Introduction

Globally, climate change is posing considerable challenges to human society, especially for energy supply and transition. Accelerating the adoption and integration of renewable energy is a crucial step in limiting global warming to within 2 °C. Photovoltaic power plants (PPPs), recognized as an effective alternative to fossil fuel power stations, have the potential to greatly reduce operational costs and carbon emissions, alongside wind energy1,2. Consequently, the development of PPPs has surged globally, with China acknowledged as the leading country in both market share and growth rate3,4. However, the establishment of PPPs requires extensive land for solar panel installation, leading to substantial changes in land use and land cover.

Remote sensing data facilitate the mapping of PPPs, which has been extensively carried out at regional5,6,7, national3,8, and global1 scales to investigate the underlying driving mechanisms behind their development. Most studies on the potential and site selection of PPPs have focused on constraining factors such as solar radiation distribution, socioeconomic conditions, natural environments, and policies9,10,11. Another focus of research examines the environmental impacts of constructing PPPs. During the initial construction phase, existing surface vegetation and biological soil crusts are often cleared, and the land is leveled to facilitate the installation of metal supports and photovoltaic (PV) panels, altering the original surface patterns. Once operational, the solar panels continuously shade portions of the land and redistribute rainfall12,13, ultimately affecting nutrient cycling14, energy balance, and biodiversity15.

Overall, studies based on satellite data and field monitoring suggest that establishing PPPs in desert regions contributes to the restoration of local vegetation16,17,18,19,20. Conversely, some studies highlight the negative impacts of constructing PPPs on local vegetation in other areas21,22,23,24,25. Consequently, stringent regulations have been implemented to optimize the spatial layout of PPPs to mitigate their negative impacts on vegetation. In China, electricity demand is concentrated in the developed coastal areas, where vegetation is abundant but solar energy resources are relatively scarce26. Therefore, the new PPP projects are often located far from end-users, necessitating large-scale centralized deployment27 and the construction of long-distance ultra-high voltage transmission networks to ensure system stability and safety. This notably increases the complexity, timeline, and overall cost of the projects. In addition, the construction of transmission and distribution infrastructure typically requires longer timelines and spans larger regions than PPPs. As a result, many areas in China have temporarily halted PPP development28,29,30,31, which could ultimately impede the progress of the power sector toward carbon neutrality.

However, the true impacts of PPPs on vegetation have not been consistently determined at the regional scale, especially concerning the vegetation under the PV panels. On one hand, the widespread distribution of PPPs, coupled with dramatic variations in geographical locations and solar panel layouts, makes it challenging to derive consistent quantitative conclusions from field monitoring studies across different stations32,33. On the other hand, research based on satellite data seems to yield relatively uniform findings34. For instance, the large-scale deployment of PPPs in certain areas has promoted desert greening, with approximately 21% of PPPs facilities in China showing an increase in greenness compared to their pre-construction levels20,23. However, those studies often overlook the unique laminated structure of PV panels and the vegetation beneath them, assuming that satellite-derived vegetation information accurately represents the true state of surface vegetation in PPPs. In fact, PPPs present a discontinuous vegetative surface, with rows of PV panels alternating with background areas such as bare soil or other vegetation, and vegetation may also grow beneath the panels (Supplementary Fig. 1)17,19,35,36. This special inter-row heterogeneity adversely affects multispectral satellite data collected by public research missions (such as Sentinel-2 and Landsat 8/9)37. Research on the vegetation retrieval capabilities of PPPs using unmanned aerial vehicle (UAV), Sentinel, and Landsat satellite data indicates that satellite images are unable to accurately distinguish different land use types within PPPs. In addition, satellite images cannot capture the differences between similar land use types both inside and outside the PPPs38. Consequently, the unique spectral information of PPPs observed by satellites is obscured by a complex signal that lacks details of the vegetation beneath the panels, which could lead to misleading conclusions if analyzed directly39.

To address these issues, the most reliable, affordable, and technically scalable solution should be the combination of continuous, free, and scientifically processed medium-resolution satellite data with UAV near-ground remote sensing data and field monitoring data. This approach leverages the high resolution of UAV imagery and the under-panel vegetation information from ground monitoring, offering a cost-effective alternative to frequent data acquisition40,41. In addition, it overcomes the limitations of both data sources in obtaining historical data for multi-temporal analyses when prior monitoring has not been conducted. However, research on inter-row layouts is limited, mainly focusing on agricultural contexts such as orchards42,43 and vineyards39,44. For instance, effective collaboration among Sentinel-2, UAV data, and field data has led to more accurate vegetation mapping, target detection45,46,47, and crop growth monitoring48. In these studies, the Normalized Difference Vegetation Index (NDVI) has been widely utilized due to its clear definition, ease of calculation, and relative robustness against issues commonly encountered in other vegetation classification methods49. Particularly in studies examining environmental impacts of PPPs, NDVI or vegetation coverage is commonly used to indicate vegetation status, as it is closely related to various phenomena such as surface albedo, biomass, leaf area index, and vegetation phenology19,20,50,51,52.

