Photovoltaic Power Station Identification Based on High-Resolution Network and Google Earth Engine: A Case Study of Qinghai Province, Northwest China

Highlights

What are the main findings?

• HRNetv2 (validation Dice = 0.9463) outperformed other models in small-sample PV segmentation via the optimized remote sensing method (integrating PCA, GEE, deep learning).
• The 2020–2024 Qinghai PV area exceeded the pre-2019 area; it mainly occupied bare land (88.7%) and promoted desert greening.

What are the implication of the main findings?

• The optimized method effectively supports PV spatial identification in arid/semi-arid regions with bare ground interference.
• PV distribution and land use guide arid-region PV-ecology synergy, and indicates future algorithm refinement for better generalizability.

Abstract

The precise identification of photovoltaic power stations is essential for advancing the assessment of energy infrastructure and for the efficient management of land resources. To address the need for spatially explicit data on photovoltaic (PV) development in arid and semi-arid regions amid green energy transitions, particularly in the context of identification challenges induced by the widespread distribution of bare ground, this study optimized a remote sensing-based identification method integrating Principal Component Analysis (PCA), automated sampling via Google Earth Engine (GEE), and deep learning models, and applied it to Qinghai Province, one of China’s largest PV regions. The results showed that HRNetv2 (validation Dice = 0.9463) outperformed UNet (0.9328), Attention UNet (0.9399), and HRNet + OCR (0.9184) in small-sample (1871 training samples) PV segmentation; the PV installed area during 2020–2024 accounted for 63.5% of the total pre-2024 area (~607 km²), exceeding the cumulative area before 2019, with projects predominantly distributed in areas with elevation less than 2500 m and slope less than 2°; bare land dominated PV land use (88.7%), followed by grassland (6.9%) and shrubland (3.9%), and PV construction contributed to desert greening by modifying microclimates. The study concludes that its optimized method effectively supports PV spatial identification, and the revealed PV distribution and land use patterns provide scientific guidance for synergistic PV development and ecological conservation in arid regions, while acknowledging limitations in generalizability to other regions due to Qinghai-specific data, suggesting future algorithm refinement and expanded research scales.

1. Introduction

With the aim of mitigating climate change, preserving the ecological environment and realizing sustainable human development, numerous countries worldwide are pursuing the United Nations’ Sustainable Development Goals [1]. Renewable energy is widely regarded as a core solution due to its multifaceted socioeconomic and ecological value [2,3]. Among these, solar energy stands out as one of the most promising clean energy sources [4], characterized by abundant reserves, widespread availability [5,6,7], zero carbon emissions, and the absence of solid or liquid waste generation [8,9,10]. Currently, photovoltaic (PV) power generation represents the main method for harnessing solar energy [11]. China, as the world’s largest carbon-emitting economy, plays a pivotal role in the global low-carbon energy transition through its carbon reduction measures [12]. In September 2020, China pledged to peak carbon emissions by 2030 and reach carbon neutrality by 2060, establishing its dual carbon strategy. Driven by this, the nation’s installed PV power generation capacity reached 886 GW by the end of 2024 [13]—more than quadrupling from 204.3 GW at the end of 2019 [14].

Since 2001, the methodology for PV power plant site selection has undergone a fundamental shift, evolving from a single economic objective optimization towards a comprehensive multi-criteria decision-making (MCDM) framework that integrates environmental, technical, and social dimensions [15]. Initially dominated by MCDM methods, which provide a structured decision-making framework, these approaches often rely on expert subjective judgment for criteria weighting [16,17]. In recent years, the integration of artificial intelligence, represented by machine learning (ML) and optimization algorithms (OA), has significantly advanced the field. ML enhances predictive accuracy by modeling complex, non-linear data relationships in challenging terrains, while OA effectively solve multi-objective coordination problems [18,19,20]. The current research frontier is moving towards a multi-model integrated site selection framework, where the synergistic combination of these methods enhances the robustness and scientific rigor of decision-making [21,22]. The construction of the site selection evaluation system commonly adopts a “restrictive-suitability” dual-layer logic [23]. Restrictive criteria function as hard constraints to eliminate unsuitable areas, whereas suitability criteria facilitate the graded evaluation and comparative optimization of potential sites using a multi-dimensional indicator system spanning technical, economic, environmental, and social factors. For dominant centralized land-based PV plants, ideal locations must possess core natural conditions such as high solar radiation, long sunshine duration, and stable, flat terrain [24,25]. Simultaneously, they must strictly avoid ecologically and culturally sensitive areas [26,27] while ensuring proximity to the power grid, accessible transportation networks, and reasonable land costs as key socio-economic factors [16,28].

