Unravelling spatiotemporal patterns of solar photovoltaic plants development in China in the 21st century

Solar energy, as an environmentally sustainable power source, is gaining increasing popularity worldwide, driving a surge in the number of solar photovoltaic (PV) plants. China, which has a prominent role in this domain, requires continuous updates to its PV plant data for spatiotemporal analyses. However, there remains an absence of a comprehensive and timely dataset of PV plants across China, leaving PV installation dates and other crucial attributes for comprehensive analyses underexplored. This study leverages Sentinel-2 data as a primary source to propose an optimized deep learning approach for PV plant extraction in China. Statistical analyses of PV plant attributes, including its installation date, size, site slope, and site land cover, were implemented from multiple data sources. Comprehensive analyses were conducted to unravel their spatiotemporal development patterns in the 21st century. The results indicate that as of 2023, China boasts 4347 PV plants, collectively spanning 4146 km2, which are predominantly concentrated in Northwest and North China. 2016 and 2017 marked substantial growth in China’s PV plants, while other years exhibited stability. These plants exhibit the distinct spatial characteristics of installing smaller PV plants on flat terrain covered by vegetation or barren land. Over time, a notable trend in the installation of China’s PV plants has been the increasing preference for establishing larger ones in smooth terrain, with a focus on preventing damage to natural resources. The results reveal China’s optimization of PV plant site selection and construction strategies, aligning with global environmental goals and sustainable energy practices.


Introduction
Solar photovoltaic (PV) plants are widely recognized as renewable energy facilities for reducing operational expenses and lowering carbon dioxide emissions, which aligns with the Sustainable Development Goals proposed by the United Nations (Nemet 2009, Creutzig et al 2017).Owing to their environmental and economic advantages, PV plants have seen a substantial increase worldwide.In the International Energy Agency's (IEA) sustainable development scenario, it is anticipated that 4240 GW of PV plant capacity will be deployed by 2040, representing about 11 000-fold increase compared to the capacity of 385 MV in 2000.
China is acknowledged as the country that holds the most substantial share of PV plants globally (Lam et al 2018, Kruitwagen et al 2021).PV plants in China have witnessed an 87.4 GW increase in installed capacity in 2022, resulting in a cumulative capacity of 392.04 GW in total.From geographical and environmental perspectives, the rapid expansion of the PV plant industry exhibits distinct patterns, traits, and trends (Wang et al 2023b).The progress of PV plants is frequently shaped by multiple factors, including policies and environmental conditions.Therefore, gaining an understanding of its spatiotemporal dynamics holds significant importance for the management of the PV industry and subsequent evaluations of its ecological and environmental impacts (Taha 2013, Hernandez et al 2015).
Satellite data is a crucial resource for effective PV plant monitoring.To date, a variety of remote sensing data sources have been employed for the extraction of PV plants.High-resolution remote sensing imagery achieves satisfactory accuracy in segmentation (Yu et al 2018, Kruitwagen et al 2021), yet their application on a large scale, such as at the national or global level, poses challenges for timely monitoring.Medium-resolution remote sensing imagery such as Landsat ( Other attributes, such as regional land cover (Kim et al 2021, Kruitwagen et al 2021, Xia et al 2022a, Tao et al 2023, Wang et al 2023a, 2023b) and slope (Charabi and Gastli 2011) of the construction site, were investigated as well.
However, we still face a deficiency in possessing the most recent dataset that provides detailed attributes of PV plants.Consequently, a comprehensive analysis of the spatiotemporal distribution of PV plant combined with multiple attributes was not specified.This significantly limits us from conducting indepth investigations into the spatial distribution patterns and development features of PV plants, as well as comprehensive analyses of construction strategies and their impacts on the ecological environment.Several characteristics of PV plants, such as their installation date, size, and site characteristics, serve as indicators for planning strategies.Investigating these attributes and conducting statistical and categorical analyses from a spatiotemporal perspective, can unveil China's planning strategies and site optimization for PV plants.
This study aims to establish a dataset encompassing China's PV plants to conclude the spatiotemporal patterns.Specifically, a two-step deep learning model was trained to collect the PV plant dataset in China from Sentinel-2 data acquired in 2023.The attributes of the extracted PV plants, including their installation date, size, and site characteristics (regional slope and land cover), were then investigated using multiple data sources.Finally, the spatiotemporal distribution and characteristics of PV plants in China in the 21st century are analysed, and the influencing factors contributing to these spatiotemporal features are discussed.