Therefore, for quantitative assessments of the impacts of PV panel installations on vegetation at the regional scale, it is crucial to accurately retrieve the true conditions of vegetation within PPPs from satellite-recorded signals. In this study, we propose a solution focused on NDVI from Sentinel-2 data by utilizing UAV images to mediate field and satellite data integration. Our objectives are to: (1) quantitatively evaluate the relationship of vegetation under and between the panels; (2) develop a framework to use UAV data to correct the Sentinel-2 NDVI data for evaluating the true conditions of vegetation in PPPs; and (3) apply our method in nine national clean energy bases. Our findings will also serve as a reference to facilitate more refined planning of PPPs in vegetation-rich areas, thereby unlocking greater potential for PPP development closer to power consumption centers.

Results

Quantifying the relationship between vegetation at BP and UP

According to the NDVI data collected manually, the average NDVI of the vegetation within PPPs was 0.25, with no significant difference between the NDVI of the vegetation under the panels (UP) and that between the panels (BP) (Fig. 1d). The NDVI distribution of vegetation in both UP and BP areas was also very similar, concentrated between −0.1 and 0.2, while it was more uniform between 0.2 and 0.8 (Supplementary Fig. 2a). Finally, 54.53% of the samples within the PV plant had an NDVI of less than 0.2 (Supplementary Fig. 2b).

Fig. 1: Relationship between NDVI of vegetation under and between photovoltaic panels using handheld and UAV data.
figure 1

a–c Liner regression between NDVI of vegetation under panels and between panels using handheld data and NDVI threshold of 0.2 for the non-photovoltaic area. a,b Regressions using data with the mean NDVI of NPV vegetation beyond 0.2 and below 0.2, respectively. c Regression using all available data. d Comparison of NDVI differences between NDVI of vegetation under panels and between panels using handheld data, raw UAV flight data, and UAV flight data after interpolation.

Full size image

Overall, there is a strong positive correlation between the vegetation at UP and BP, with an R² value of 0.821 (Fig. 1c). However, based on the distribution of handheld NDVI data (Supplementary Fig. 2) and the UAV-derived NDVI of vegetation outside PPPs, we observed a relatively distinct boundary in vegetation distribution of PPPs. Therefore, we considered categorizing PPPs into high-NDVI and low-NDVI groups to better understand the vegetation relationship between UP and BP in PPPs of different vegetation conditions. To this end, we performed a series of regression analysis within the NDVI range of 0.1 to 0.3, utilizing threshold increments of 0.05 as the gradient (Table 1). As the division threshold increased, the R² value of the regression model for high-NDVI plants decreased, while the R² value for low-NDVI plants increased. Furthermore, the RMSE value of the regression model for high-NDVI plants continuously increased, whereas the RMSE value for low-NDVI plants initially decreased and then increased. Finally, we chose 0.2 as the NDVI threshold to balance the simulation of low-NDVI and high-NDVI plants. The fitting result of high-NDVI plants decreased slightly compared to that before the division, while the fitting result of low-NDVI plants significantly improved, with the RMSE decreasing from 0.112 to 0.079 (Fig. 1a–c). After interpolating the vegetation beneath the panels, the average NDVI of the PPPs increased from 0.248 ± 0.158 to 0.298 ± 0.193, which aligned more closely with data obtained from field surveys (Fig. 1d).

Table 1 Details of the linear regression between NDVI of vegetation under panels and between panels using different NDVI thresholds of the non-photovoltaic area
Full size table

Comparisons of NDVI values between model outputs and raw data

Our modified deep learning models demonstrated exceptionally stable performance and high precision. After 75 epochs of training, the final DeepLabV3+ model achieved excellent results, with a mean Intersection over Union (mIoU) of 94.90 and a mean Accuracy (mAcc) of 0.983 (Supplementary Fig. 3 and 4). After 150 epochs of training, the modified U-Net model achieved a test accuracy of 0.945 (Supplementary Fig. 5).

