ 2721]. Similarly,years from in 1988this study, to 2009 NDSI [27] was. This generated index isand helpful applied in the in thisselection field to of identify areas suitable the amount for controlof sand depositionmeasures [27] on. solar Similarly, panels. in Itth isis computedstudy, NDSI as was the ratiogenerated of the and measured applied intensities in this field in theto identify short-wave the infraredamount (SWIRof sand 2) deposition and red (R) on spectral solar panels bands,. It usingis computed Equation as (1).the ratio of the measured intensities in the short-wave infrared (SWIR 2) and red (R) spectral bands, using Equation (1). NDSI = (SWIR 2 Red)/(SWIR 2 + Red) (1) NDSI = (SWIR 2 – R−ed)/(SWIR 2 + Red) (1) 3.2.2. Ratio Normalized Difference Soil Index (RNDSI) 3.2.2. Ratio Normalized Difference Soil Index (RNDSI) Apart from the sand index discussed above, we used another index to accomplish the purpose. The RNDSIApart from can characterizethe sand index the discussed sandy desert, above, which we used can helpanother in determiningindex to accomplish the level the of purpose. fine sand deposition.The RNDSI Itcan was characterize developed the for sandy separating desert soil, which from can impervious help in determining surface areas the andlevel vegetation of fine sand so thatdeposition. it can serve It was as develo an inputped for for the separating land use soil/land from cover impervious model (LULC) surface [26 area]. Unlikes and vegetation vegetation, so string that spectralit can serve responses as an input do not for exist the forland soil us duee/land to itscover complex model physical (LULC) and[26]. chemical Unlike vegetation, composition string and regionalspectral diresponsesfferences do [26 not]. The exist two for indices soil due used to areits normalizedcomplex physical differential and chemical soil index composition and tasseled and cap transformationregional differences 1 (TC1) [26] [.26 The]. Here, two indices RNDSI used takes are into normalized consideration differential the brightness soil index eff andect t ofasseled the image cap thattransformation is derived from1 (TC1) the [26] tasseled. Here, cap RNDSI transformation takes into consideration [26]. As a first the task, brightness the soil effect index of is the calculated image usingthat is a derived green band from and theshort tasseled wave cap infra-red transformation 2 bands [26] from. As the aLandsat first task, 8 satellitethe soil data.index is calculated using a green band and short wave infra-red 2 bands from the Landsat 8 satellite data. NDSI2 = (SWIR 2 Green)/(SWIR 2 + Green) (2) NDSI2 = (SWIR 2 - Green)− / (SWIR 2 + Green) (2)

TheThe above above result result i.e. i.e.,, NDSI NDSI2 is2 furtheris further transformed transformed to normalized to normalized NDSI2 or NDSI NNDSI2 or2 ( NNDSIsee equation2 (see Equation3). Then (3)).the NNDSI Then the2 is NNDSI divided2 is by divided a tasseled by a tasseledcap transformation cap transformation brightness brightness factor factor(NTC) (NTC).. The Themathematical mathematical expression expression of RNDSI of RNDSI is as isfollows as follows::

NNDSI2 = (NDSI2– NDSI2min) / (NDSI2max– NDSI2min) (3) NNDSI = (NDSI NDSI )/(NDSI NDSI ) (3) 2 2 − 2min 2max− 2min and,

Remote Sens. 2020, 12, 1466 10 of 26

and, NTC1 = (TC1 TC1 )/(TC1max TC1 ) (4) − min − min RNDSI = (NNDSI2)/(NTC1) (5)

NDSI = 1; NDSI = 1; TC1 = 10,000; TC1max = 50,000. 2min − 2max min The range of RNDSI values depends on the TC1min and TC1max values. In this case, the range of values selected for tasseled cap transformation brightness was from 10,000 to 50,000. Consequently, the ﬁnal output, that is, the RNDSI has values that vary from 0 to 3. The higher the value of RNDSI, the higher is the amount of soiling.

3.2.3. Land Surface Temperature (LST) Landsat 8 thermal bands (band 10 and 11) measure the energy emitted from the earth’s surface and can be used to retrieve the LST. The presence of sand particles on PV panels can show the change in LST [37]. The LST depends on the land surface type because of variations in the emissivity from heterogeneous surfaces such as vegetation cover, built-up, and bare soil. The clean PV panels and arid sand dunes are continuously exposed to sunlight for a long time. This makes the surface temperature of panels and bare lands to rise. However, when fine sand particles carried by a sandstorm or wind are accumulated on the panel surface, the overall panel surface temperature changes. Wind carried sand particles have comparatively low temperatures; therefore, settlements of sand particles make the region cooler. With this principle, a variation on solar panel surface value can be linked with the presence of the sand layers. However, as the Landsat 8 derived LST has a coarse resolution in comparison to the size of the PV panels, we analyze the pattern of the LST rather than specific values. In this study, a two-step process was used to generate LST from the Landsat 8 data [38]. In the first step, the atmospheric correction was carried out by computing necessary atmospheric parameters. In the second-step, land surface emissivity was calculated based on studies carried by [39–41]. Furthermore, LST was obtained by using TIRS band 10 with a mono-window algorithm [42].

3.2.4. Dry Bare Sand Index (DBSI) The Dry bare soil index (DBSI) is used to identify bare soil from arid and semiarid regions—especially areas having a dry climate. A recent study [25] developed this index to map built-up and bare areas in dry climates from the Landsat 8 data. The DBSI algorithm was used in determining barren land and differentiating it from other built-up areas [25]. In this study, the same algorithm was implemented on GEE values by using Sentinel-2 data. Sentinel-2 images have higher spatial resolution compared to Landsat; hence, it was more precise in detecting sand deposition layers. The proposed equation for bareness area in a dry climate is the inverse of the modified normalized difference water index [25].

DBSI = (SWIR 1 - Green)/(SWIR 1 + Green) NDVI − (6) Where, NDVI = (NIR - Red)/(NIR + Red)

The DBSI results’ values range from 2 to +2, and higher numbers represent a high degree of the − bareness of soil [25]. In previous studies, the threshold value was used for diﬀerentiating bare soil and non-bare soil areas in the city of Erbil, Iraq [25]. Based on a test carried out with a sample of bare soil pixels, a DBSI value of 0.26 and higher was delineated as bare soil [25]. With this background, we used DBSI with the assumption that it can diﬀerentiate the soiled pixels and clean pixels in our study area. Remote Sens. 2020, 12, 1466 11 of 26

4. Results

4.1. Spatial Correlation between NDSI, and RNDSI with LST NDSI, RNDSI and LST are mapped over the spatial domain to detect soiling on the panels. Figure6 shows the images taken during September 2017. Consequently, this year the panels were expectedRemote to Sens. receive 2020, 11 a, x highFOR PEER level REVIEW ofsoil deposition since self-cleaning systems were not11 of 26 installed (in the initial stages). In the initial stages, the cleaning process was carried manually using water. The Ecoppiaexpected E4 to panelreceive cleaning a high level robots of soil were deposition installed since andself-cleaning operational systems on were some not of installed the panels (in since the initial stages). In the initial stages, the cleaning process was carried manually using water. The February 2018 [43]. The company deployed around 2000 units of Ecoppia E4 robots in two diﬀerent Ecoppia E4 panel cleaning robots were installed and operational on some of the panels since February phases2018 of the [43] park. The [company43]. These deployed robots around move over2000 units the surfaceof Ecoppia of theE4 robots panels in andtwo different sweep awayphases dust of from it [44].the Figure park6 [43]a,c.show These NDSIrobots move values over based the surface on Landsat of the panels 8 data, and acquired sweep away in September dust from it 2017. [44]. High valuesFigure correspond 6a,c show to the NDSI high values level based of sand on Landsat deposition 8 data, on ac thequired surface in September of panels 2017. [27 ].High Some values part of the study areacorrespond shows to low the valueshigh level of of NDSI, sand whichdeposition corresponds on the surface to theof panels clean [27] panels.. Some Uneven part of the distribution study of valuesarea can shows be seen low over values panel of NDSI surfaces., which Thesecorresponds variations to the clean with panels. high values Uneven can distribution be visually of values ascribed as can be seen over panel surfaces. These variations with high values can be visually ascribed as sand sand deposition. The sand deposition also aﬀects temperature. Hence, a change in temperature is also deposition. The sand deposition also affects temperature. Hence, a change in temperature is also expectedexpected [37]. Figure[37]. Figure6b,d 6 showb,d show the the LST LST based based on on thethe Landsat Landsat 8 data. 8 data. The The visual visual comparison comparison between between the landthe surface land surface temperature temperature and and NDSI NDSI of of the the studystudy area area can can be be seen seen in Figure in Figure 6. 6.