Among the factors mentioned above, the most critical data for determining the site selection of a PV power plant is the regional solar radiation potential value [29,30,31]. Solar resources in China are unevenly distributed, with the arid and semi-arid regions in the northwest concentrating the majority of solar resources. These areas are regarded as China’s future energy bases [32]. The national plan for 2022 calls for constructing large-scale wind and PV bases primarily in desert areas of the northwest, aiming to achieve a total installed capacity of 455 GW by 2030 [33]. Deploying PV facilities in arid regions holds three additional significant implications beyond advancing energy transition. First, the harsh environmental conditions in these areas severely impede socioeconomic development, making PV projects a vital poverty alleviation measure. In 2021, the PV poverty alleviation initiative covered all 39 impoverished counties in Qinghai Province, generating revenues of 570 million yuan and lifting 77,000 households and 283,000 people out of poverty [34]. Second, PV installations in arid regions can also yield numerous positive ecological benefits. Field studies indicate that solar panels can help regulate air and ground humidity, thereby maintaining high soil moisture levels and alleviating heat stress, which accelerates vegetation recovery in arid areas [35,36]. Furthermore, they can reduce wind-blown dust and sand [37], and mitigate soil erosion [32]. Third, they enhance the utilization of degraded land, enabling spatial symbiosis between the PV industry and other sectors such as agriculture and animal husbandry [11,38], thereby achieving synergistic benefits for multi-use solar sites [39]. For instance, agrivoltaics systems provide shade to lower intense solar radiation, boosting crop yields. Simultaneously, the evapotranspiration from underlying crops lowers panel temperatures, thereby improving energy production efficiency—creating a win-win outcome [39].

Qinghai Province is a typical arid and semi-arid region within China, and is also a key province for PV development [32]. Under the support of multiple policies [40] the PV industry in Qinghai Province has achieved rapid development. However, the province features a fragile ecological environment, with degradation phenomena occurring in some regions [41]. In the short term, PV construction may damage surface vegetation and soil layers [32,42,43]; in the long run, nevertheless, it may facilitate the gradual restoration of the ecological environment by altering microclimates [35,36]. To conduct a comprehensive analysis of the ecological impacts brought about by the large-scale development of PV, it is essential to establish a robust identification method and this method should provide highly accurate geographical location and boundary information of PV facilities. Such precise data is crucial for policymakers to formulate more optimized planning decisions based on the analysis results [44]. The scientific community employs remote sensing techniques to identify power plants, primarily through manual interpretation of high-resolution imagery or by training deep learning models with encoder–decoder structures to predict features from raw satellite images [45]. Both visual interpretation and sample collection undermine the efficiency of PV identification, as they incur significant labor and time costs. When it comes to accuracy, challenges initially emerge from the images themselves, and most studies only carry out cloud removal and compositing on original images [45], with residual noise left unprocessed, which can hinder recognition. These problems are further compounded by the drawbacks of recognition models: during the training of the classic encoder–decoder model, it tends to lose both deep-level and shallow-level feature information of PV systems, which leads to poor model stability and further causes issues such as adjacent adhesion and misidentification of other ground objects, making high-precision recognition of PV systems a significant challenge [46].

To fill these gaps, we developed an automated method integrating Google Earth Engine (GEE) to generate high-quality sample datasets, the image samples underwent principal component analysis (PCA) processing. Our study aims to (1) assess the performance of different deep learning models, including encoder–decoder architectures and high-resolution-maintaining models, to screen out the optimal one for PV plant identification; (2) utilize the selected best model to identify PV plant locations in Qinghai Province; and (3) analyze the spatial distribution characteristics and corresponding land use types of the identified PV plants in Qinghai Province, provide comprehensive considerations and scientific guidance for the coordinated development of PV plants and ecological conservation in the region. To our knowledge, this study systematically evaluates multiple deep learning models for PV plant identification in Qinghai Province. Integrating the optimal model with spatial distribution and land use analysis of PV plants, it offers reliable technical reference for large-scale, accurate PV plant identification in arid and semi-arid regions, supporting the formulation of science-based strategies for sustainable PV industry development and ecological risk mitigation.

2. Materials and Methods

2.1. Study Area

The study area is located in Qinghai Province (Figure 1, 31°36′–39°19′N, 89°35′–103°04′E), within the northeastern region of the Qinghai–Tibet Plateau in China, and encompasses two prefecture-level cities along with six autonomous prefectures. The region experiences extended periods of sunshine and elevated levels of solar radiation intensity, with annual total radiation reaching 5860 to 7400 MJ/m², elevations range from 1708 to 6452 m, with an average altitude of 4061 m [47]. Geomorphological units such as the Qaidam Basin and the Gonghe Basin are characterized by flat, open terrain with extensive tracts of unused desert and Gobi land. Therefore, Qinghai possesses world-leading potential for solar energy development [48], and this sparsely populated land has become a testing ground for China’s solar energy advancement [49]. Taratan in Gonghe County, once ravaged by sandstorms, has now been transformed into the world’s largest PV park. In recent years, efforts have focused on utility-scale PV construction [50]. By the end of 2024, the total installed capacity across the region had reached 36.317 GW [51]. The widespread presence of bare ground, seasonal snow cover, and other background features in this region often forms a “different object-same spectrum” characteristic alongside PV panels, providing an excellent test for the model’s training effectiveness. Furthermore, as a typical arid and semi-arid area, the results obtained here can also reveal ecological patterns for PV installations in other arid zones, offering prior knowledge for further research on integrating PV construction with ecological conservation across arid regions as a whole.