Mapping PV plants
Obtaining the precise locations and boundaries of all PV plants is fundamental for exploring their development patterns in China.Despite the existence of datasets covering the distribution of PV plants in China, their accuracy is compromised, primarily due to inaccuracies in extracting boundaries (Kruitwagen et al 2021).Additionally, these datasets exhibit limited temporal coverage and low timeliness (Zhang et al 2022).Consequently, we propose employing the year 2023 as the reference year, utilizing Sentinel-2 Level-1C imagery with 10 m resolution, which underwent rigorous radiometric and geometric correction, covering China, as the primary data source for mapping PV plants (Butler et al 2015).All Sentinel-2 images within the studied period underwent cloud removal preprocessing on the Google Earth Engine (GEE) platform, followed by median compositing.The composited median image with three visible bands (RGB), one near-infrared band (NIR), and two short-wave infrared bands (SWIR1 and SWIR2) was then downloaded and used as the data source for the PV plant extraction task.
For the extraction of small-scale targets, such as small PV plants in the 10 m resolution Sentinel-2 imagery, an image classifier proved effective in precise localization and reducing misclassification for subsequent segmentation (Yu et al 2018).Therefore, in this study, a two-step deep learning model, including an image classifier model and a segmentation model, was developed for PV plant mapping (figure 1).A meticulously selected ResNet-50 classifier was used to discern the presence of PV plants within a given patch.The semantic segmentation model UNet++, utilizing MobileNet V3 as its backbone, was chosen to accurately identify the boundaries of PV plants.The input data of model encompassed six variables corresponding to the data source bands.To train the deep learning models, training and test sets were constructed using a series of processes.Specifically, Sentinel-2 imagery samples were initially located using the Global PV inventory proposed by Kruitwagen et al (2021), followed by careful sample selection and manual labelling.The dataset was divided into training set and test set of 80% and 20%, respectively.To avoid overlooking the samples that contained PV plants, the ResNet-50 classifier was trained with 2804 samples in total, including 1804 adequate positive samples with distinctive features and 1000 negative samples, with a size of 128 × 128 × 6.The UNet++ segmentation model was trained using more samples (5400 image samples measuring 128 × 128 × 6, including a small proportion of negative samples) to avoid under-fitting and ensure a balanced representation of both positive and negative pixels in the dataset.All the training sets were subjected to data augmentation before training.
With the data source for prediction, the PV plants were predicted using adeptly trained models.First, the median composited Sentinel-2 data for prediction were processed into clips.Subsequently, the classifier was applied to each clip to determine whether it contained the PV plants.The classifier was designed to minimize the omission of clips containing PV plants, which might result in higher commission errors.Thus, a visual inspection was performed to ensure the accuracy of the results.Finally, the clips containing the PV plants were processed using the segmentation model to obtain the accurate PV plants boundaries.The PV plant vectors were systematically combined to ascertain the total number of PV plants.Specifically, any two PV vectors located within a distance of 1 km or less are merged into a single PV plant.