The proposed correction method substantially improved the numerical accuracy and authenticity of the Sentinel data distribution. It effectively corrected NDVI values in PV plant areas while leaving those in non-PV regions unchanged. Compared to the interpolated UAV data, the NDVI bias decreased from −0.052 ± 0.150 to −0.002 ± 0.038, representing a reduction of 16.98% (Supplementary Fig. 6). From the raw Sentinel-2 data, we could only distinguish wide roads and dense vegetation areas with difficulty; the intervals, PV panels, and roads were not easily observable, unlike in UAV images (Fig. 2 a4, d4). The interpolation process reduced the low-value belt caused by the PV panels, with the effect varying on the original vegetation state and the PV panel occupancy rate (Fig. 2b–d). After the interpolated UAV data were mean-aggregated to match the resolution of Sentinel-2, the intervals and PV panels could no longer be distinguished. However, a clearer distribution pattern could be detected with greater precision. Although the prediction by the U-Net model did not achieve the same precision as the interpolated and mean-aggregated UAV data, the results were significantly better than those of Sentinel-2 NDVI in both value and distribution (Fig. 2 and Supplementary Fig. 6). Our method also significantly enhanced the accuracy of Sentinel data in assessing vegetation changes in PPPs. Using the raw Sentinel-2 NDVI data, most of the observations showed a decrease, with an average reduction of −0.089 (Fig. 3). Only 6 out of 129 observations indicated an increase, with an average gain of 0.063. However, it is important to note that the number of decreased observations is actually 64, with an average reduction of −0.153, while the number of increased observations is 26, with an average increase of 0.070, as demonstrated by the raw UAV-derived NDVI data. After interpolating the UP data from the UAV-derived NDVI, the number of decreased observations was further reduced. Only 48 observations showed a decrease in vegetation NDVI, with the average decrease narrowing to −0.118. Meanwhile, the number of increased observations rose significantly to 46, with an average increase of 0.099. The mean aggregation did not significantly alter these results. The outputs of the U-Net model achieved similar outcomes, with a slight reduction in the decrease range, narrowing to −0.112.

Fig. 2: Samples of different NDVI results of three different photovoltaic plants.
figure 2

RGB_UAV (a1–f1): UAV true color detail images. NDVI_UAV (a2–f2): UAV NDVI detail images. NDVI_UAV_int (a3–f3): NDVI detail images after UAV NDVI interpolation. NDVI_sen (a4–f4): Sentinel-2 NDVI detail images. NDVI_UAV_agg (a5–f5): NDVI detail images after mean aggregation of NDVI_UAV_int at Sentinel-2’s resolution. NDVI_sen_pre (a6–f5): NDVI detail images predicted by deep learning model.

Full size image
Fig. 3: Changes in NDVI after the construction of photovoltaic power plants, using 0.02 as the threshold to compare the area within photovoltaic power plants to the surrounding area.
figure 3

NDVI_sen: Sentinel-2 NDVI data. NDVI_UAV_agg: NDVI data from UAV interpolated and aggregated to Sentinel-2’s resolution. NDVI_sen_pre: NDVI data predicted by the deep learning model. NDVI_UAV: UAV NDVI data. NDVI_UAV_int: NDVI data after UAV NDVI interpolation. Whiskers denote 1.5 times the interquartile range, and the remaining points are outliers. The bold blue number next to each box indicates the number of independent observations in that category.

Full size image

Regarding the predictions of turning points, both the raw Sentinel data and the UAV data significantly underestimated the NDVI threshold, at 0.113 and 0.123, respectively (Fig. 4a, d). After interpolating the data under the PV panels, the NDVI turning point significantly increased to 0.212. When using the final output of our method, the turning point decreased to 0.201 (Fig. 4c).

Fig. 4: Details of the liner regression between ΔNDVI and the NDVI of the surrounding non-photovoltaic area.
figure 4

a Sentinel-2 NDVI data. b NDVI data after UAV NDVI interpolation. c NDVI data predicted by deep learning model. d UAV NDVI data. e NDVI data from UAV interpolated and aggregated to Sentinel-2’s resolution. ΔNDVI represents the changes in NDVI between the photovoltaic power plant areas and the surrounding non-photovoltaic area. NPV: Non-photovoltaic area.

Full size image

Applications of our method in nine clean energy bases

By 2020, the nine clean energy bases had approximately 10,963 PV polygons, covering an area of approximately 1155.51 km2. In terms of polygon area, 84.29% of the polygon areas were less than 0.1 km², and 13.46% of the polygon areas were between 0.1 and 1 km². The construction was primarily concentrated in grassland, bare land, and cropland, occupying 541.94, 390.06, and 173.54 km², respectively (Supplementary Fig. 7). Forest areas experienced the least development, with only 10.11 km². In terms of average area per polygon, bare land was the largest at 0.555 km², while forest areas were the smallest at 0.026 km² (Table 2). Spatially, 79.93% of the polygons were distributed in Bases 2, 3, and 5, accounting for approximately 69.00% of the total area.