2.2. Sample Preparation

In order to precisely identify the distribution of PV panels in the study area for 2024, this paper utilizes 2021 vector data, which systematically records the geographic coordinates and morphological characteristics of the majority of PV arrays in 2021, for sample preparation. To circumvent the inefficiency of extensive manual labeling, an automated methodology was adopted to directly generate experimental samples and their corresponding annotation information from the vector dataset. Nevertheless, the inherent limitations of vector data, such as incomplete records for a small subset of entries and the interspersion of extensive bare ground areas within specific PV array clusters, may introduce confounding interference during subsequent identification processes, the construction of a qualified sample set necessitated the implementation of a multi-stage quality screening protocol. The core procedures of this protocol are outlined as follows (Figure 2):

• Consolidation of PV vector polygons. Dispersed vector polygons were aggregated by applying a distance threshold of 12,000 m, which was determined through repeated experiments to achieve the optimal visual effect, and all polygons within this distance range were classified into the same cluster. In the present study, a total of 374 vector polygons were initially grouped into 29 clusters. Experimental findings revealed that excessively small cluster areas led to an insufficient number of pixels in PV regions, which made the spectral characteristics of these regions statistically insignificant. Consequently, clusters with an area smaller than 1.8 hectares (ha) were excluded, and the final consolidated data consisted of 19 clusters. To support subsequent processing that specifically targets vector polygon areas (rather than the entire Qinghai Province), a 2 km buffer zone was established around the centroid of each consolidated cluster.
• Satellite imagery data synthesis. Sentinel-2 satellite imagery (10 m resolution) acquired in 2021 was retrieved and synthesized through the GEE platform. An excessively long temporal span might result in capturing PV facilities at different construction stages, thereby causing ambiguity between completed and unfinished areas; on the contrary, an overly short temporal span could lead to an insufficient quantity of high-quality images with minimal cloud coverage. To balance these constraints, the imagery was synthesized at monthly, bimonthly, and quarterly intervals, respectively. The Best Clear-Sky Composition Method was adopted for image compositing, which involves selecting pixels with the lowest cloud cover at each location across all available image scenes. Initially, eleven bands [52] were selected for analysis, including NDBI, NDPI, Band 8, Band 2, slope, Band 12, elevation, SAVI, NDVI, Band 4, and Band 3 (the last of which was included for visualization purposes). To reduce computational burden and improve training efficiency in subsequent modeling processes, PCA was applied to these eleven bands. Finally, the top five principal components were retained, which preserved 95.6% of the original information of the bands.
• Calculation of indicators for evaluating image quality. To evaluate the quality of images, this study further synthesized a PV high-probability distribution map based on the PCA-processed images. The generation of this map entailed manually selecting 20 PV samples and 20 non-PV samples within the GEE platform (Figure 3), followed by the application of the platform’s built-in Random Forest algorithm. For each generated map, the proportion of the total vector area occupied by high-probability PV zones was calculated. Considering that PV zones generally maintain stability over a one-year period, a significant decrease in this proportion usually indicates that the imagery failed to accurately identify the PV areas; therefore, this proportion was employed as a proxy indicator for assessing image quality.
• Acquisition of the sample set via dual screening (Figure 4). High-quality imagery was first selected, and within the GEE platform, a 256 × 256 grid cell was generated using the buffer zones (from step 1) as spatial boundaries. For each grid cell that contains PV vector surface area, two screening criteria were applied: the first criterion is whether the proportion of high-probability PV area within the grid cell relative to the vector surface area exceeds 0.8; the second criterion is whether the area classified as high-probability PV outside the vector regions accounts for less than 10% of the entire grid cell. These threshold parameters were determined empirically through multiple experimental trials, aiming to balance sample quality and sample quantity. Only the grid cells that meet both criteria had their corresponding imagery and binary label masks exported, which constitute the final sample set. The size of the exported sample set shows a positive correlation with sample quality.
• Sample set expansion. Owing to the limited number of samples obtained from the 2021 dataset, this study adopted five synthesis methods (selected from 2021), specifically including September (single-month synthesis), August-September and September-October (bimonthly synthesis), as well as July-August-September and September-October-November (trimonthly synthesis), in line with the identical methodology outlined in step (4), sample sets were generated for each of the other four years (2019, 2020, 2022, and 2023). This process ultimately resulted in a total of 1871 training samples and 569 validation samples, corresponding to an approximate training-to-validation ratio of 80:20.

Figure 2. Sample preparation and subsequent procedures.

Figure 2. Sample preparation and subsequent procedures.

Figure 3. Sample identification. The red four-pointed stars and blue circles in the figure represent PV and non-PV point samples, respectively. Some points are accompanied by enlarged diagrams, with the yellow areas inside indicating high-probability PV distribution zones.

Figure 3. Sample identification. The red four-pointed stars and blue circles in the figure represent PV and non-PV point samples, respectively. Some points are accompanied by enlarged diagrams, with the yellow areas inside indicating high-probability PV distribution zones.