Dating PV plants
The installation dates of PV plants constitute a highly significant attribute that is intricately linked with subsequent temporal analyses.To precisely ascertain the installation dates spanning two decades, a spectrum reflectance-based algorithm LandTrendr employed on GEE was utilized (figure 2).The LandTrendr algorithm is a spectral time-segmentation approach used for break-point detection in a long time-series of Landsat series imagery, which has been proven to be effective in various break-point detection tasks (Kennedy et al 2010).The used Landsat images underwent cloud, shadow and snow removal by the LandTrendr API to reduce noise.
Subsequently, the near-infrared (NIR) band of the Landsat imagery was used to generate a time-series as a pronounced reduction in NIR can reflect PV plant construction (Fan and Huang 2020).Finally, the break-point was detected using LandTrendr as an indicator of the installation date.
To ensure the accuracy of subsequent analyses, the installation date of the PV plant was defined as the latest year when a significant drop followed by a stable series of NIR can be seen in the timeseries.During the experiments, two distinct scenarios were found to frequently appear (figure 2).In the first scenario, characterized by the absence of a break-point, LandTrendr identifies points with apparent changes that subsequently stabilize, and this method was observed to be accurate (figure 2(a)).However, in the second scenario, where a break-point exists, LandTrendr identifying them as installation date points cannot be convincing, as they lack reliability in observation (figure 2(b)).Consequently, in the second scenario, the installation date is determined as one year after the break-point to ensure the overall accuracy of the results.
The pixels within the PV plant area were scrutinized to generate the NIR temporal mean values.The LandTrendr algorithm with the parameters for detecting the most recent break-points was then applied to the NIR time-series.However, in many cases, the beak-point detector incorrectly identified the installation dates, such as the years 2000 and 2023, owing to the lack of Landsat imagery in these sites.Therefore, a visual inspection of the NIR time-series was conducted on the installation dates for the cases to further ensure accuracy.

Investigating attributes of PV plants
Following the acquisition of the locations and installation dates of PV plants, we employed additional data sources to extract more attributes, including the area of PV plants, and the slope and land cover of their construction sites.
The PV plant area was directly calculated from the segmentation results.The PV plant slope was calculated using the Shuttle Radar Topography Mission (SRTM) DEM data on GEE.Specifically, we computed and averaged the slopes of all covered pixels to derive the slope value for each PV site.The land cover of the PV plant site was determined after the dating process, which was defined as the land cover before the PV plant was installed.The land cover map was quarterly generated on Landsat imagery within the extent of PV plants using the Random Forest algorithm, dividing pixels into five classes: cropland, water, barren land, vegetation and built-up area.The sample points used to train the classifier were generated aided by the dataset of Yang and Huang (2021).The land cover of each PV plant was then assigned as a class occupying the majority of its site.

Validation
The confusion matrix was constructed using the prediction and ground-truth of the test set to evaluate the performance of the segmentation model, which shows a mean intersection-over-union (IoU) of 90.23% (figure 3(a)).Noted that the capacity of PV plants can be derived from its area (Kruitwagen et al 2021), the correlation between the predicted (provincial area of PV plants by 2022) and measured data (provincial capacity of PV plants by 2022 measured by the National Energy Administration) was generated to further evaluate the accuracy of the extracted PV plant data.As a result, the R 2 for their relationship is 0.902 (figure 3(b)), showing a strong correlation between the mapping results and the government statistical data, indicating validate data of PV plants on provincial scale.These results signify the satisfactory performance of the algorithms in PV segmentation, affirming the validity of further analysis.
To assess the performance of the installation date detection algorithm, we randomly chose 100 PV vector samples using Google Earth to ascertain the actual installation dates of these vector samples.A confusion matrix was constructed by combining predicted and actual results to evaluate the algorithm (figure 3(c)).The results show an acceptable performance with an average accuracy of 91%.The time gaps between the predicted and observed dates range from 0 to 2 years, with a mean RMSE of 0.41.
The assessment of the site land cover was conducted using test sample points to directly generate the confusion matrix on the GEE platform.The results show an overall accuracy of 90.04%, signifying an acceptable performance.