(a) L8 NDSI (9th September 2017) (b) L8 LST (9th September 2017)

FigureFigure 6. Visual 6. Visual comparison comparison of sandyof sandy deposition deposition on on aa solar solar farm farm (described (described within within the black the border black). border). High values of normalized difference sand index (NDSI) shown in (a) and (c) within the black border High values of normalized diﬀerence sand index (NDSI) shown in (a) and (c) within the black border indicates the location of sandy deposition. Landsat 8 generated land surface temperature (LST) indicatespattern the locationduring the of same sandy period deposition. shows the Landsat lowest temperature 8 generated (bluish land tints) surface for temperaturethe sandy deposition (LST) pattern duringareas thesame on top period of the panels shows (b the) and lowest (d). temperature (bluish tints) for the sandy deposition areas on top of the panels (b) and (d). The pattern of LST variation is quite similar to the variation of pixels values on the panel surface. Regions having low temperature in the LST map coincide with regions having high NDSI values. A

Remote Sens. 2020, 12, 1466 12 of 26

The pattern of LST variation is quite similar to the variation of pixels values on the panel surface. Regions having low temperature in the LST map coincide with regions having high NDSI values. Remote Sens. 2020, 11, x FOR PEER REVIEW 12 of 26 A case study experiment done by Márquez and Ramírez [37], reveals that thermographic analysis is ancase eﬀ studyective experiment method in done the detection by Márquez of accumulated and Ramírezdust [37], particlesreveals that carried thermographic by the wind analysis on the is solar an panels.effective Based method on emissivity in the detection analysis, of the accumulated value of the dust emissivity particles of carried dust was by revealedthe wind to on be the very solar low. Aspanels. a result, Based the on brightness emissivity temperature analysis, the of value the surfaceof the emissivity falls. Hence, of dust the was temperature revealed decreasesto be very whenlow. theAs dusta result, and the sand brightness gets accumulated temperature [37 ],of and the undersurface the falls. larger Hence, areas the with temperature dust, the temperaturedecreases when dips furtherthe dust [37 and]. A sand similar gets pattern accumulated is observed [37], and in the under results the shownlarger areas in Figure with6 .dust Correspondingly,, the temperature Figure dips7 showsfurther the [37] NDSI. A similar of Landsat pattern 8 imagesis observed acquired in the in results May 2018shown and in JuneFigure 2018 6. Correspondingly which have a lower, Figure sand deposition.7 shows the The NDSI solar of Landsat power plant 8 images and acquired its surrounding in May area2018 showand June a low 2018 variation which have in NDSI a lower values sand on thedeposition. surface ofThe panels solar comparedpower plant to and images its surrounding taken in September area show 2017. a low The variation NDSI andin NDSI LST values results on are indicativethe surface of of uniformity panels compared along these to images surfaces. taken This in implies September that the 2017 panels. The were NDSI cleaner and LST in 2018 results because are ofindicative the useof of robots uniformity or by other along means these [surfaces.45]. This implies that the panels were cleaner in 2018 because of the use of robots or by other means [45].

(a)L8 NDSI (23rd May 2018) (b) L8 LST (23rd May 2018)

(c)L8 NDSI (8th June 2018) (d) L8 LST (8th June 2018)

FigureFigure 7. 7.Visual Visual comparison comparison of of sand sand deposition deposition on theon solar the solar farm farm based based on NDSI on andNDSI LST and after LST cleaning. after Thecleaning ﬁgure. The shows figure a similarity shows betweena similarity NDSI between (Figure 7NDSI( a), 7( (cFigure)) with 7sand(a), deposition7(c)) with variationsand deposition and LST patternvariation of and Figure LST7( patternb), 7(d), of respectively. Figure 7(b), 7(d), respectively.

WeWe furtherfurther investigatedinvestigated RNDSIRNDSI toto seesee thethe patternpattern ofof soilingsoiling inin thethe studystudy area,area, asas shownshown inin FigureFigure8 . The8. The results results from from RNDSI RNDSI show show similar similar outcomes outcomes as as shown shown inin NDSINDSI based analysis analysis.. Since Since RNDSI RNDSI is is a ratio of soiling index and brightness, it dampens the built-up effect and emphasizes the soiling effect. The images acquired in December 2017 and January 2018, show high values of RNDSI varying between 0.99 and 1.87, as shown in Figure 8a,c. On the other hand, during May 2018 and June 2018, the values of RNDSI show lower values, which vary between 0.61 and 0.82, as shown in Figure 8e,g.

Remote Sens. 2020, 12, 1466 13 of 26 a ratio of soiling index and brightness, it dampens the built-up eﬀect and emphasizes the soiling eﬀect. The images acquired in December 2017 and January 2018, show high values of RNDSI varying between 0.99 and 1.87, as shown in Figure8a,c. On the other hand, during May 2018 and June 2018, the values of RNDSI show lower values, which vary between 0.61 and 0.82, as shown in Figure8e,g. A similarRemote distribution Sens. 2020, 11, x pattern FOR PEER of REVIEW NDSI values are observed in 2017 and 2018, displayed13 in of Figures26 6 and7. InA Figuresimilar 8distribution, the RNDSI-generated pattern of NDSI greyscale values are rasterobserved images in 2017 highlight and 2018, some displayed bright in Figure clusterss 6 in the study area,and 7 which. In Figure represents 8, the RNDSI highly-generated soiled g regions.reyscale raster We further images comparedhighlight some the bright mean clusters value i ofn the NDSI and RNDSIstudy in the area, study which site represents from September highly soil 2017ed region to Februarys. We further 2019 compared as shown the in mean Figure value 10. of NDSI and RNDSI in the study site from September 2017 to February 2019 as shown in Figure 10.

(a) RNDSI 30th December 2017 (b) L8 LST 30th December 2017

(e) RNDSI 23rd May 2018 (f) L8 LST 23rd May 2018

Figure 8. Cont.

Remote Sens. 2020, 12, 1466 14 of 26 Remote Sens. 2020, 11, x FOR PEER REVIEW 14 of 26