Figure 4. Illustration of selection and rejection criteria in the screening workflow using five representative grid examples.

Figure 4. Illustration of selection and rejection criteria in the screening workflow using five representative grid examples.

2.3. Deep Learning Algorithm

2.3.1. Model Selection

Given the two challenges of overfitting risks and feature utilization inefficiency inherent in small-sample binary segmentation tasks, it is crucial to maximize the utilization of limited labeled data to extract the features. Furthermore, to compare the differences in PV boundary extraction performance between the encoder–decoder structure and the full-path high-resolution structure, two representative models were selected from each of these two structural frameworks. Correspondingly, the classical UNet, Attention UNet, HRNetv2, and HRNet + OCR models were chosen as the core deep learning architectures for this study. The key advantages of each model are summarized in the table below (

Table 1. Summary of advantages of semantic segmentation models.

Semantic Segmentation Models	Advantages	References
UNet	Simple structure, easy to implement, and fast inference speed.	[53]
Attention UNet	Precision targeting of designated areas, adapting to dynamic target conditions.	[54]
HRNetv2	Maintain high-resolution features throughout the process, precisely preserve spatial details, aggregate multi-resolution features, and fully extract global-local information.	[55]
HRNet + OCR	Accurately capture object-level context, overcoming the limitations of local features.	[56]

).

2.3.2. Training and Prediction

Each model was deployed using the same parameters, including epoch number, process metrics, loss type, learning rate strategy and best model indicator; see

Table 2. Model deployment parameters.

Parameters	Deployment
epoch	50
process metrics	Dice coefficient, IoU coefficient, accuracy, loss
loss type	binary cross-entropy loss
learning rate strategy	adaptive
best model indicator	validation set Dice coefficient

for details.

We conducted a survey of PV construction across all districts and counties in Qinghai Province to narrow down the identification area. Synthetically combined imagery from July, August, and September 2024 was processed using PCA dimensionality reduction, resulting in 545 images that conformed to the input specifications of the deep learning model for predictive analysis.

2.3.3. Data Post-Processing and Validation

A total of 545 predicted binary images of PV arrays were subjected to image fusion and post-processing procedures (Figure 5). Specifically, internal voids within the binary images were filled to ensure spatial integrity, while external noise was eliminated to reduce interference from non-target features. Subsequently, polygon approximation algorithms were employed to optimize and straighten the edge contours of the PV array regions, followed by vectorization operations to convert the processed raster images into vector polygon datasets for subsequent spatial analysis. To validate and correct the vectorization results, visual inspection was conducted within a Geographic Information System (GIS) platform. In order to ensure the reliability of the dataset presented above, a cross-validation was conducted against the Global Power Plant Database (GPPD), whose latest update was timestamped on 16 October 2024 (Table S1).

2.3.4. Attributes Calculation and Analysis

To further characterize the PV arrays, the construction year of each vectorized PV array was inferred by integrating the 2024 PV power plant inventory data with Sentinel-2 multispectral remote sensing imagery, this integration enabled cross-validation of temporal information and spatial location to ensure the reliability of the inferred construction years. Meanwhile, the spatial extent of each PV array was calculated and quantified (ha) as a key metric for assessing its land occupation scale. Furthermore, to retrospectively analyze the land use types prior to PV construction, the Dynamic World land cover dataset was incorporated into the analysis framework. For each vectorized PV array polygon, land cover data corresponding to the two-year period preceding the inferred construction year were extracted and analyzed. This retrospective analysis allowed for the identification of the original landcover types of the areas occupied by PV power plants, thereby providing a comprehensive basis for evaluating the land use change processes associated with PV power plant development. Additionally, Gross domestic product (GDP) data in 2024 were collected for each county-level administrative unit, geographic location data of major urban centers were collected and spatially mapped, so as to clarify the socioeconomic associations of each PV plant.

3. Results

3.1. Methodological Evaluation and Results Validation

Prior to sample collection, PCA was conducted on the raw imagery, aimed to reduce the number of bands and eliminate noise. The contribution rates of the principal components are presented in Figure 6a. The first five principal components (PC1 to PC5) were selected, which collectively accounted for a contribution rate of 95.6%. To realize automation, the relationship between the proportion of high-probability PV areas obtained via different synthesis methods and the number of exported samples was tested, Figure 6b illustrates their positive correlation. Figure 6c shows the proportion of high-probability PV areas corresponding to 33 synthesis methods in 2021, where the gray dashed line represents the threshold of 0.75. Synthesis methods that exceeded this threshold were used to generate candidate samples, and the top five synthesis methods highlighted in the red box achieved proportions exceeding 0.9, which were adopted as candidate synthesis methods for other years. Figure 7 presents the process metrics of the four models, with each model converging independently but reaching different convergence limits. The two models with UNet as the backbone network exhibited outstanding performance on the training set: their loss approached to 0, and the Dice coefficient, Intersection over Union (IoU), and accuracy all exceeded 0.97. In contrast, the two models with HRNet as the backbone network showed slightly inferior performance on the training set, with loss exceeding 0.2, though their Dice coefficient, IoU, and accuracy still surpassed 0.94. However, on the validation set, the performance of the UNet-based models decreased significantly, while the HRNet-based models maintained stable performance, among the four models, HRNetv2 achieved the highest validation Dice score of 0.9463.