Table 2 NDVI of photovoltaic power plants across different land use types
Full size table

The comparison of the final output of our proposed method to the original Sentinel data revealed that using the original Sentinel data to assess vegetation change in PPPs can be seriously misleading in terms of scale and spatial layout (Fig. 5). Using the raw Sentinel-2 NDVI data, 94.24% of the polygons were categorized as either decreased or no-change (Fig. 6c). The polygons with decreased NDVI were concentrated in Base 5 and the southern part of Base 3, with an average decrease of −0.050 (Fig. 5a). The polygons with increased NDVI were scattered across Bases 2, 3, 4, and 9, with an average increase of 0.030. Ultimately, 70.58% of the polygons had a ΔNDVI of less than 0, accounting for 1,115.07 km² of the total area (Fig. 6). Conversely, the percentage of polygons with increased NDVI rose to 31.04% from 0.06%, with an average increase of 0.049 based on the model-predicted data. This change was primarily attributed to the NDVI of most polygons that originally showed a decrease or no change but now showed an increase in Bases 3 and 5 (Fig. 6b). Meanwhile, the percentage of polygons with decreased NDVI dropped to 14.95% from 34.18%, with an average decrease of −0.042. These polygons with decreased NDVI were primarily scattered across Bases 2, 3, 4, and 5. Finally, 61.59% of the polygons had a ΔNDVI greater than 0, accounting for 375.74 km² of the total area (Fig. 6). When categorized by land use type, using the original Sentinel data, the increase in NDVI was highest in bare land areas after PPPs construction, reaching 0.037. However, only 46.35% of the area in PPPs located in forest regions showed an increase in NDVI, while only a quarter of the area in PPPs on other land use types displayed a similar increase (Supplementary Fig. 7). With model-predicted data, over half of the PPPs across all land use types displayed a similar increase. PPPs in forest areas showed the highest NDVI increase and proportion, reaching 0.058 and 71.54% (Supplementary Fig. 7).

Fig. 5: Spatial distribution of ΔNDVI between photovoltaic power plant areas and the 100–300 m buffer zones in the nine clean energy bases.
figure 5

a Sentinel-2 NDVI data. b NDVI data predicted by deep learning model. c–h Enlarged paired panels showing different photovoltaic power plant configurations.ΔNDVI represents the changes in NDVI between photovoltaic power plant areas and the 100–300 m buffer zone.

Full size image
Fig. 6: Details of the NDVI change in the photovoltaic power plants areas across the nine clean energy bases.
figure 6

a Distribution of photovoltaic power plant areas. b Proportional distribution of photovoltaic power plant patches. c Distribution of ΔNDVI between photovoltaic (PV) and surrounding non-photovoltaic (NPV) areas. NDVI_sen: Sentinel-2 NDVI data. NDVI_sen_pre: NDVI data predicted by deep learning model. Whiskers denote 1.5 times the interquartile range, and the remaining points are outliers.

Full size image

The numerical underestimation is also evident in the turning point projections for the nine clean energy bases when using raw Sentinel data (Fig. 7). Analyses of the raw Sentinel-2 data revealed that the NDVI threshold for non-photovoltaic (NPV) areas in the nine clean energy bases was 0.058 for the construction of PPPs (Fig. 7a). Meanwhile, the ΔNDVI changed slightly at the same NDVI levels of NPV locations. However, when using the model-predicted data, the turning NDVI of NPV increased to 0.155, and the distribution became more scattered (Fig. 7b). When categorized by land use type, the threshold was noticeably higher in forest areas at 0.076, while grasslands had the lowest threshold at 0.054. In addition, croplands, bare land, and impervious surfaces showed higher R² values. Similarly, when using model-predicted results, R² values decreased significantly, with forest areas dropping to 0, and the turning points for each land type increased substantially (Supplementary Fig. 8).

Fig. 7: Details of the liner regression between ΔNDVI and the NDVI of the 100–300 m buffer zone.
figure 7

a Sentinel-2 NDVI data. b NDVI data predicted by deep learning model. ΔNDVI represents the changes in NDVI between photovoltaic power plant areas and the 100–300 m buffer zone.

Full size image

Discussion

The proposed multi-scale method not only enables the matching and calibration of field data with UAV data but also facilitates the alignment and calibration of satellite data across regions. Our results demonstrated high prediction accuracy at various study sites, providing a reliable foundation for enhancing the true NDVI of PPPs using Sentinel-2 data. Our findings reveal that vegetation growth beneath PV panels was only slightly lower than that between panels. More importantly, the development of PPPs enhanced local vegetation growth in over half of the areas examined, with no significant limitations related to land use types or regional characteristics. These results highlight substantial potential to integrate PPP development with ecological restoration. Below we discuss the causes of NDVI changes, the evaluation bias for future improvement, as well as insights for further application of our method to vegetation monitoring of PPPs.