(g) RNDSI 8th June 2018 (h) L8 LST 8th June 2018

FigureFigure 8. Landsat 8. Landsat 8 spatial 8 spatial distribution distribution of of soilingsoiling pattern patternss using using ratio ratio normalized normalized difference diff erencesoil index soil index (RNDSI) along with LST at different times after the installation of panels. The figure shows results of (RNDSI) along with LST at different times after the installation of panels. The figure shows results of RNDSI in Figure 8(a), 8(c), 8(e), 8(g) with LST in Figure 8(b), 8(d), 8(f), 8(h) respectively. RNDSI in Figure8( a), 8(c), 8(e), 8(g) with LST in Figure8( b), 8(d), 8(f), 8(h) respectively. Although Landsat 8 based LST cannot provide precise information about exact temperatures, it Althoughhas been widely Landsat used 8 basedto see the LST temperature cannot provide pattern and precise their variations. information When about we consider exact temperatures, entire it has beensolar farm widelys, certain used variations to see theof temperature temperature patterns pattern were andseen theiron the variations.panel surfaces When. Due to we the consider entire solarcoarse farms, resolution certain of LST variations data when of temperaturecompared to the patterns size of were solar seen PV panels, on the in panel this study surfaces., the Due to the coarsetemperature resolution pattern of LST was dataqualitatively when compared analyzed rather to the tha sizen quantified of solar as PV a specific panels, temperature in this study, the temperaturerange [37] pattern. Continuous was qualitatively exposure of clean analyzed PV panels rather and fine than sand quantified particles to assolar a specificlight for a temperature long time causes an increase in the temperature of sand particles [46]. However, when the fine sand range [37particles]. Continuous carried over exposure through ofa sandstorm clean PV or panels wind andare accumulated fine sand particles on the panel to solarsurface light, the foroverall a long time causespanel an increase surface in temperature the temperature changes. of Furthermore, sand particles the [46 wind]. However, carrying whenthe sand the particles fine sand has particles carriedcomparatively over through low aer sandstorm temperature ors, and wind therefore, are accumulated the settlements on of the sand panel particles surface, make thethe region overall panel surfacesignificantly temperature cool changes.er [37]. Subsequently Furthermore,, through the the wind use carryingof this principle, the sand a comparison particles of has the comparatively results lower temperatures,with LST was helpful and therefore,for validation. the The settlements variation of oftemperature sand particless observed make in 2018 the–2019 region indicate significantlys a more uniform texture when compared to 2017. This implies that there was less sand deposition in cooler [37]. Subsequently, through the use of this principle, a comparison of the results with LST 2018–2019. The inference of this observation is also closely related to the results obtained from the was helpfulNDSI. for validation. The variation of temperatures observed in 2018–2019 indicates a more uniform texture when compared to 2017. This implies that there was less sand deposition in 2018–2019. The inference4.2. Sand of Layer this D observationetection Using isDBSI also closely related to the results obtained from the NDSI. Figure 9 indicates the result based on the dry bare soil index (DBSI) using Sentinel-2 data. We 4.2. Sandtried Layer various Detection values Usingof DBSI DBSI to identify the threshold values to detect soiling on the PV panels. The Figurethreshold9 indicates value of the DBSI result is 0.26 based, as onhigh thelighted dry barein red soil in indexFigure (DBSI)9. It is usingan easy Sentinel-2 and clear data.way of We tried information extraction and analysis. Visually, it can be inferred through the results as some patches variousand values strips of of DBSI affected to identify areas are the noticeable threshold. In values January, to February,detect soiling March on, and the May PV panels. of 2018, The many threshold value ofconcerned DBSI is areas 0.26, are as readily highlighted visible. However in red in, the Figure results9 .for It other is an months easy and, such clear as June way 2018, of August information extraction2018, and November analysis. 2018 Visually,, and February it can be2019 inferred, reveal clean through panels the with results the except as someion patchesof some bright and strips of affectedpatches areas.are noticeable. In January, February, March, and May of 2018, many concerned areas are readily visible. However, the results for other months, such as June 2018, August 2018, November 2018, and February 2019, reveal clean panels with the exception of some bright patches.

Remote Sens. 2020, 12, 1466 15 of 26 Remote Sens. 2020, 11, x FOR PEER REVIEW 15 of 26

(a) DBSI January, 2018 (b) DBSI February, 2018

(e) DBSI June, 2018 (f) DBSI August, 2018

Figure 9. Cont.

Remote Sens. 2020, 11, x FOR PEER REVIEW 16 of 26 Remote Sens. 2020, 12, 1466 16 of 26 Remote Sens. 2020, 11, x FOR PEER REVIEW 16 of 26

(g)( gDBSI) DBSI November November, 201, 20188 ((hh) )DBSI DBSI February February, 201, 2019 9

FigureFigureFigure 9. 9.Sentinel-2 Sentinel 9. Sentinel-2 Spatial-Spatial2 Spatial distribution distribution of ofof soiling soiling pattern patternsss usingusing using dry dry dry bare bare bare soil soil soilindex index index (DBSI (DBSI (DBSI)) at) differentat different at diﬀ erent timestimestimes after after after the the installation the installation installation of panels. of of panels. panels. Images ImagesImages highlight highlight highlight the athetheﬀected affectedaffected area area that area thatis that distinguished is is distinguished distinguished by applying by by thresholdapplyingapplying value.threshold threshold Figure value. value.9( a Figure), Figure 9(b), 9( 9( 9(aca),),), 9( 9(9(bdb),)), 9( show9(cc),), 9(9( highd) show soiling highhigh comparedsoiling soiling compared compared to Figure to toFigure9 (Figuree), 9(9(fe ),9(), 9(e9(),gf ),9(), f 9(),h ). 9(g),9( 9(g),h 9(). h). 4.3. Time Series Behaviour of Sand Indices 4.3.4.3. Time Time Series Series Behavio Behaviouru ofr ofS andSand I ndicesIndices Along with the visual interpretation, we also performed a zonal statistical calculation to display Along with the visual interpretation, we also performed a zonal statistical calculation to display the quantitativeAlong with informationthe visual interpretation for clarity., Thewe also graphical performed representation a zonal statistical of NDSI, calculation RNDSI, to anddisplay DBSI the quantitative information for clarity. The graphical representation of NDSI, RNDSI, and DBSI was the quantitative information for clarity. The graphical representation of NDSI, RNDSI, and DBSI was wasable able to to better better track track the behavior the behavior of soiling of soilingsince the since period the when period the plant when was the established. plant was The established. four- able to better track the behavior of soiling since the period when the plant was established. The four- The four-studystudy site of site the ofsolar the farm solar as farm shown as in shown Figure in 2 Figurewas considered2 was considered as one unit as and one monitored unit and. Figure monitored. study site of the solar farm as shown in Figure 2 was considered as one unit and monitored. Figure Figure10 10illustrates illustrates the mean the mean NDSI NDSI and RNDSI and RNDSI from September from September 2017 to February 2017 to 2019. February 2019. 10 illustrates the mean NDSI and RNDSI from September 2017 to February 2019. 1.8 0.15 NDSI 1.8 0.15 NDSI RNDSI RNDSI 1.5 0.12 1.5 0.12 1.2 0.09 1.2

0.09

NDSI RNDSI 0.9

NDSI 0.06 RNDSI 0.9 0.06

0.03 0.6

0.00 0.3

Jul-18 Jan-19

0.00 Jan-18 0.3

Jun-18

Oct-17 Oct-18

Feb-18

Sep-17 Sep-18

Dec-17 Dec-18

Apr-18

Mar-18

Aug-18

Nov-17 Nov-18

May-18

Jul-18

Jan-18 Jan-19

Jun-18

Oct-17 Oct-18

Feb-18

Sep-17 Sep-18

Dec-17 Dec-18

Apr-18

Mar-18

Aug-18 Nov-18 Figure 10. ComparisonNov-17 of mean NDSI values and mean RNDSI values of study site from September May-18 2017 till February 2019. FigureFigure 10. 10.Comparison Comparison ofof meanmean NDSI values values and and m meanean RNDSI RNDSI values values of of study study site site from from September September 20172017The till till February distributionFebruary 2019.2019 of. values of indices throughout the time frame in Figure 10 is uneven. Such variations arise because the cleaning process was not carried out daily. The cleaning processes scheduleTheThe distributiondistribution and performance of of values values are decidedof of indices indices by throughoutrespective throughout plant the ownerstime the timeframe [29]. frame Furthermore,in Figure in Figure10 Landsatis uneven. 10 8is data uneven.Such Suchvariationare variations nots availablearise arise because on because consecutive the thecleaning cleaning days, process hence process monthlywas wasnot not data carriedcarried were out considered. outdaily daily.. The TheAs cleaning a cleaning result, processes high processeser scheduleschedule and and performance performance are are decided decided by by respective respective plant plant owners owners [ 29[29]]. Furthermore,. Furthermore, Landsat Landsat 8 8 data data are notare available not available on consecutive on consecutive days, hencedays, hencemonthly monthly data were data considered. were considered. As a result, As a higher result, variations higher were observed in the graphical readings. However, a change in the trend of the lines is observed at the Remote Sens. 2020, 12, 1466 17 of 26 Remote Sens. 2020, 11, x FOR PEER REVIEW 17 of 26 latervariations stages. were The meanobserved value in ofthe NDSI graph isical higher readings. until AprilHowever, 2018. a Thechange decreasing in the trend pattern of the of the lines mean is valueobserved of NDSI at the is observedlater stage untils. The February mean value 2019. of NDSI is higher until April 2018. The decreasing patternUnlike of the NDSI, mean the value RNDSI of graphNDSI displaysis observed a nondrastic until February change 2019. in the values. Furthermore, the RNDSI valuesUnlike show consistencyNDSI, the RNDSI during graph the same display monthss a nondrastic of the year. change Additionally, in the values there is. Furthermore another similarity, the likeRNDSI graphs values with show NDSI. consistency A continuous during increase the same in the months value of of the RNDSI year. isAdditionally, seen from September there is another 2017 to Januarysimilarity 2018. like The graph RNDSIs with value NDSI reaches. A itscontinuous maximum increase mean at in an the index value value of of RNDSI 1.5. After is seen that period,from aSeptember continuous 2017 decrease to January in the 2018 values. The until RNDSI July value 2018 reach is observed.es its maximum This dip mean in the at graphan index significantly value of coincides1.5. After with that theperiod, drop a observedcontinuous in decrease NDSI during in the thevalues same until period. July 2018 Nonetheless, is observed. its This values dip keep in the on increasinggraph significantly thereafter. coincides with the drop observed in NDSI during the same period. Nonetheless, its valuesFigure keep 11 shows on increasing the percentage thereafter. area covered by soiling using the DBSI index from September 2017 toFigure February 11 shows 2019. Therethe percentage are few peaks area observedcovered by in soiling the first using half ofthe the DBSI year index in 2018, from this September indicates a high2017 level to February of soiling, 2019. as There shown are in few Figure peaks 11 .observed Around Julyin the 2018, first ahalf drop of the in theyear DBSI in 2018, value this near indicates to zero wasa high observed. level of Low soiling soiling, as shown was observed in Figurefrom 11. Around July 2018 July to 2018, February a drop 2019 in the when DBSI compared value near to to January zero 2018was to observed June 2018.. Low This soiling may was be due observed to the usefrom of July robotic 2018 systems to February for cleaning 2019 when the compared PV panels to to January improve effi2018ciency. to June The 2018. DBSI This graph may also be showsdue to the a similar use of trendrobotic to system NDSI ands for RNDSI.cleaning The the changePV panels in natureto improve of the graphefficiency. in all The three DBSI indices graph is observedalso shows around a similar February, trend to March, NDSI and and RNDSI. April in The 2018, change with in a decreasingnature of trendthe graph thereafter. in all three indices is observed around February, March, and April in 2018, with a decreasing trend thereafter.