According to the GPPD records, there are 144 entries of solar power plants within Qinghai Province (Figure 8). Among these, 119 entries showed a direct intersection with the dataset generated by this study. For the remaining 25 non-overlapping entries in the GPPD, a systematic visual verification was implemented to resolve discrepancies, and only 3 of these 25 entries corresponded to actual solar power plants that were not detected by the methodology employed in this study. Additionally, this study successfully identified a total of 441 PV surface units in Qinghai Province, representing an increment of 297 entries compared to the GPPD’s records.

3.2. Spatial and Temporal Distribution of PV Plants

From a temporal perspective, the PV area installed between 2020 and 2024 accounted for 63.5% of the total PV area deployed prior to 2024, which exceeded the cumulative PV area installed before 2019. Figure 9a–f illustrates the evolutionary process of the PV landscape from 2019 to 2024. Before 2019, a number of large-scale PV bases were developed along the edges of the Qaidam Basin and within the Gonghe Basin, while small-scale PV projects were constructed in the Hehuang Valley and impoverished regions at that time. After 2020, in addition to the expansion of existing large-scale PV facilities, new PV parks [57] emerged in areas such as the Qaidam Basin. Figure 9g presents the annual growth rate of PV areas, which decreased from 2020 and reached a low of 6.6% in 2022, followed by a sharp acceleration to 30.6% by 2024.

The results show that most PV installations are distributed in areas with an elevation below 2500 m and a slope gradient of less than 2 degrees, the distribution of aspects is relatively extensive, except for west-facing slopes, PV installations are evenly distributed across all other aspects (Figure 10). From the economic perspective (Figure 11), Qinghai classifies its economic development into three tiers based on whether the annual GDP exceeds 2 billion yuan and 10 billion yuan. There is a positive correlation between the installed area of PV systems and economic development. Specifically, the first tier has the largest PV installed area, reaching 52,397 ha; the second tier ranks second with a PV installed area of 7823 ha; and the third tier has the smallest PV installed area, merely 80 ha. In terms of county-level administrative units, the three counties/cities with the largest PV installed areas all belong to the first echelon; namely, Gonghe County (29,156 ha), Golmud City (14,342 ha) and Delingha City (7865 ha). Additionally, 68.6% of PV power stations are located within 20 km of urban centers, while less than 7% are situated more than 40 km away from urban centers. Therefore, based on the above statistical results, the most common site selection criteria for PV stations in Qinghai are summarized as follows: elevation below 2500 m, slope gradient less than 2°, non-west-facing slope orientation, location within 20 km of urban centers, and economically developed regions within the province.

Figure 9. Spatial and temporal distribution of PV power plants. In order from (a–f), the evolutionary process of the PV landscape from 2019 to 2024 are shown respectively, and the PV is plotted at four times the scale of the cluster’s center, making it possible to obtain both location and shape information simultaneously. Graph (g), which combines a bar chart and a line chart, the bars represent the annual newly added PV area (ha), where “2019 and Before” denotes the total area of all PV facilities built in 2019 and earlier (ha), the line chart shows the annual growth rate, calculated as (newly added area/total area of all previously built facilities).

Figure 9. Spatial and temporal distribution of PV power plants. In order from (a–f), the evolutionary process of the PV landscape from 2019 to 2024 are shown respectively, and the PV is plotted at four times the scale of the cluster’s center, making it possible to obtain both location and shape information simultaneously. Graph (g), which combines a bar chart and a line chart, the bars represent the annual newly added PV area (ha), where “2019 and Before” denotes the total area of all PV facilities built in 2019 and earlier (ha), the line chart shows the annual growth rate, calculated as (newly added area/total area of all previously built facilities).

Figure 10. Terrain characterizes of PV, including elevation, slope and aspect. (a) is a map synthesized from three bands: elevation, slope, and aspect, the red circles and green circles mark the PV facilities built in 2019 and earlier, and those built in 2020 and later, respectively, the size of the circles represents the area (hm²). (b) below follows the same principle and presents statistical distributions across three dimensions. The two numbers highlighted in red within the figure are circled to prevent any misinterpretation as a single digit.

Figure 10. Terrain characterizes of PV, including elevation, slope and aspect. (a) is a map synthesized from three bands: elevation, slope, and aspect, the red circles and green circles mark the PV facilities built in 2019 and earlier, and those built in 2020 and later, respectively, the size of the circles represents the area (hm²). (b) below follows the same principle and presents statistical distributions across three dimensions. The two numbers highlighted in red within the figure are circled to prevent any misinterpretation as a single digit.