Overall, the NDVI changes produced by our method show a similar value distribution to the original Sentinel-2 data, but the decreased and increased sections undergone a reversed transformation (Supplementary Fig. 7). When applied at a regional scale, our findings partially align with previous research, indicating that the impact of PPPs on vegetation varies with climate and land use type. However, unlike previous reports suggesting that PPPs lead to widespread vegetation loss, our results indicate that more than half of PV installations on various land use type contributed to vegetation restoration. Notably, the increase in NDVI was smaller in bare land and grassland areas, while more substantial gains were observed in forested and cropland regions20,23,34. This may be related to the strict site selection strategy during the construction of PPPs, where vegetation in the construction areas is typically more sparse than that in the surrounding areas23. Even in forest and cropland regions, the baseline NDVI values (around 0.09) are close to those of bare land and grassland (Table 2). However, due to relatively more favorable climatic conditions in forest and cropland areas, vegetation growth after PPPs construction is better than that in bare land and grassland areas. Multi-site monitoring in humid regions show that although PV panels reduced wind speed and photosynthetically active radiation, vegetation could actively adapt (e.g., by reducing water consumption) to maintain basic biomass production. As a result, vegetation beneath PV panels did not significantly differ from that between panels or in external control areas35. In addition, the more dispersed distribution of ΔNDVI indicates that actual vegetation growth is influenced by multiple factors (Fig. 7).

In the assessment of local vegetation changes at PPPs, two factors can affect the accuracy of this process. One is the accuracy of the vector boundaries of PPPs. When extracting boundaries using medium-resolution satellite images, the polygons often include NPV areas at the edges in flat areas, such as deserts and grasslands. In contrast, in mountainous regions with shrubs and forests, PPPs are more scattered and fragmented (Table 2), making it difficult to distinguish adjacent NPV areas and potentially resulting in the omission of small PV arrays1,3,53. Therefore, we used the extraction results from 2 m high-resolution satellite images to define the boundaries of the PPPs. The other is the selection of control areas. Previous studies commonly used a buffer zone of over 500 m away from the PV area as control areas, which may introduce considerable errors21,34,54. Using Landsat 8 data at a 30 m resolution, NDVI and EVI values for 3991 PPPs in China were compared23. Their results show that the values stabilized beyond 90 m, suggesting that the influence of PPPs construction is typically localized. Thus, our study also selected 118 PPPs for comparison using Sentinel data at a 10 m resolution, revealing that ΔNDVI initially increased with distance and then gradually decreased. In other words, control areas that are too far from the PPPs may exaggerate the negative impacts of their construction. As a result, we selected a uniform buffer zone of 100–300 m as the control area (Supplementary Fig. 9). Our findings suggest that conducting classification studies across different land use types and regions may be necessary to define research boundaries and select appropriate buffer zones more accurately.

The spatial and temporal transferability of our method is a key aspect of its applicability. Independent validations revealed that model performance varied across different PPPs. At the pixel level, the corrected outputs significantly reduced the discrepancy between Sentinel-2 NDVI and ground truth, though perfect agreement was not achieved (Supplementary Fig. 10). However, at the plant level, the corrected NDVI closely matched ground truth values (Supplementary Fig. 6). Overall, Sentinel-2 NDVI tended to overestimate values in low-NDVI plants and underestimate them in high-NDVI ones. Furthermore, the model trained on the full dataset (i.e., with internal validity) did not consistently outperform those trained with the six test sites excluded (i.e., in extrapolative settings), although their results remained largely comparable across most PPPs (Supplementary Fig. 10). These findings highlight the potential of more detailed and multi-level intra-site datasets (e.g., panel layouts, vegetation types, and heterogeneous distribution patterns) to enable more refined analyses and further enhance the robustness of our proposed method. Time series analysis of a power plant located in Yueqing, Zhejiang Province demonstrated that all data sources exhibited similar trends over both the entire plant and the PV area throughout the year (Supplementary Fig. 11). It is particularly evident in the final output of our method and the UAV data after interpolating the data under PV panel. However, the data consistency was relatively poor for NPV areas, which may result from the limited number of training samples in the NPV region and lower image quality, particularly during haze-affected periods in February. Existing studies also showed that the high revisit frequency of Sentinel-2 enabled the temporal gap between satellite and ground observations to be limited to within one week. In our comparative analysis, a 3-day versus 7-day temporal gap did not significantly affect the performance of our proposed method (Supplementary Figs. 12 and 13). However, the consistency between UAV data and Sentinel data varies across different species and growth stages55,56,57. These findings underscore the importance of conducting regular field observations across a broader range of sites and time periods to support the robustness and generalizability of our proposed method. Moreover, these datasets enable more granular investigations to improve numerical accuracy by assessing the stepwise impacts of differences in viewing angles and temporal mismatches. In addition, UAV data can capture changes or phenological events that may be overlooked by satellite data40. When sufficient suitable paired data are available, UAV data can serve as an effective complement to long-term monitoring data of PPPs, similar to its application in crop damage assessment58.