28% Area covered by soil 24%

20% 15.90% 17.43% 16% 16.03% 13.75% 12%

8% 3.85% 6.53% 4% 2.77% 2.69% 2.38% 2.22% 1.53% 1.06% 0.45% 0.00%0.15% 0.01% 0.00% 0.06%

Jul-18

Jan-18 Jan-19

Jun-18

Oct-17 Oct-18

Feb-18 Feb-19

Sep-17 Sep-18

Dec-17 Dec-18

Apr-18

Mar-18

Aug-18

Nov-17 Nov-18 May-18

FigureFigure 11.11. DBSIDBSI derived percentage area area covered covered by by soil soil..

FigureFigure 12 12 describes describes the the index index values value fors for each each PV parkPV park in the in study the study area duringarea during dry and dry wet and seasons. wet Plotsseason P1,s. P2, Plots P3, P1, and P2, P4 P3 are, and described P4 are described in Figure in2 with Figure subscripts: 2 with subscripts‘d’ and: ‘‘wd’ and for dry ‘w’ for and dry wet and seasons, wet respectively.seasons, respectively. For the dry For season, the dry the season, mean the value mean of value indices of wasindices calculated was calculated for December, for December, January, February,January, March,February, and March April., Duringand April. this During period this the studyperiod area the experiencesstudy area experience a low levels ofa low precipitation. level of Onprecipitation. the other hand, On the the mean other value hand, ofthe indices mean was value calculated of indices for June, was calculated July, and September for June, July during, and the wetSeptember season, whenduring the the precipitation wet season, when is high. the Inprecipitation the case of is NDSI high. and In the RNDSI, case of the NDSI mean and index RNDSI, value the of plotsmean in index the dry value season of plots are higher in thethan dry thatseason in theare wethigher season. than However,that in the the wet mean season. values However, of DBSI the are almostmean values similar of for DBSI both are the almost dryand similar wet for seasons. both the Still, dry whenand wet compared seasons. toStill, the when wet season,compared the to mean the DBSIwet showsseason, athe lesser mean variation DBSI shows in the a drylesser season. variation in the dry season.

Remote Sens. 2020, 12, 1466 18 of 26 Remote Sens. 2020, 11, x FOR PEER REVIEW 18 of 26

(a) NDSI (b) RNDSI

Figure 12 12.. MeanMean value value of of ( (aa)) NDSI, NDSI, ( (bb)) RNDSI RNDSI and and ( (cc)) DBSI DBSI of of four four plots plots in in the the study study area area during during the the dry and wet wet season. P1d, P1d, P2d, P2d, P3d P3d and and P4d P4d are are four four PV PV p parksarks in in the the dry dry season season and and P1w, P1w, P2w, P2w, P3w P3w andand P4w in the wet season season.. All three indices show similarity in their graphical representation of soiling during the whole All three indices show similarity in their graphical representation of soiling during the whole study period, i.e., 2017–2019, as shown in Figure 12. For each index shown in Figure 12a–c, the peak study period, i.e., 2017–2019, as shown in Figure 12. For each index shown in Figure 12a–c, the peak and lowest values were observed in the dry and wet periods, respectively. From Figures 10 and 11, and lowest values were observed in the dry and wet periods, respectively. From Figure 10 and Figure it can be observed after June 2018, results show a gradual decrease in the indices values. Figure3b 11, it can be observed after June 2018, results show a gradual decrease in the indices values. Figure shows that the average rainfall in the study area is highest around July and August, which is the 3(b) shows that the average rainfall in the study area is highest around July and August, which is the same period when the lowest values in indices were recorded. The reason for this lies in one of the same period when the lowest values in indices were recorded. The reason for this lies in one of the properties of sandy soil [47,48]. The sand dunes have high moisture absorbing capacity [48]. Therefore, properties of sandy soil [47,48]. The sand dunes have high moisture absorbing capacity [48]. during the high precipitation periods, sand dunes lose their cohesion. Hence, it is less aﬀected by wind Therefore, during the high precipitation periods, sand dunes lose their cohesion. Hence, it is less erosion [47,48]. Dijk et al. [49], reported that moisture in the atmosphere dampens the ability of wind to affected by wind erosion [47,48]. Dijk et al. [49], reported that moisture in the atmosphere dampens carry ﬁne sand particles. Although more wind is expected in July and August as displayed in Figure3b, the ability of wind to carry fine sand particles. Although more wind is expected in July and August as displayed in Figure 3b, the amount of soil drifted by it is lesser. Thus, a low level of sand dune deposition is detected in July and August during the rainy season. In the case of NDSI, the soiling

Remote Sens. 2020, 12, 1466 19 of 26

Remote Sens. 2020, 11, x FOR PEER REVIEW 19 of 26 the amount of soil drifted by it is lesser. Thus, a low level of sand dune deposition is detected in July andlevel August in September during the 2017 rainy is different season. Inthan the in case the of same NDSI, month the soiling of 2018. level Therefore, in September this difference 2017 is diff erentin the thansame in season the same exists month as there of 2018.were no Therefore, self-cleaning this dirobotsfference installed in the in same September season 2017 exists [43 as]. thereThe meag wereer nochanges self-cleaning among robots the values installed exist in September because of 2017different [43]. The index meager gauge changess with amongvarying the soil values types exist and becauseproperties. of di Nonetheless,fferent index the gauges nature withs of varying the three soil indice typess and are properties.considerably Nonetheless, related to each the natures other and of theare three useful indices in detect areing considerably soiling on related PV panels. to each other and are useful in detecting soiling on PV panels.