Spatiotemporal distribution of PV plants in China
PV plants exhibit a widespread distribution in the Chinese region: a total of 22 425 PV arrays merged into 4347 PV plants, spanning 4146 km 2 .The density map of PV plants in China illustrates distinct spatial patterns (figure 4(a)).Notably, regions such as Xinjiang, Inner Mongolia, and Gansu exhibit a high density, yet extensive areas remain entirely devoid of PV plants, indicating a relatively concentrated distribution.In contrast, regions including Hubei, Anhui, and Zhejiang display lower density but a more uniform dispersion of PV plants.These conspicuous spatial features are also reflected in the various geographical divisions within China.For instance, Northwest China boasts the highest proportion of PV plant area (33%), characterized by a dense concentration, while North China follows closely behind the secondhighest proportion (22.9%), showing a more uniform and dispersed distribution (figures 4(a) and (b)).Furthermore, significant differences in PV plant area between provinces are discernible.In vast provinces such as Xinjiang and Inner Mongolia, areas can span up to 300 km 2 , whereas smaller provinces, such as Hong Kong and Macau, show no detectable presence of PV plants (figure 4(c)).These marked spatial disparities in PV plant distribution underscore the existence of distinct construction strategies across different regions, thus highlighting the diversity in solar energy utilization among various areas.
In China, the construction of PV plants was relatively limited before 2012.However, since 2012, there has been a continuous increase in PV plant construction, with the annual construction area displaying a pattern of initial growth followed by a subsequent decrease (figure 5

Spatiotemporal characteristics of PV plants in China
When considering the PV plant distribution across all geographic divisions of China with different attributes in terms of their size, slope, and land cover, the results show various characteristics (figure 6).
PV plant size exhibits noteworthy consistency as 0-1 km 2 PV plants are uniformly distributed across various regions (figure 6(a)), which also accounts for the majority of PV plants in all regions (figure 6(d)).Conversely, PV plants with an area exceeding 5 km 2 are primarily concentrated in Northwest China, accounting for 68% of such installations, presenting notable differences (figure 6(a)).
Slope exhibits both disparity and consistency as well.For example, Northwest China features PV plants constructed in areas with 0 • -3 • slopes, encompassing nearly half of this slope category's total area (46%), whereas in Southwest China PV plants are predominantly situated in areas with slopes greater than 12 • , covering the majority of this category's total area (56%), presenting significant differences (figure 6(b)).However, the fact that all regions exhibit a preference for constructing PV plants in areas of Land cover, however, witnesses substantial disparities across different geographical divisions (figures 6(c) and (f)).For instance, there is no uniform dominant land cover type across all geographic divisions (figure 6(c)).In Northwest China, PV plants are predominantly situated on land covered by vegetation or barren land, accounting for nearly the entire PV plant area (96%).In Northern China, vegetation PV plants represent the majority of PV plant areas (62%).In other regions, cropland is the dominant land cover type.These distinctions in the distribution of PV plants across China's geographical divisions reflect the varying considerations in their construction strategies within each region.
Between 2012 and 2022, temporal variation was observed in terms of the size and site characteristics of PV plants, which could be further concluded into multiple temporal patterns (figure 7).
Over time, China's PV plants exhibited distinct distribution trends in terms of size (figure 7(a)).Notably, there was a 3% reduction in the proportion of 0-1 km 2 PV plants, while 1-2 km 2 PV plants saw a 3% increase.These trends can be extended to the slope (figure 7(b)).For instance, there was a 6% expansion in PV plants' slopes of 0 • -3 • , whereas PV plants built on steeper terrains witnessed a reduction in their proportional representation.However, additional alterations were observed for the land cover of PV plants (figure 7(c)).The proportions of areas covered by water, vegetation, and

Methods
Establishing a precise and timely geospatial dataset is pivotal for scrutinizing the spatiotemporal dynamics of PV plants.Remote sensing data has emerged as an effective tool for extracting specific targets across expansive areas.Sentinel-2 imagery stands out as particularly advantageous for realtime updates over extensive regions.Despite its rich spectral information and acceptable data volume, it is imperative to acknowledge that the accuracy of the extraction results may face limitations due to its resolution constraints.Leveraging spectral reflectance-based break-point detection methods, such as LandTrendr employed in this study (figure 2), offers a rapid and efficient approach for extracting the installation date of PV plants.However, the assurance of accuracy often necessitates the application of well-crafted postprocessing techniques.While this approach typically yields a generally satisfactory result (figure 3(c)), optimized methods are advised for temporal analysis, such as applying shorter time intervals instead of years.