However, as a data fusion method, our approach is constrained by certain physical limitations, primarily in two aspects. First, the spectral response functions of different sensor hardware are not entirely consistent (Supplementary Fig. 14). Second, the shooting times are not completely synchronized, meaning the observed subjects may not be exactly the same. Regarding the first aspect, our approach needs to address the differences between two sets of hardware. The first is the difference between the sensor used to photograph the field pot and the onboard sensor of the UAV51,59. To mitigate this, we 3D-printed a customized stand and manually lifted the UAV used for flight data collection, enabling effective vegetation photography while also reducing the cost of regular monitoring (Supplementary Fig. 15). The second set of differences arises from the onboard sensors of the UAV and the satellite, as well as their different operating environments. Fortunately, the UAV and Sentinel sensors exhibit similar spectral coverage and response characteristics of each band (Supplementary Fig. 14), supporting more effective data integration compared to Landsat 838,60. In addition, extensive research in precision agriculture has demonstrated strong NDVI consistency between UAV and Sentinel satellite data57,61,62. Meanwhile, several companies are developing UAV platforms equipped with sensors that match the physical properties of Sentinel and Landsat satellite sensors, aiming to achieve better data consistency in future research63. For the second aspect, since UAV remote sensing often requires multiple shots to cover the target area, internal consistency must also be considered59. The flights must be timed to ensure optimal leaf conditions (dry or fresh), avoid shadows, and maintain a similar solar zenith angle as satellite photography64,65. These impacts are diminishing as UAVs become faster with longer flight durations and the multiple-shot mosaic algorithm continues to improve.

Aside from the impacts of satellite imagery resolution and sensor differences, greater attention is needed for PPPs in densely vegetated areas where the construction patterns are more complex (e.g., varying array sizes, orientations, and irregular configurations). Observations in these areas are more susceptible to local meteorological events such as rainfall, cloud cover, and fog, which can introduce errors. Although our study employed convolutional neural networks to better establish the connection between complex structural targets and various sensor and reference data type66, more frequent and comprehensive observations can provide a broader set of characteristic parameters to further enhance this advantage. With advancements in visual positioning and automated control technology, fully automated drone monitoring has increasingly been implemented in the energy industry. However, current applications primarily focus on using thermal infrared sensors to detect faults in PV panels rather than to monitor vegetation67. Another key aspect of vegetation monitoring is the acquisition of ground data. In fact, within the same PPP, there is considerable heterogeneity in vegetation conditions68, necessitating more sample points or richer sample information to accurately predict the vegetation data beneath the PV panels. Moreover, the impact of PPPs on vegetation is a long-term and ongoing process. Therefore, long-term monitoring of PPPs, as a unique man-made land use type, will help better understand their lasting impacts on vegetation, soil, hydrology, and other ecological components.

Ultimately, our proposed method demonstrates that current assessments of vegetation impacts of PPPs based solely on satellite data can be significantly biased at the regional scale. This bias reinforces the notion that constructing PPPs in the arid regions of western and northwestern China is more advantageous. Consequently, it has partially led to stricter policy constraints on PPP development, resulting in a mismatch of regional electricity resources and an increased burden on transmission infrastructure. Our research suggests that a more balanced spatial layout for PPP development can be achieved by considering the specific impacts on vegetation. In addition, our study reinforces the idea that the “PV +” modes, such as agrophotovoltaic and PV ecological restoration power plants, may represent a win-win strategy for achieving both electricity generation and food provision or vegetation restoration.

Conclusions

Using satellite remote sensing images for vegetation information extraction is a crucial step in land use planning and environmental protection. However, PPPs present a challenge in this regard, as the vegetation information within PPPs is discontinuous and partially obscured. Neglecting these complexities may lead to incorrect conclusions and misleading interpretations. Here, we propose a potential solution that integrates ground data and UAV data to calibrate satellite data. Through this application, we find that current research results overestimate the negative impact of PPPs on vegetation. Our findings indicate that when accounting for vegetation growth under PV panels, the development of PPPs enhances local vegetation growth in more than half of the surveyed areas. Our study further reveals that the true vegetation conditions in PV plants are influenced by multiple factors and could not be reliably predicted solely based on the pre-existing local vegetation conditions. Nevertheless, with the growing application of UAV surveys, our proposed method offers an efficient, accurate, and scalable solution for long-term monitoring of vegetation in PPPs across diverse regions.

Methods

Study area

To develop a reliable method, we selected mainland China as our study area. We collected data to analyze the relationships among field data, UAV data, and satellite data. For method validation, we targeted key PPP development areas in different climatic regions across China as demonstration areas.

During the method development phase, we aimed to cover a wide range of PPPs across diverse climate zones and land use types in China. To support this, we conducted multiple rounds of field investigations (Fig. 8). These included monthly investigations at three sites in Zhejiang Province throughout 2022, when travel within China was still restricted due to COVID-19, along with expanded surveys in July, August, and December to capture variations across different vegetation growth stages, representing PPP in humid regions with well-developed vegetation. After travel restrictions were eased in 2023, we conducted two cross-region investigations to improve temporal and spatial representativeness: one in April and May, representing the growing season, and another in August and September, representing the non-growing season. Based on the number of valid observations in this study, PPP samples were classified into three UAV data groups (single flight, two flights, and ten flights) and two field survey groups (single field trip and two field trips) (Fig. 8, Supplementary Table. 1). All investigated power plants were pre-selected based on publicly available inventories of PPPs1,3.