4.4.4.4. ComparisonComparison ofof NDSI,NDSI,RNDSI RNDSIand andDBSI DBSI ExplicitExplicit and and sharp sharp images images with with a a highhigh spatialspatial resolution resolution were were acquiredacquired fromfromPlanetScope PlanetScope toto monitormonitor the the soiling soiling phenomenon phenomenon more more closely. closely. The The three-meter three-meter spatial spatial resolution resolution PlanetScope PlanetScope data with data fourwith bands four (RGB, bands NIR) (RGB, was NIR) also wasanalyzed also to analyzed see the soiling to see pattern. the soiling Figure pattern. 13 shows Figur thee comparison13 shows the of NDSI,comparison RNDSI of and NDSI DBSI, RNDSI with PlanetScope and DBSI with data PlanetScope for the validation. data for Figure the validation 13a shows. F aigure magnified 13a shows view a ofmagnified a particular view section of a particular of the PV panelssection withof the sand PV panels deposition. with Withsand keendeposition observation,. With keen a slight observation, variation isa observedslight variation on the is surface. observed This on sectionthe surface. has aThis flimsy section white has trail a flimsy and has white patches trail and on itshas rectangular patches on base,its rectangular as shown base in Figure, as shown 13a. Thein Figure same 1 section3a. The onsame the section NDSI rasteron thelayer, NDSI displayedraster layer in, displayed Figure 13 b,in highlightsFigure 13b slightly, highlight highers slightly values. higher Similarly, values. a comparison Similarly, ofa comparison RNDSI with of PlanetScope RNDSI with is alsoPlanetScope displayed is inalso Figure displayed 13a,c. in Unlike Figure NDSI, 13a,c. Unlike the values NDSI, of the pixels values along of pixels the rectangular along the rectangular base are rather base are diff erent.rather Thedifferent. dark patch The darkon the patch high-resolution on the high image-resolution has high image values has onhigh the values RNDSI on layer. the ItRNDSI indicates layer. that It theindicates region that has athe high region level has of sanda high deposition level of sand compared deposition to other compared parts. The to other result parts. of, DBSI The shown result inof, FigureDBSI shown13d however, in Figure is similar13d however, to NDSI, is showingsimilar to a NDSI, high index showing value a high on bright index patches. value on All bright three patch indiceses. practicallyAll three indices adhere practically to the data. adhere However, to the a data. finer However, observation a fine revealedr observation that NDSI revealed and DBSI that NDSI are more and sensitiveDBSI are to more visual sensitive observation. to visual The observation. dark patch onThe high-resolution dark patch on PlanetScope high-resolution data P waslanet contrastingScope data withwas RNDSI contrasting results. with The RNDSI area with results. the dark The andarea bright with patchesthe dark identified and bright with patch PlanetScopees identified data was with furtherPlanetScope investigated data was to seefurther the spectral investigated reflectance to see properties the spectral of reflectanc Landsat 8e data properties at a wavelength of Landsat ranging 8 data fromat a 0.4 wavelengthµm to 2.2 rµangingm. The from study 0.4μm of spectral to 2.2μm. responses The is study also usefulof spectral to validate responses minute is dialsofferences useful into thevalidate sand layerminute to comparedifferences accuracy in the sand between layer NDSI, to compare DBSI, andaccuracy RNDSI between qualitatively. NDSI, DBSI, and RNDSI qualitatively.

(a)Planet RGB composite (31st January 2018) (b)L8 NDSI (31st January 2018)

Figure 13. Cont.

Remote Sens. 2020, 12, 1466 20 of 26 Remote Sens. 2020, 11, x FOR PEER REVIEW 20 of 26 Remote Sens. 2020, 11, x FOR PEER REVIEW 20 of 26

(c)L8 RNDSI (31st January 2018) (d)L8 DBSI (31st January 2018) (c)L8 RNDSI (31st January 2018) (d)L8 DBSI (31st January 2018) FigureFigure 13. 13. ComparisonComparison of of (b (b) )NDSI, NDSI, (c (c) )RNDSI RNDSI and and (d (d) )DBSI DBSI with with (a (a)) high high-resolution-resolution red, red, green, green, blue blue ((RGB)FigureRGB) PlanetScope PlanetScope13. Comparison data. data. of The The (b )bright brightNDSI, patches ( patchesc) RNDSI on on panelsand panels (d )created createdDBSI with by by sand sand(a) high deposition deposition-resolution are are red,highlighted highlighted green, blue by by high(highRGB value value) PlanetScope on on NDSI, NDSI, data. RNDSI RNDSI The and andbright DBSI DBSI patches results. results. on panels created by sand deposition are highlighted by high value on NDSI, RNDSI and DBSI results. FigureFigure 1414 shows shows the the spectral spectral reflectance reflectance graph graph of of the the photovoltaic photovoltaic solar solar panels panels in in different different soiling soiling conditions.conditions.Figure As 14 As showsdisplayed displayed the spectralin in Figure Figure reflectance 14a, 14a, t thehe cleaner graph cleaner of PV PV the solar solar photovoltaic panels panels have have solar a a panelslow low reflectance reflectance in different value value soiling of of 0.17conditions.0.17 and and 0.16 0.16 As in in displayedthe the NIR NIR and and in FigureSWIR SWIR-2- 214a, bands bands, the, cleanerrespectively. respectively. PV solar On On thepanels the other other have hand, hand, a low Figure Figure reflectance 14b14b shows shows value that that of dusty0.17dusty and solar solar 0.16 panels panels in the with with NIR dark darkand patchSWIR patcheses-2 havebands have high higher, respectively.er reflectance reflectance On values valuesthe other of of 0.18hand, 0.18 and and Figure 0.23 0.23 in14b in the the shows NIR NIR andthat and SWIRdustySWIR-2- 2solar bands bands, panels, respectively. respectively. with dark Figure Figurepatches 14 14 havecc shows shows high theer the reflectancespectral spectral reflectance reflectance values of of0.18 of dusty dusty and solar0.23 solar inpanels panels the NIR with with and a a brightSWIRbright -patch.2 patch. bands It It, alsorespectively. also shows shows a a highFigure high reflectance reflectance 14c shows at at thethe the spectralbright bright patch patchreflectance with with values valuesof dusty of of 0.21 0.21solar and and panels 0.30 0.30 within in the the a NIRbrightNIR and and patch. SWIR SWIR-2 It-2 also band bands, showss, respectively. respectively. a high reflectance It It can can be be atseen seen the from frombright F Figureigure patch 1414 withbb,c,c that thatvalues the the darkof dark 0.21 patch patch and and and0.30 bright brightin the patchpatchNIR and have have SWIR a a difference di-ff2 erencebands, in inrespectively. their their reflectance reflectance It can patterns. patterns.be seen from In In the the F iguredark dark 14patch, patch,b,c that NIR NIR the and and dark SWIR SWIR-2 patch-2 reflectance reflectanceand bright valuesvaluespatch haveare are relatively relativelya difference low low incompared compared their reflectance to to the the bright brightpatterns. patch patch. In. In theIn the thedark case case patch, of ofsand, sand,NIR the and the spectral spectralSWIR reflectance-2 reflectancereflectance is highvaluesis high in are inNIR NIRrelatively and and SWIR SWIR-2low- 2compared wavelength wavelengths to thes [5 bright [050]. ]. Figure Figurepatch 14. 14Inc cindicatesthe indicates case of that sand, that the the the bright bright spectral patch patch reflectance has has high high is reflectancehighreflectance in NIR value values ands in SWIR NIR - and2 wavelength SWIR-2SWIR-2 comparedcompareds [50]. Figure toto thethe dark14darkc indicates patch.patch. This This that implies implies the bright that that the the patch bright bright has patch patch high is isreflectanceformed formed by by avalue higha highs amountin amountNIR and of sandof SWIR sand deposition-2 depositioncompared which towhich the has dark ahas high patcha high reflectance. This reflectance implies in NIR inthatand NIR the SWIR-2 andbright SWIR bands.patch-2 bandsisNDSI formed. andNDSI DBSIby and a werehigh DBSI ableamount were to detectable of sandto this detect deposition sensitive this sensitive di ffwhicherence difference has on thea high panel on reflectance the surface panel as surface in compared NIR as and compared to SWIR RNDSI-2 tovalues.bands RNDSI. NDSI Therefore, values and. DBSITherefore NDSI were and, able DBSINDSI toproved anddetect DBSI tothis be proved sensitive more accurate to difference be mo methodsre accurateon the to panel detect method surface sands to deposition asdetect compared sand on depositiontosolar RNDSI panels. onvalues solar. Therefore panels. , NDSI and DBSI proved to be more accurate methods to detect sand deposition on solar panels.

(a) Spectral reflectance of clean solar panels. (a) Spectral reflectance of clean solar panels.

Figure 14. Cont.