Spatiotemporal analyses
The spatial distribution of PV plants across various regions of China exhibits a significant disparity (figure 4), underscoring the multifaceted factors influencing their siting.First, natural factors, such as solar radiation ( Šúri et al 2005, Markvart et al 2006, Huld et al 2012, Yu et al 2018), stand out as pivotal determinants in the selection of PV plant locations.For instance, the abundant solar resources found in Northwest China substantiate the clustering of a considerable number of PV plants in these areas.Secondly, socio-economic considerations, particularly land use conflicts, hold substantial sway over PV plant siting.The expansive deserts and barren lands of Northwest China lend themselves to large-scale PV plants, whereas the densely populated zones in East and Central China are less amenable to widespread PV plants.Finally, it was discovered that regional policies within China exert a notable impact on PV plant location decisions (Liao et al 2022).
The temporal analysis conducted in this study revealed significant trends in the development of PV plants in China (figure 5).These trends in the PV plant area in China exhibit clear patterns, because regional PV construction trends are closely related to the external economic climate and internal policies (Luan and Lin 2022).Notably, the sluggish commercial performance of PV industry in 2012-2013 can be predominantly attributed to anti-dumping and anti-bribery (AD & AB) investigations.It was not until early 2014 that the industry experienced a widespread resurgence, only to face restrictions imposed by the '531' policy in 2018 (Luan and Lin 2022).The impact of the policy exhibits a certain degree of latency, resulting in the continued stagnation of PV industry in China in 2019.Wang et al 2023b).The slope is another criterion for assessing PV planning strategies.For instance, when a region leans towards constructing PV plants in steep terrains, it reflects a distinct planning characteristic, since building PV plants in such areas increases maintenance challenges (Charabi and Gastli 2011, Uyan 2013).In addition, the land cover of PV plant sites reveals the region's construction strategy and attitude towards the environmental impact of PV plants (Dhar et al 2020, Kruitwagen et al 2021).Constructing PV plants involves altering the previous land cover of the site, resulting in a massive land cover disruption.This study discovered that when considering factors such as the PV plant size, slope, and land cover, it becomes evident that China has optimized its construction strategies.Site selection increasingly prioritizes flatter regions to minimize costs, and prioritizes non-natural land types to minimize environmental impacts.This transition in China's PV plant construction trend mirrors the country's evolving policies and technological advancements in PV plant.

Limitations and future research
The method used in this study and the analysis of results also have some limitations.Firstly, the 10 m spatial resolution of the used Sentinel-2 imagery resulted in the inability to detect small objects, thus excluding the extraction and analysis of small-scale PV plants and rooftop PV in this study.Secondly, some important information about the PV plants cannot be directly extracted from remote sensing imagery, and further integration with other data is necessary.
From a future perspective, the extraction process developed in this study can be applied to global monitoring, enabling a comparative analysis of PV development trends and corresponding policies across different countries.Additionally, the PV monitoring conducted in this study can be effectively integrated with PV potential assessment to evaluate the suitability of PV construction.Finally, further attention should be given to the environmental impact of PV plant, such as the effects of PV construction on the surrounding environment and issues related to the waste management of PV materials (Aman et al 2015).