Fig. 8: Spatial distribution and data acquisition overview for photovoltaic power plant monitoring.
figure 8

a,b Locations of photovoltaic power plants for UAV flight data and field data collection. c,d Representative vegetation conditions. e,f Pairwise handheld UAV data acquisition.

Full size image

During the method application phase, we selected the nine national clean energy bases and the remaining areas of Gansu Province as demonstration areas (Fig. 5, Fig. 8). These nine national clean energy bases are highlighted for PPP development in the 14th Five-Year Plan for National Economic and Social Development of China and the Outline of the 2035 Vision Goals27. We also included the remaining areas of Gansu Province to encompass more core regions where PPPs have been developed, ranging from mountainous areas to grasslands. This selection (hereinafter referred to as the nine clean energy bases) spans 11 provinces in China and covers approximately 2.24 million km2 and provides a more balanced representation of varying vegetation conditions. Overall, Base 1 is dominated by desert bare land, with an average temperature of 9.5 °C and precipitation of 116.9 mm. Bases 2, 3, 6, and 7 are primarily grasslands, with Base 7 experiencing shorter sunshine duration (2370 h) and a higher precipitation (1771.9 mm). Bases 4, 8, and 9 are predominantly forested, with an average NDVI exceeding 0.7. while Base 5 is mainly cropland (Fig. 8).

Data collection

During each field investigation, we simultaneously collected the pairwise handheld orthophoto data that represents vegetation growth within PPPs and the ultra-high-resolution UAV data that covers each PPP and its surrounding area. Then, we obtained corresponding medium-resolution satellite data to align with the UAV data.

During field investigations, we used the DJI Phantom 4 Multispectral quadcopter to capture pairwise handheld orthophotos and the DJI Mavic 3 Multispectral quadcopter to collect ultra-high-resolution aerial remote sensing data. Both quadcopters feature identical image sensors and an incoming sunlight irradiance sensor, although the Mavic 3 provides a longer flight duration. Each quadcopter is equipped with a color sensor for visible light (RGB) imaging and four monochrome sensors for multispectral imaging: namely green (G) 560 ± 16 nm; red (R) 650 ± 16 nm; red edge (RE) 730 ± 16 nm; and near infrared (NIR) 840 ± 26 nm. While capturing images, the top-mounted incoming sunlight irradiance sensor on the quadcopter measures incident light intensity in real time (Supplementary Fig. 15b), enabling preliminary spectral calibration and improving the accuracy of multispectral information.

In each field investigation, UAV flights were conducted on clear days with good visibility and relatively low wind speeds to ensure even coverage of the study area, including a 150 m outer buffer zone. All UAV flights were scheduled to fly at a height of 200 m, with 80% front overlap and 70% side overlap, achieving optimal stitching quality and balance between image coverage and spatial resolution (9.22 cm pixel-1). Moreover, onboard DJI Real-Time Kinematic Positioning module enabled centimeter-level positioning, ensuring precise geolocation of the UAV images. Then, three greyscale reference panels with reflectance values of 25, 50, and 75% were placed during each flight for image calibration (Supplementary Fig. 16). Simultaneously, we conducted approximate stratified random sampling within each PPP, selecting plots based on preliminary field observations and satellite imagery to gather balanced samples across different levels of vegetation saturation. Handheld orthophotos were manually captured in pairs to compare vegetation BP and UP (Fig. 8e, f, and Supplementary Fig. 15b). Due to access restrictions at some PPPs, we ultimately gathered 3295 pairs of BP and UP images from 76 field observations.

After completing the UAV data collection, we used the DJI Terra software to pre-process the data. The processed output included orthomosaics for each individual band. We calculated the NDVI as the normalized difference between the NIR and red bands:

$${NDVI}=\,\frac{{NIR}-{Red}}{{NIR}+{Red}}$$
(1)

For satellite data, we selected the Sentinel-2 series due to its high spatial and temporal resolution, as well as free availability for wide application. We sourced Level-2 Sentinel-2 data from the AI Earth platform69, which provides products corrected for orthorectification, geometry, and atmospheric effects, starting from March 2017. We used four spectral bands, including blue (492.4 ± 66 nm), green (559.8 ± 36 nm), red (664.6 ± 31 nm), and NIR (832.8 ± 106 nm) bands. The acquisition dates were selected to ensure comparability with UAV data (Supplementary Table 1). By filtering out satellite data compromised by adverse weather conditions (primarily cloud cover) or failing to meet the temporal alignment requirement (within one week) of UAV acquisition, we retained a final set of 129 valid image pairs from 82 geographically representative PPPs across China (Fig. 8 and Supplementary Table 1).