Remote Sens. 2020, 12, 1466 21 of 26 Remote Sens. 2020, 11, x FOR PEER REVIEW 21 of 26

(b) Spectral reflectance of dusty solar panels having a dark patch.

Figure 14. SpectralSpectral reflectancereflectance of of (a ()a clean) clean panels, panels, (b) ( dustyb) dusty panels panels with with dark dark patch patch and ( cand) dusty (c) panelsdusty panelswith bright with patch.bright patch Circle. Circle highlights highlights the di thefferences differences in reflectance in reflectance of near-infrared of near-infrared and short-waveand short- waveinfrared infrared -2 wavelength. -2 wavelength.

4.5. Accura Accuracycy A Assessmentssessment We also quantiﬁedquantified the performance of NDSI, RNDSI RNDSI,, and DBSI against the LST value derived fromfrom the Landsat thermal bands. For this, we selected 100 random samples within the the solar panel boundaries and and extracted extracted the the pixel pixel values values of of NDSI, NDSI, DBSI, DBSI, RNDSI RNDSI,, and and LST. LST. Later, Later, a threshold value of 40ᵒ 40◦ CelsiusCelsius was was applied applied (based (based on on observed observed changes) changes) to to the the LST LST data data for for representing representing the the change ( 40 C) or nonchange (<40 C) pixels in temperature. Then the performance was evaluated based on (≥40≥ ᵒ◦C) or nonchange (<40 ᵒC◦ ) pixels in temperature. Then the performance was evaluated based on various matrices in includingcluding accuracy, accuracy, kappa statistics, true positive positive (TP) (TP) and and false false positive positive (FP) (FP) rates, rates, and the area under receiver operating characteristics curve function function (AUC). (AUC). The The results results show that DBSI alone can can detect detect the the change change in in temperature temperature with with an an accuracy accuracy of of 76% 76%,, as as displayed displayed in in Table Table 33.. Table3 3 also also showsshows thatthat thethe combinationcombination ofof inputinput datadata suchsuch asas DBSIDBSI togethertogether withwith NDSINDSI increasesincreases thethe performance of detection significantly signiﬁcantly,, producingproducing anan accuracyaccuracy ofof 80%.80%.

Remote Sens. 2020, 12, 1466 22 of 26

Table 3. Performance evaluation metrics of various indices (DBSI, NDSI, and RNDSI) against the land surface temperature derived from Landsat OLI images.

Weighted Index Accuracy Kappa AUC MCC TP Rate FP Rate DBSI 76% 0.45 0.76 0.33 0.69 0.42 NDSI 67% 0.28 0.67 0.38 0.7 0.28 RNDSI 61% 0.14 0.61 0.46 0.65 0.15 DBSI+NDSI 80% 0.56 0.8 0.25 0.8 0.56 DBSI+RNDSI 74% 0.44 0.7 0.29 0.79 0.44 NDSI+RNDSI 72% 0.38 0.72 0.34 0.79 0.38 DBSI+NDSI+RNDSI 79% 0.54 0.79 0.24 0.85 0.54

Similarly, we quantiﬁed the performances of the three indices against PlanetScope data. In this case, we selected 100 points within the solar panel boundaries of four plots and compared them with the observed data points from PlanetScope images. The evaluated performance based on the kappa values and accuracy matrix is shown in Table4. The DBSI showed the highest accuracy of 89%(0.77 kappa) among the three indices shown in Table4.

Table 4. Performance evaluation metrics of various indices (DBSI, NDSI, and RNDSI) in detecting the soiled and clean pixel as veriﬁed from PlanetScope Images.

Weighted Index Accuracy Kappa AUC MCC TP Rate FP Rate DBSI 89.6% 0.77 0.89 0.12 0.86 0.77 NDSI 87.9% 0.73 0.87 0.15 0.88 0.73 RNDSI 86.2% 0.70 0.86 0.14 0.88 0.70

5. Discussion The spectral indices approach has been widely used to highlight various land cover types. Various indices have been developed for vegetation and water in past decades. However, few indices have been developed to monitor sand directly. The main reason is the complexity of sand properties and their spectral reflectance. The spectral reflectance varies with the moisture, texture, and other physical properties. From the results, it can be noticed that soiling can be detected by a customized interpretation of indices by visual cross-verifications. Al-Quraishi [27] monitored the sand dune accumulations and encroachment using NDSI with an accuracy of 90.8% in Iraq. The NDSI based method provided impressive results for mapping and monitoring sand dunes [27]. Its effective monitoring of sand dunes is useful for the selection of suitable control procedures and the prevention of further expansion. Similarly, empirically derived RNDSI highlights soil information and suppresses the noises (e.g., brightness factor caused by built-up). It is an effective technique in separating soil form the panel surfaces. RNDSI has a limitation when it comes to separating moist soil from dark impervious surfaces [26] and as RNDSI calculation is based on tasseled cap transformation of a particular image, it might cause less robust application in detecting soiling. Since the study area is covered with arid soil, it can be effectively identified on dark panel surfaces. Moreover, such arid soils are very dry, so the use of DBSI was helpful to particularly focus on bare soil. A comparative analysis was performed between NDSI, DBSI and RNDSI only since they belong to the same data lineage. Data taken at the same time is imperative for comparative purposes. Moreover, Landsat 8 images were considered for this analysis because of the availability of thermal bands which were further required for LST based validation. The DBSI index generated from a relatively high-resolution Sentinel-2 data is found to be an effective technique and a viable alternative, and it could be used to highlight the high soiling area by applying an arbitrary threshold after visual verification [25]. The images acquired by high-resolution PlanetScope provides a true color composite Remote Sens. 2020, 11, x FOR PEER REVIEW 23 of 26

Moreover, Landsat 8 images were considered for this analysis because of the availability of thermal bands which were further required for LST based validation. The DBSI index generated from a Remote Sens. 2020, 12, 1466 23 of 26 relatively high-resolution Sentinel-2 data is found to be an effective technique and a viable alternative, and it could be used to highlight the high soiling area by applying an arbitrary threshold afterview visual of the PVverification panels that [25] helps. The us images in detecting acquired the by substantial high-resolution soiling inPlanetScope the panels. pr Bothovides qualitative a true colorand quantitative composite view approaches of the PV were panels used that to comparehelps us thein detecting accuracy the of indicessubstantial to evaluate soiling theirin the ability panels to. Bothdetect qualitative minute di ffanderences quantitative in soiling approach on the PVes panels. were used to compare the accuracy of indices to evaluatThee their field ability survey to also detect revealed minute important differences information in soiling on that the supports PV panels the. findings of this study. As perThe the field standard survey operatingalso revealed manual important of the SECI,information the solar that plants support in Bhadlas the findings is being of divided this study into. Asdiff pererent the sections standard consisting operating of 400manual to 800 of panels.the SECI, Our the interviews solar plants with in the Bh projectadla is managersbeing divided and solarinto differentplant operators sections of co dinsistingfferent solarof 400 units to 800 of thepanels. four Our parks interviews in Bhadla with concluded the project that managers the solar farms and solar here plantexperience operators a high of different soiling period solar units starting of the from four January parks tillin Bhadla May. The concluded atmospheric that the temperature solar farm durings here experiencedaytime rises a high above soiling 50◦C period from March starting to from May. January Frequent till dust May. storm The eventsatmospheric are also temperature noticed during during this daytimeperiod which rises causesabove more50°C from frequent March soiling. to May Sudden. Frequent drops dust in soiling storm were events also are reported also noticed from July during until thisDecember. period which Manual causes cleaning more of frequent each section soiling of. Sudden some plants drop withs in soiling water w isere conducted also reported in 10 from to 15-day July untilcycles, December. while others Manual with cleaning Ecoppia E4of each robots section perform of some cleaning plants operations with water daily. is conducted Figure 15a,b in shows 10 to 15 the- dayphotographs cycles, w capturedhile others during with the Eco fieldppia survey E4 robots for the perform cleaning cleaning of PV panels operations using waterdaily.spray Figure systems, 15a,b showand Ecoppias the photographs E4 robot systems captured mounted during on the the field PV survey panels, for respectively. the cleaning of PV panels using water spray systems, and Ecoppia E4 robot systems mounted on the PV panels, respectively.