Conclusion
This study presents a deep learning algorithm for the extraction of PV plants in China.The essential attributes of the extracted PV plants, such as the installation date, area, and site characteristics (slope and land cover) were also investigated, in order to perform an extensive spatiotemporal analysis for the development of Chinese PV plants in the 21st century.The evaluation results demonstrate the capability to extract various types information about PV plants, and the analysis reveals a distinct spatiotemporal distribution of China's PV plants.In Zhang et al 2022, Xia et al 2022a, Wang et al 2023a) and Sentinel-1/2 (Xia et al 2022b, Jiang et al 2023, Tao et al 2023) is more commonly utilized because of its rapid data updates and cost-free accessibility.In addition, PV plant extraction methods have been widely discussed in previous studies.For example, manual annotation (Bradbury et al 2016, Dunnett et al 2020), machine learning techniques (Zhang et al 2022, Xia et al 2022b) and deep learning methods (Yu et al 2018, Kruitwagen et al 2021) have been commonly used when tackling PV plant extraction tasks.In comparison to research focused on the extraction of PV plant locations with satellite data, few studies have examined the attribute characteristics of PV plants.Most of the works utilized recorded data for the installation date of PV plants (Jordan and Kurtz 2011, Trapani and Redón Santafé 2014, Stainsby et al 2020), but they were greatly limited by data accessibility.Methods for investigating the installation date were also explored, such as date detection based on spatial classification results (Xia et al 2022a, 2022b) and spectrum-reflectance-based methods, such as Continuous monitoring of Land Disturbance (COLD) algorithm (Tao et al 2023).

Figure 1 .
Figure 1.The methodology of the deep learning methods for extracting PV plants.

Figure 2 .
Figure 2. Illustrations of the installation date search method based on LandTrendr, and corresponding Google Earth images.(a) A case does not contain a break-point.(b) A case contains a break-point.

Figure 3 .
Figure 3. Evaluation results presenting (a) the performance of the deep learning methods by the confusion matrix, (b) the correlation of predicted area of PV plants by 2022 and measured capacity from National Energy Administration by 2022, and (c) normalized confusion matrix constructed by observed and predicted dates of test set of PV plants.
(a)).Most Chinese provinces have experienced steady growth in PV plant construction (figure 5(b)).Notably, provinces with significant fluctuations, such as Xinjiang and Gansu, underwent a surge in PV plant construction in 2016, leading to a sudden boost in overall growth (figure 5(a)).By 2018, the growth had gradually stabilized (figures 5(b) and (c)).Additionally, only Shandong and Qinghai provinces witnessed substantial PV plant construction in 2017 and 2021, respectively (figure 5(b)).In summary, the temporal pattern of PV plant development in China indicates an overall state of stable growth, which is likely influenced by China's PV plant policies.

Figure 4 .
Figure 4.The spatial distribution of PV plants in China.(a) PV plant area density map (PV plant area per unit area).(b) PV plant area proportion in geographic divisions, in which NW represents Northwest China, N represents North China, C represents Central China, E represents East China, SW represents Southwest China, S represents South China and NE represents Northeast China.(c) Provincial PV plant area.

Figure 5 .
Figure 5. distribution of PV plants in China spanning years 2000-2022 with results from 2000 to 2011 merged into one unit.(a) Total number and area of PV plant variation.(b) Maps of the temporal distribution of PV plants.(c) Heatmap of provincial PV plant area variation.

Figure 6 .
Figure 6.The spatial distribution of the PV plants in China.(a) Individual PV plant area distribution, (b) PV plant site slope distribution and (c) PV plant site land cover is subdivided by geographic divisions.PV plant distribution in geographic divisions is subdivided by (d) six categories of individual PV plant area, (e) five categories of PV plant site slope and (f) five categories of PV plant site land cover.