Figure 11. Relationship between economic development and PV. (a) shows the relationship between the county-level GDP and PV area of each county in Qinghai Province in 2024, the vertical axis is arranged from top to bottom in descending order of GDP, the green bars represent GDP, and the yellow bars represent PV area. Counties and cities enclosed in red, blue, and purple boxes are divided into the first, second, and third tiers by 2 billion yuan and 10 billion yuan, respectively; the basemap of (b) is the 2024 GDP. Red circles and green circles mark PV facilities built in 2019 and earlier, and those built in 2020 and later, respectively, with the size of the circles representing the area. “Towns” denotes major towns; (c) shows the distance distribution between PV facilities and their nearest towns.

Figure 11. Relationship between economic development and PV. (a) shows the relationship between the county-level GDP and PV area of each county in Qinghai Province in 2024, the vertical axis is arranged from top to bottom in descending order of GDP, the green bars represent GDP, and the yellow bars represent PV area. Counties and cities enclosed in red, blue, and purple boxes are divided into the first, second, and third tiers by 2 billion yuan and 10 billion yuan, respectively; the basemap of (b) is the 2024 GDP. Red circles and green circles mark PV facilities built in 2019 and earlier, and those built in 2020 and later, respectively, with the size of the circles representing the area. “Towns” denotes major towns; (c) shows the distance distribution between PV facilities and their nearest towns.

3.3. Land Use for PV Plants

Figure 12 illustrates the various land use types associated with newly constructed PV facilities from 2020 to 2024. Notably, bare land constituted the predominant category, comprising 88.7% (34,188.3 ha) of the total land converted for PV utilization. This was followed by grassland, shrubland, and cropland, which accounted for 6.9% (2650.6 ha), 3.9% (1492 ha), and 0.5% (214.8 ha), respectively. From a chronological perspective, the proportion of bare land used for PV facilities decreased from 2020 and reached its lowest point in 2022, after which it increased rapidly, this trend is consistent with the overall annual growth rate of PV installations. Grassland began to be converted into PV areas starting in 2022, with annual increases in the converted area in subsequent years. Shrubland experienced minor conversions to PV use in 2021 and 2023. Cropland showed the smallest extent of conversion, with such conversion occurring only in 2024 and distributed across Huanzhong District, Gangcha County, and Jianzha County.

4. Discussion

Methodologically, this study introduced enhancements in the identification of PVs across three key dimensions: the processing of raw data, the approach to sample generation, and the selection of model structure. First, it adopted PCA for raw image processing. Although PCA is a conventional data processing technique, its application to deep learning datasets remains infrequent, primarily due to the substantial computational resources and time required for this procedure. Nevertheless, its effectiveness is indisputable: this approach reduces the number of input bands while suppressing noise, thereby enabling the successful generation of high-probability PV distribution maps, only using raw multi-band imagery with 20 positive and 20 negative samples would be insufficient, which underscores the critical role of PCA in feature dimensionality reduction. However, PCA emerged as the most time-consuming step in the PV identification process, taking from 2 to 7 h and consuming extensive computational resources on the GEE platform. Memory constraints necessitated strict control over the size of the study area to ensure the successful completion of PCA operations, which highlights the importance of prior knowledge regarding PV plants, as it allows for the exclusion of areas where PV stations definitely do not exist and focuses only on probable or planned sites. Second, the sample acquisition process is fully automated. Through experimental determination of thresholds and multi-stage filtering, manual sample collection is avoided, thus reducing labor costs. Specifically, samples were generated using 2021 data, while supplementary imagery from 2019, 2020, 2022, and 2023 was incorporated. This approach was adopted because the weather conditions and cloud cover during the corresponding months of each year are approximately consistent, rendering the supplementation analogous to a data augmentation technique that enriches spectral data within the same region. Furthermore, only the five synthesis methods with the highest quality were selected, and a dual filtering process was implemented to further ensure sample quality. Compared with manual annotation, this method demonstrates a marked increase in speed and efficiency. However, due to inherent limitations in the quality of raw images, its accuracy may not match the flexibility of manual annotation (Figure 13). Third, in terms of model selection, four models spanning two architectural frameworks, including encoder–decoder and full high-resolution structures, were compared. Previous studies have predominantly utilized encoder–decoder models, which may compromise boundary extraction accuracy to a certain extent [45,58]. The HRNetv2 model was identified as the optimal solution, as it exhibited superior performance on the validation set compared to the classical UNet, Attention UNet, and HRNet + OCR models, demonstrating excellent resistance to overfitting in small-sample semantic segmentation tasks. HRNetv2 maintains high-resolution representations throughout the network via a multi-branch parallel architecture, which enables continuous fusion of features across multiple resolutions, this design preserves fine-grained spatial details and alleviates the information loss commonly associated with dimensionality reduction processes. More importantly, this multi-resolution fusion mechanism allows the model to learn generalized semantic patterns rather than merely memorizing features of the training samples, which effectively mitigates overfitting and ensures consistent segmentation performance—an advantage particularly prominent in small-sample learning scenarios. The underperformance of the HRNet + OCR model relative to the standalone HRNetv2 is presumably due to its excessive parameter size and complexity, which cannot be adequately supported by limited training samples. Future studies with larger and higher-quality datasets may facilitate further exploration of such complex architectures. Consistent with the findings of Chang et al. [58], who demonstrated the advantage of HRNetv2 in pixel-level classification tasks for automated land cover classification of high-resolution remote sensing imagery, achieving a higher mean Intersection over Union (mIoU) of 77.5% compared to FCN (62.4%), UNet (67.9%), and DeeplabV3+ (76.9%), our results further confirm its superiority in pixel-level binary segmentation. This consistency indicates the model’s robustness across related pixel-wise prediction tasks.