(a) Manual cleaning (b) Robotic cleaning

FigureFigure 15 15.. ((aa)) A A tractor tractor vehicle vehicle mounted mounted with with a a water water spray spray system system used used for for manual manual cleaning cleaning;; ( (bb)) EcoEcoppiappia E4 E4 robot robot mounted mounted on on one one of of the the panel panel rows. rows.

FromFrom a a future future perspective, perspective, the the accuracy accuracy of the of thesoil soildetection detection can be can improved be improved by using by high using- resolutionhigh-resolution satellite satellite data datamore more precisely precisely,, as noticed as noticed in the inthe case case of ofDBSI DBSI from from Sentinel Sentinel-2-2 10 10 m m images images.. However,However, high high tem temporalporal resolution resolution data data is is also also required required to to distinguish distinguish the the soiling soiling phenomenon phenomenon more more frequently.frequently. Thus Thus,, a a viable viable alternative alternative will will be be the the use use of of u unmannednmanned a aerialerial v vehiclesehicles (UAVs) (UAVs) technology, technology, thatthat provide provide high high spatial spatial and and temporal temporal resolution resolution data data.. However, However, more more monetary monetary and and technical technical investmentsinvestments are are required required to to conduct conduct such such studies. studies. Recently, Recently, with with technical technical advancement advancement in in drone drone and and imagingimaging engineering, engineering, both both thermal thermal and and optical optical cameras cameras can can be be mounted mounted on on a a UAV, UAV, thus thus generat generatinging resultsresults with with higher higher pre precision.cision. Further Further research research is is needed needed to to improve improve the the threshold threshold optimization optimization for for DBSIDBSI based based detection detection by by conducting conducting studies studies in in different diﬀerent regions regions with with varying varying meteorological meteorological and and landscapelandscape conditions. conditions.

6. Conclusions This study demonstrates the use of optical satellite data to monitor the soiling phenomenon on solar panels. The GEE tool is effectively used for rapid computation, processing, generation, and extraction of satellite data with semiautomated processing to detect soiling on PV panels cost-effectively. The frequent generation of results for a longer time frame was necessary for this study. This was achieved easily because of automated processes performed on the GEE platform. The NDSI, RNDSI and DBSI were used to detect the soiling phenomenon at the temporal scale. All of these indices could detect soiling on solar panels of PV solar farms. Moreover, a comparison of results with the LST and Remote Sens. 2020, 12, 1466 24 of 26 high-resolution images with the sand and soil indices could qualitatively verify the results. The time series analysis of various indices revealed the amount of sand depositions right from the start of the project. Based on the accuracy assessment conducted on the three indices used in this study, DBSI proved to be the most accurate in identifying the soiled areas. The soiling was more frequent from January to May and less frequent from July to August. This change is mainly because of change in the weather conditions, as precipitation in the July and August months is significantly different when compared to the other months. The monitoring of soiling on existing solar farms can discern information about soiling periods and intensity. This information can be useful to adopt a proper cleaning mechanism at required intervals. The approach used in this study is efficient and cost-effective. The Indian solar companies currently adopt a fixed schedule for cleaning which is not realistic. The methods discussed in this study could be used as an alternative source of information and provide a profound view of the field survey. However, these approaches have some limitations, and further development of the detection techniques is necessary.

Author Contributions: Conceptualization, H.S. and R.A.; methodology, H.S., R.A., A.P.Y., A.K.; software, H.S., A.G., A.A.R. and R.A.; formal analysis, H.S., V.S., J.D. and R.A.; investigation, H.S., C.K.S., O.T., A.K. and R.A.; writing—original draft preparation, H.S., A.P.Y., A.K. and R.A.; writing—review and editing, H.S., R.A., D.S., A.G., A.P.Y., J.D., A.R., G.M., S.K.C., V.S., C.K.S., O.T., and A.K. All authors have read and agreed to the published version of the manuscript. Funding: This research received no external funding. Acknowledgments: Authors would like to thank Hokkaido University and Japan Student Services Organization (JASSO) for providing fellowship. The authors also express gratitude towards Officers and Staff of the Rajasthan Renewable Energy Corporation Limited (RRECL) for granting permission to conduct the field survey in Bhadla solar farm. Furthermore, the authors are thankful to Planet lab, Copernicus Sentinel-2 data hub and the United States Geological Survey (USGS) for providing satellite data. We also acknowledge the support of Minh, Stanley, Mustafizur, Manish Soongra and Anil and appreciate the contribution made by the anonymous reviewers. Conflicts of Interest: The authors declare no conflict of interest.

References

1. Avtar, R.; Tripathi, S.; Aggarwal, A.K.; Kumar, P. Population–Urbanization–Energy Nexus: A Review. Resources 2019, 8, 136. [CrossRef] 2. Avtar, R.; Sahu, N.; Aggarwal, A.K.; Chakraborty, S.; Kharrazi, A.; Yunus, A.P.; Dou, J.; Kurniawan, T.A. Kurniawan Exploring Renewable Energy Resources Using Remote Sensing and GIS—A Review. Resources 2019, 8, 149. [CrossRef] 3. Avtar, R.; Tripathi, S.; Aggarwal, A.K. Assessment of Energy–Population–Urbanization Nexus with Changing Energy Industry Scenario in India. Land 2019, 8, 124. [CrossRef] 4. Raturi, A.K. Asia and the Paciﬁc Renewable Energy Status Report; REN21: Suva, Fiji, 2019. 5. Devabhaktuni, V.; Alam, M.; Shekara Sreenadh Reddy Depuru, S.; Green, R.C.; Nims, D.; Near, C. Solar energy: Trends and enabling technologies. Renew. Sustain. Energy Rev. 2013, 19, 555–564. [CrossRef] 6. Sahu, B.K. A study on global solar PV energy developments and policies with special focus on the top ten solar PV power producing countries. Renew. Sustain. Energy Rev. 2015, 43, 621–634. [CrossRef] 7. PV Solar Power around the World. Available online: https://en.wikipedia.org/wiki/Solar_power_by_country (accessed on 5 April 2019). 8. Yap, W.K.; Galet, R.; Yeo, K.C. Quantitative analysis of dust and soiling on solar pv panels in the tropics utilizing image-processing methods. In Proceedings of the 2015 Asia-Paciﬁc Solar Research Conference, Brisbane, Australia, 9 December 2015. 9. Wilson, N.R.; Norman, L.M.; Villarreal, M.; Gass, L.; Tiller, R.; Salywon, A. Comparison of remote sensing indices for monitoring of desert cienegas. Arid Land Res. Manag. 2016, 30, 460–478. [CrossRef] 10. Li, D. Using GIS and Remote Sensing Techniques for Solar Panel Installation Site Selection. Master’s Thesis, University of Waterloo, Waterloo, ON, Canada, 2013. 11. Bergin, M.H.; Ghoroi, C.; Dixit, D.; Schauer, J.J.; Shindell, D.T. Large reductions in solar energy production due to dust and particulate air pollution. Environ. Sci. Technol. Lett. 2017, 4, 339–344. [CrossRef] Remote Sens. 2020, 12, 1466 25 of 26