Figure 7 .
Figure 7.The temporal distribution of PV plant attributes by 2012, 2017 and 2022, in which (a) shows the accumulated number of PV plants in six classes of individual PV plant size, (b) shows the accumulated area of PV plants in five classes of PV plant site slope and (c) shows the accumulated area of PV plants in five PV plant site land cover types.
The COVID-19 pandemic has emerged as another factor significantly impeding the development of the Chinese PV industry by impacting the industry's overall progress, particularly in 2020 (Song et al 2022).The future trajectory of the PV industry likely continues to be influenced by multiple factors, necessitating diverse means of achieving timely updates and interpretations of PV plants.The attributes of PV plants, such as their size and site characteristics (slope and land cover), serve as indicators of industry development.The size of PV plants correlates closely with various factors, including technology, labour, and demand for PV-generated electricity (Khatib et al 2013, summary, PV plants are predominantly clustered in Northwest and North China, with provinces such as Xinjiang and Inner Mongolia boosting most PV plants.Notable surges are found in PV plant areas occurring in 2016 and 2017, whereas other years present stability.These installations exhibit a preference for flat terrain featuring vegetation or barren land.Over time, China has fine-tuned its selection of installation sites to align with the Sustainable Development Goals (SDG) outlined by the UN.Luan R and Lin B 2022 Positive or negative?Study on the impact of government subsidy on the business performance of China's solar photovoltaic industry Renew.Energy 189 1145-53 Markvart T, Fragaki A and Ross J N 2006 PV system sizing using observed time series of solar radiation Sol.Energy 80 46-50 Nemet G F 2009 Net radiative forcing from widespread deployment of photovoltaics Environ.Sci.Technol.43 2173-8 Song Y, Liu T, Li Y and Ye B 2022 The influence of COVID-19 on grid parity of China's photovoltaic industry Environ.Geochem.Health 44 2847-62 Stainsby W, Zimmerle D and Duggan G P 2020 A method to estimate residential PV generation from net-metered load data and system install date Appl.Energy 267 114895 Šúri M, Huld T A and Dunlop E D 2005 PV-GIS: a web-based solar radiation database for the calculation of PV potential in Europe Int.J. Sustain.Energy 24 55-67 Taha H 2013 The potential for air-temperature impact from large-scale deployment of solar photovoltaic arrays in urban areas Sol.Energy 91 358-67 Tao S, Rogan J, Ye S and Geron N 2023 Mapping photovoltaic power stations and assessing their environmental impacts from multi-sensor datasets in Massachusetts, United States Remote Sens. Appl.: Soc.Environ.30 100937 Trapani K and Redón Santafé M 2014 A review of floating photovoltaic installations: 2007-2013 Prog.Photovolt.23 524-32 Uyan M 2013 GIS-based solar farms site selection using analytic hierarchy process (AHP) in Karapinar region, Konya/Turkey Renew.Sustain.Energy Rev. 28 11-17 Wang X, Xiao X, Zhang X, Ye H, Dong J, He Q, Wang X, Liu J, Li B and Wu J 2023a Characterization and mapping of photovoltaic solar power plants by Landsat imagery and random forest: a case study in Gansu Province, China J. Clean.Prod.417 138015 Wang Y et al 2023b Accelerating the energy transition towards photovoltaic and wind in China Nature 619 761-7 Xia Z, Li Y, Chen R, Sengupta D, Guo X, Xiong B and Niu Y 2022a Mapping the rapid development of photovoltaic power stations in northwestern China using remote sensing Energy Rep. 8 4117-27 Xia Z, Li Y, Guo X and Chen R 2022b High-resolution mapping of water photovoltaic development in China through satellite imagery Int.J. Appl.Earth Obs.Geoinf.107 102707 Yang J and Huang X 2021 The 30 m annual land cover dataset and its dynamics in China from 1990 to 2019 Earth Syst.Sci.Data 13 3907-25 Yu J, Wang Z, Majumdar A and Rajagopal R 2018 DeepSolar: a machine learning framework to efficiently construct a solar deployment database in the United States Joule 2 2605-17 Zhang X, Xu M, Wang S, Huang Y and Xie Z 2022 Mapping photovoltaic power plants in China using Landsat, random forest, and Google Earth Engine Earth Syst.Sci.Data 14 3743-55