The validation results indicated a high degree of consistency between the two sources for the overlapping samples. For the 25 non-overlapping entries in the GPPD, except for 3 entries that were not detected, the other 22 entries were confirmed to be data anomalies (e.g., positional mismatches, or outdated entries) in the GPPD, thus excluding them from valid comparison. 297 newly detected entries, coupled with the validation results of non-overlapping samples, collectively demonstrates that the dataset presented in this paper exhibits superior spatial granularity (i.e., capturing more localized PV installations) and enhanced temporal currency (i.e., reflecting the latest status of solar power infrastructure) within the study area.

From a chronological perspective on the deployment of PV installations, our results indicate that the cumulative PV installed capacity commissioned between 2020 and 2024 (~38,547 ha) has exceeded the total installed capacity prior to 2019 (~22,168 ha). This phenomenon explicitly reflects the accelerated expansion trend of PV infrastructure construction in recent years. Before 2019, large-scale PV projects in the region were geographically concentrated in the Qaidam Basin and Gonghe Basin. A typical case is the Taratan PV Park [59], which currently holds the status of the largest PV park in Asia. These basin areas possess inherent geographical and climatic advantages, characterized by flat and open terrain as well as abundant solar radiation resources, thereby meeting the fundamental site requirements for large-scale PV installations [32]. In contrast, small and medium-sized PV projects during the same period were mainly distributed in the Hehuang Valley and some impoverished regions at that time. This spatial distribution pattern was largely driven by the “PV poverty alleviation” initiative [34], a key policy implemented in the national poverty eradication campaign, which aimed to integrate renewable energy development with regional poverty alleviation goals. Since 2020, with the formal implementation of the national dual-carbon strategy and the release of the 2022 plan [33] for the development of large-scale wind and solar power bases in northwest China, the development focus of the PV industry has further shifted toward large-scale base construction, site selection for PV projects has shown a natural tendency to concentrate in the intersection area of the vast basin regions and first-tier economic zones within the province. In the specific site selection process, priority is consistently given to areas with lower elevations and flat terrain, as such geographical conditions can effectively reduce engineering difficulties and construction costs. The aforementioned site selection logic further explains the spatial agglomeration of PV projects in the Qaidam Basin, where most installations are distributed along the basin’s periphery rather than its interior. The primary reason for this pattern is the combination of convenient infrastructure access and proximity to urban centers in peripheral areas: these locations not only enable more feasible and cost-effective construction of supporting infrastructure and power transmission facilities but also benefit from shorter distances to urban centers, supporting project operation and power integration. In this study, only PV completed between 2020 and 2024 were confirmed, primarily due to the insufficient availability of Sentinel-2 imagery prior to 2018. In future research, an attempt could be made to determine the completion time using Landsat imagery, as the confirmation of completion time does not require high spatial resolution like Sentinel-2.