12. Avtar, R.; Aggarwal, R.; Kharrazi, A.; Kumar, P.; Kurniawan, T.A. Utilizing geospatial information to implement SDGs and monitor their Progress. Environ. Monit. Assess. 2020, 192, 35. [CrossRef] 13. Avtar, R.; Kumar, P.; Oono, A.; Saraswat, C.; Dorji, S.; Hlaing, Z. Potential application of remote sensing in monitoring ecosystem services of forests, mangroves and urban areas. Geocarto Int. 2017, 32, 874–885. [CrossRef] 14. Saidan, M.; Albaali, A.G.; Alasis, E.; Kaldellis, J.K. Experimental study on the effect of dust deposition on solar photovoltaic panels in desert environment. Renew. Energy 2016, 92, 499–505. [CrossRef] 15. Salim, A.; Huraib, F.; Eugenio, N. PV power-study of system options and optimization. In Proceedings of the EC photovoltaic Solar Conference, Florence, Italy, 8–9 May 1988; pp. 688–692. 16. Jain, A.; Kapoor, A. Exact analytical solutions of the parameters of real solar cells using Lambert W-function. Sol. Energy Mater. Sol. Cells 2004, 81, 269–277. [CrossRef] 17. Jiang, H.; Lu, L.; Sun, K. Experimental investigation of the impact of airborne dust deposition on the performance of solar photovoltaic (PV) modules. Atmos. Environ. 2011, 45, 4299–4304. [CrossRef] 18. Gupta, J. Wind erosion of soil in drought-prone areas. In Desertification and Its Control in the Thar, Sahara and Sahel Regions; Scientific Publisher: Jodhpur, India, 1993. 19. Schill, C.; Brachmann, S.; Koehl, M. Impact of soiling on IV-curves and efficiency of PV-modules. Sol. Energy 2015, 112, 259–262. [CrossRef] 20. Maghami, M.R.; Hizam, H.; Gomes, C.; Radzi, M.A.; Rezadad, M.I.; Hajighorbani, S. Power loss due to soiling on solar panel: A review. Renew. Sustain. Energy Rev. 2016, 59, 1307–1316. [CrossRef] 21. Zea-Cabrera, E.; Iwasa, Y.; Levin, S.; Rodríguez-Iturbe, I. Tragedy of the commons in plant water use. Water Resour. Res. 2006, 42. [CrossRef] 22. Karnataka: 100% Water Tariff Hike Ups Production Cost for Heavy Industries. Available online: https://economictimes.indiatimes.com/industry/indl-goods/svs/steel/karnataka-100-water-tariff-hike-ups -production-cost-for-heavy-industries/articleshow/65167007.cms?from=mdr (accessed on 15 April 2019). 23. Avtar, R.; Herath, S.; Saito, O.; Gera, W.; Singh, G.; Mishra, B.; Takeuchi, K. Application of remote sensing techniques toward the role of traditional water bodies with respect to vegetation conditions. Environ. Dev. Sustain. 2014, 16, 995–1011. [CrossRef] 24. Minh, H.V.T.; Avtar, R.; Mohan, G.; Misra, P.; Kurasaki, M. Monitoring and Mapping of Rice Cropping Pattern in Flooding Area in the Vietnamese Mekong Delta Using Sentinel-1A Data: A Case of An Giang Province. IJGI 2019, 8, 211. [CrossRef] 25. Rasul, A.; Balzter, H.; Ibrahim, G.; Hameed, H.; Wheeler, J.; Adamu, B.; Ibrahim, S.; Najmaddin, P. Applying built-up and bare-soil indices from Landsat 8 to cities in dry climates. Land 2018, 7, 81. [CrossRef] 26. Deng, Y.; Wu, C.; Li, M.; Chen, R. RNDSI: A ratio normalized difference soil index for remote sensing of urban/suburban environments. Int. J. Appl. Earth Obs. Geoinf. 2015, 39, 40–48. [CrossRef] 27. Al-Quraishi, A. Sand Dunes Monitoring Using Remote Sensing and GIS Techniques for Some Sites in Iraq; International Society for Optics and Photonics: Sanya, China, 2013. 28. Rudiyanto; Minasny; Shah; Soh; Arif; Setiawan Automated Near-Real-Time Mapping and Monitoring of Rice Extent, Cropping Patterns, and Growth Stages in Southeast Asia Using Sentinel-1 Time Series on a Google Earth Engine Platform. Remote Sens. 2019, 11, 1666. [CrossRef] 29. Bhadla Solar Park, Rajasthan. Available online: https://www.nsenergybusiness.com/projects/bhadla-solar-p ark-rajasthan/ (accessed on 24 April 2019). 30. Pandey, S. Success in Scaling-up Solar Energy in Rajasthan, India. 2013. Available online: http://re.indiaenvironmentportal.org.in/files/file/Success%20in%20Scaling-up%20Solar%20Energy %20in%20Rajasthan,%20India.pdf (accessed on 5 May 2019). 31. Publications Division. INDIA 2019: A Reference Mannual, 1st ed.; Ministry of Information & Broadcasting, Government of India: New Delhi, India, 2018; Volume 63. 32. Kar, S.K.; Sharma, A.; Roy, B. Solar energy market developments in India. Renew. Sustain. Energy Rev. 2016, 62, 121–133. [CrossRef] 33. Climate Data. Available online: https://en.climate-data.org/asia/india-129/ (accessed on 3 May 2019). 34. Adani Group. Available online: https://www.areprl.com/solar-resource/#solarbhadla (accessed on 12 June 2019). 35. Google Earth Engine. Available online: https://earthengine.google.com/ (accessed on 14 April 2019). Remote Sens. 2020, 12, 1466 26 of 26

36. GIS Course. Available online: https://www.giscourse.com/ (accessed on 21 April 2019). 37. Márquez, F.P.G.; Ramírez, I.S. Condition monitoring system for solar power plants with radiometric and thermographic sensors embedded in unmanned aerial vehicles. Measurement 2019, 139, 152–162. [CrossRef] 38. Bonafoni, S. Downscaling of Landsat and MODIS land surface temperature over the heterogeneous urban area of Milan. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2016, 9, 2019–2027. [CrossRef] 39. Saradjian, M.R.; Jouybari-Moghaddam, Y. Land Surface Emissivity and temperature retrieval from Landsat-8 satellite data using Support Vector Regression and weighted least squares approach. Remote Sens. Lett. 2019, 10, 439–448. [CrossRef] 40. Hofierka, J.; Gallay, M.; Onaˇcillová, K.; Hofierka, J., Jr. Physically-based land surface temperature modeling in urban areas using a 3-D city model and multispectral satellite data. Urban Clim. 2020, 31, 100566. [CrossRef] 41. Rahman, M.; Avtar, R.; Yunus, A.P.; Dou, J.; Misra, P.; Takeuchi, W.; Sahu, N.; Kumar, P.; Johnson, B.A.; Dasgupta, R. Monitoring Effect of Spatial Growth on Land Surface Temperature in Dhaka. Remote Sens. 2020, 12, 1191. [CrossRef] 42. Wang, F.; Qin, Z.; Song, C.; Tu, L.; Karnieli, A.; Zhao, S. An Improved Mono-Window Algorithm for Land Surface Temperature Retrieval from Landsat 8 Thermal Infrared Sensor Data. Remote Sens. 2015, 7, 4268–4289. [CrossRef] 43. Israel’s Ecoppia To Deploy 2000 E4 Solar Panel Cleaning Robots for SB Energy’s Installations in Bhadla Solar Park. Available online: http://taiyangnews.info/business/ecoppia-secures-580-mw-order-for-bhadla-solar- park/ (accessed on 24 April 2019). 44. Patil, P.; Bagi, J.; Wagh, M. A review on cleaning mechanism of solar photovoltaic panel. In Proceedings of the 2017 International Conference on Energy, Communication, Data Analytics and Soft Computing (ICECDS), Chennai, India, 1–2 August 2017; pp. 250–256. 45. Ecoppia Expands Bhadla Park Cloud-Based Robotic Cleaning Footprint with Additional 580MWp. Available online: https://www.ecoppia.com/press-releases/ecoppia-expands-bhadla-park-cloud-based-robot ic-cleaning-footprint-with-additional-580mwp/ (accessed on 15 May 2019). 46. PV Solar Panel Field Inspection with UgCS. 2013. Available online: https://www.ugcs.com/solar-panel-inspe ction-with-ugcs (accessed on 25 June 2019). 47. Kidron, G.; Yair, A. Rainfall–runoff relationship over encrusted dune surfaces, Nizzana, Western Negev, Israel. Earth Surf. Process. Landf. J. Br. Geomorphol. Group 1997, 22, 1169–1184. [CrossRef] 48. Tsoar, H. Sand dunes mobility and stability in relation to climate. Phys. A Stat. Mech. Appl. 2005, 357, 50–56. [CrossRef] 49. Van Dijk, P.; Stroosnijder, L.; De Lima, J. The influence of rainfall on transport of beach sand by wind. Earth Surf. Process. Landf. 1996, 21, 341–352. [CrossRef] 50. O’Neill, A.L. Reflectance spectra of microphytic soil crusts in semi-arid Australia. Int. J. Remote Sens. 1994, 15, 675–681. [CrossRef]

© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).