Data processing and analysis

To enhance the accuracy of vegetation condition assessment in PPPs, we developed a multistage method to interpolate the NDVI at BP and correct Sentinel-2 data (Fig. 2). Before interpolating the NDVI at BP, we utilized a DeepLabV3+ model with a pretrained ResNet101 backbone to segment PV panel arrays from the background70. We manually labeled PV panels on 36 of 129 UAV orthomosaics (Supplementary Fig. 17), resulting in a total of 16,025 RGB patches of 1024 × 1024 pixels. Of these, 80% were used for training and 20% were reserved as the test set. The training of the DeepLabV3+ model was conducted on the AI Earth platform. Subsequently, the final model was applied to segment PV panel arrays in the remaining plants.

After segmenting the PV panels and the background, we followed the workflow outlined in Supplementary Fig. 18b to interpolate the NDVI at BP using customized Python codes. First, we divided the PV panel array into left and right sections. Second, we utilized a 3 × 3 pixel moving window to calculate the mean NDVI at BP, ensuring greater numerical stability. Thirdly, we employed a linear regression model to simulate the relationship between BP and UP. During this process, we removed outliers from the pairwise handheld data using a studentized residual threshold of 2. To eliminate bias and better capture the relationship across different regions, we used the mean NDVI of the buffer zone as a proxy. We then divided the data into two independent sets and built separate models for regions with better vegetation and those with poorer vegetation. Finally, we applied the regression model to predict the NDVI value of the corresponding pixel at UP.

Visual inspection confirmed that the UAV data acquired using RTK technology were well aligned with the geometrically corrected Level-2 Sentinel imagery. To minimize the impact of resolution differences between the datasets, the interpolated UAV data were resampled to match the spatial resolution of the Sentinel-2 imagery, thereby ensuring improved geospatial matching.

Upon completing the interpolation, we modified a U-Net model to calibrate the Sentinel-2 NDVI to better capture the complex relationship between UAV and satellite data (Supplementary Fig. 19). To minimize human-guided feature engineering and reduce interference with the model’s ability to autonomously discover task-relevant spectral interactions from raw satellite data, while also reducing data complexity and redundancy, we used the red, green, blue, and near-infrared bands of Sentinel-2 as inputs. We utilized the interpolated and resampled UAV NDVI data to the resolution of Sentinel-2 as outputs. Ultimately, due to the coarse resolution of the Sentinel images and the relatively small area of the PV plants, we selected 5054 patches of 64 × 64 pixels as inputs for model training, with 80% used for training and 20% for validation. To further evaluate the model’s generalization capability, we tested it on six PPPs across China (Fig. 8). We then trained the modified U-net model using all available data, which was subsequently used to calibrated the Sentinel-2 NDVI and mosaicked the predicted NDVI for each PPP for further analyses.

Applications of our method

Typically, there are two methods to test whether a constructed PPPs has affected the local vegetation or not and to what extent. For PPPs with a known construction date, a before-after comparison is often used (see Eq. 2). For PPPs where the exact construction date is unknown, a comparison to the buffer zone is utilized (see Eq. 3). We focused on the vector boundaries of the constructed PPPs by the end of 2020, which were more accurately extracted from Gaofen-2 satellite data (2 m resolution). We then used Eq. 3 to calculate ΔNDVI. In addition, to avoid inter-annual bias, we employed the average NDVI data from Sentinel for the three years from 2021 to 2023 in our calculations. To further investigate the impacts of PPP construction on different land use types, we categorized those types based on the dominant land use within the 0–300 m peripheral buffer zone from 2016 to 202071. To determine an appropriate buffer zone, we manually determined the vector boundaries of 118 newly constructed PPPs in 2021 and conducted a comprehensive comparison using Supplementary Equation 1. Ultimately, we identified an area ranging from 100 to 300 m outside the PPPs as the control zone. We then set ΔNDVI values between −0.02 and 0.02 as indicative of no change, allowing us to analyze the impact of PPPs on vegetation using various data sources.

$$\varDelta {NDVI}={{NDVI}}_{{PV}}^{{after}}-{{NDVI}}_{{PV}}^{{before}}$$
(2)
$$\varDelta {NDVI}={{NDVI}}_{{PV}}-{{NDVI}}_{{buffer}}$$
(3)

Another important application is to determine whether new PPPs will impact local vegetation, which involves the identification of NDVI threshold for such impacts. Such assessment helps quickly determine whether a specific area is suitable for constructing PPPs and whether any post-construction artificial maintenance measures for vegetation are necessary. Therefore, we employ a simple linear regression model to predict the ΔNDVI using the NDVI of NPV, indicating the rate of decrease and potential turning point.