Our results found that PV installations primarily use four types of land: bare land, grassland, shrubland, and cropland. Bare land is the most common, making up 88.7% of the land used for PV projects. Our study focuses specifically on the land use patterns of newly constructed PV facilities during the period 2020–2024, noting that the preference for bare land is due to a strategic move towards building large-scale PV bases. Qinghai Province, as a typical arid and semi-arid region, features extensive distributions of sandy land, saline-alkali land, deserts, and Gobi terrain, all of which are categorized as bare land in land use classifications. These land types inherently align with the demand for large, flat areas required by large-scale PV projects, thus explaining the overwhelming proportion of bare land in PV land use. A growing body of research reports [60] has confirmed that PV facility construction in desert regions contributes to desert greening. The deployment of PV panels creates microenvironmental conditions conducive to vegetation survival: for instance, in Hainan Tibetan Autonomous Prefecture, ecological restoration measures (including the establishment of windbreak and sand-fixation forests, roadside landscape vegetation, and large-scale planting of drought-tolerant species such as poplars and tamarisks) were implemented at the Taratan PV Park, resulting in a vegetation coverage rate of over 80% within the park [61]. Mechanistically, PV panels modulate the spatiotemporal distribution of light radiation, thermal regimes, and soil moisture dynamics via physical shading effects and interception of environmental resources, and this modulation induces perturbations in the local microenvironmental conditions, thereby exerting a selective filtering effect on vegetation communities, favoring the colonization and proliferation of plant species with ecological niches compatible with the modified microhabitats, which in turn promotes the recovery of drought-tolerant and shade-tolerant vegetation [62,63]. In the Hehuang Valley region, grassland serves as the primary land use type for PV development, small-to-medium-sized PV projects prevailed in this area prior to 2019. According to research by Xia et al. [57], between 2007 and 2019, grassland accounted for 61% of PV land use across Qinghai and Ningxia Hui Autonomous Region [57]. This finding may point to the possibility that the historical contribution of grassland to PV land use is more substantial than current statistics indicate. However, the conversion of grassland to PV sites disrupts the original vegetation community structure, potentially driving vegetation succession toward shade-tolerant species. Ecologically, this transformation may lead to a decline in biodiversity under PV panels, while the inter-panel areas may exhibit enhanced vegetation recovery potential due to water redistribution and shading effects. For bare land and grasslands in the process of gradual ecological recovery, the rangevoltaic model emerges as a proactive and effective approach. This model enhances soil carbon sequestration capacity, supports pollinator habitats, strengthens soil erosion resistance, and improves water retention [39]. Additionally, it reduces wind and water erosion, mitigates dust emissions from disturbed soils, and prevents dust accumulation on PV panels, vegetation growth reduces the water demand for panel cleaning and dust suppression, with excess moisture further nourishing understory vegetation, eventually achieving a synergistic balance between grassland ecological conservation and solar energy production. Shrubland, as a more drought-resistant vegetation type than grassland in arid regions, benefits from PV panel shading, which enhances the survival rate of shrub species. While farmland constitutes the largest share of PV land use globally [39], it accounts for merely 0.5% of PV land use in the study region. This low proportion is attributed to the arid climate of the area and the inherent scarcity of arable land, which collectively drive the preference for degraded desert areas for PV development. Although the agrivoltaics system can create synergies between solar energy generation and agricultural production, the ecological and economic trade-offs associated with this model remain insufficiently explored. In desert-dominated regions such as the study area, prioritizing the integration of PV development with desertification control is therefore deemed a better strategy. In summary, PV project development on any land use type presents both advantages and inherent challenges, construction in fragile arid and semi-arid ecosystems inevitably imposes additional environmental disturbances. Under such circumstances, a rigorous assessment of trade-offs between national energy strategy implementation and the maintenance of ecosystem services is imperative. The pathways and intensity of the ecological impacts of PV power plants are closely linked to their geographic location and the pre-construction land use type, thus differentiated ecological management strategies must be formulated and implemented throughout the entire lifecycle of PV development, with each strategy tailored to the regional ecological baseline conditions.

While this study focused specifically on Qinghai Province, a representative region of arid and semi-arid environments, its findings are constrained in generalizability by two key factors: the limited spatial extent of the study area, and the rapid global expansion of PV infrastructure alongside significant regional variations in natural endowments, ecological baselines, and development models. Consequently, the application of these conclusions to other arid and semi-arid regions requires adaptive adjustments based on specific local contexts. Given the urgent academic and practical need to clarify the ecosystem impacts of large-scale PV development, particularly amid efforts to advance carbon peaking and carbon neutrality goals, future research should prioritize two core tasks. First, it should refine the algorithms developed in this study to enhance their accuracy and adaptability. Second, it should expand the research scope to broader geographical scales. Such efforts will enable a more comprehensive assessment of the ecological effects of PV construction, thereby supporting more scientifically robust site selection and management strategies for PV projects. This approach will provide a stronger foundational basis for achieving synergistic development between PV energy exploitation and ecological conservation in arid regions.

5. Conclusions

This study optimizes a remote sensing-based framework integrating PCA, GEE automated sampling, and deep learning to address PV power station identification challenges in arid and semi-arid regions, with Qinghai Province as the case study. The optimized method proves effective for small-sample PV segmentation, with the HRNetv2 model outperforming others (validation Dice = 0.9463); for example, it identified 441 PV units in Qinghai, 297 more than the GPPD, thus showing superior spatial-temporal accuracy. Qinghai’s PV distribution reflects both policy and geographical influences: installations from 2020 to 2024 account for 63.5% of the total pre-2024 area, aligning with China’s “dual carbon” strategy, and PV projects cluster in areas with elevation less than 2500 m, slope less than 2°, proximity to cities, and economically developed regions (GDP more than 10 billion yuan) hold 52,397 ha of installations. In terms of land use, 88.7% of PV development in 2020–2024 utilized bare ground, contributing to desert greening, while 6.9% used grassland and only 0.5% used cropland. When developing PV on bare land, it is imperative to implement subsequent under-panel land management strategies to enhance both ecological and economic values. Priority should be given to avoiding PV installation in cultivated areas; if such installation is unavoidable, the crops grown on these cultivated lands must not be adversely affected by the PV facilities. In summary, the development of the PV industry should not compromise the environment or sacrifice economic benefits. Instead, comprehensive schemes should be formulated to achieve a win-win scenario for both PV development and ecological-economic sustainability. Supported by national policies, the PV industry in Qinghai has witnessed robust development momentum. However, PV construction exerts substantial impacts on the original ecological pattern; as an ecologically fragile region, Qinghai requires extra prudence in PV development. Limitations include the method’s limited generalizability and constraints from small sample sizes on complex models, so future work should refine algorithms and expand the research scope to global arid zones. Overall, this study provides a robust PV identification framework and offers guidance for synergistic PV development and ecological conservation in Qinghai and beyond.