Satellites Reveal Spatial Heterogeneity in Dryland Photovoltaic Plants' Effects on Vegetation Dynamics

Large‐scale photovoltaic (PV) plants are growing rapidly in drylands because of the rich solar radiation and vast unutilized land. The transformation of landscapes in dryland has threatened local fragile vegetation. Existing studies have investigated the issue by field observations and satellite data, yet the spatial differences in vegetation changes due to dryland PV plants deployment and underlying driving mechanisms remain poorly understood. In this study, Landsat Normalized Difference Vegetation Index data were used to assess the vegetation changes disturbed by PV plants in China's drylands. We further identified environmental factors affecting vegetation changes based on random forest regression model. Our findings reveal the spatial heterogeneity in the impact of PV plants on vegetation dynamics—PV plant deployment promoted the growth of vegetation in the vast majority of arid and hyper‐arid drylands, while it tends to cause vegetation decline in the sub‐humid and semi‐arid drylands. The impact of PV plants on vegetation dynamics depends on local environmental conditions. We found that deploying PV plants in areas with sparse vegetation, low humidity, and long sunshine duration is more likely to promote vegetation restoration. The findings and data maps with highly detailed information can help guide solar energy operators in siting and ecological restoration to enhance techno–ecological synergies in the future.


Introduction
Dryland cover approximately 40% of the earth's land surface, where the annual precipitation is far below the annual potential evapotranspiration (Huang et al., 2017;Reynolds et al., 2007).As water-stressed biomes, dryland ecosystems are regarded to be fragile and easily affected by land use and climate change (Burrell et al., 2020;Ren et al., 2022).Globally, 10%-20% of these lands have experienced severe degradation, with unsustainable land use being the important driver (Newbold et al., 2015;Reynolds et al., 2007).In the pursuit of transitioning society from carbon-intensive fossil fuels to renewable energy, global solar energy generation has grown exponentially in recent decades (De Marco et al., 2014;Hernandez et al., 2019).Taking advantage of the overlapping nature of unused land resources and abundant solar resources, a large number of utility-scale photovoltaic (PV) plants have been installed in drylands (Xia, Li, Chen, et al., 2022).However, extensive landscape modification raises concerns regarding land degradation risk (Kim et al., 2021).It is urgent to gain a more comprehensive insight into the implications of land use changes arising from PV expansion for dryland vegetation.
Many studies have recently investigated the possible impact of PV plants on the local environment throughout their lifespan (Wu et al., 2022).During the PV site preparation process, topsoil and native vegetation were typically removed at the site (Choi et al., 2020).Such thorough changes in the soil environment will destroy the inherent biological soil crusts and facilitate soil erosion (Wu et al., 2014).These impacts are obvious in dryland, as restoration of depleted and denuded landscapes from natural succession processes will take a long time, or it may even be impossible to restore the landscapes (van den Berg and Kellner, 2005;Wu et al., 2014).Despite the negative impacts, others found PV panels can also bring positive benefits.For instance, the physical presence of PV panels could modify microclimatic conditions by adjusting solar radiation, wind speed, water condition and temperature (Suuronen et al., 2017;Tanner et al., 2020).Available field observations indicate that the shading effect of the PV panels reduces soil temperature and elevates soil moisture content at the PV-plant level (Armstrong et al., 2016;Yue et al., 2021).These impacts are considered beneficial in dryland ecosystems, as the panel shade could relieve heat and water stress and thus promote vegetation restoration (Wu et al., 2022).In arid sandy areas and grassland, higher vegetation coverage, biomass, and biodiversity were observed in PV plants than that in the outside (Bai et al., 2022;Liu et al., 2019).These possible ecological benefits of dryland PV plants offer potential techno-ecological synergetic solutions for the restoration of degraded drylands (Hernandez et al., 2019).
In northern China, the government has been encouraging the integration of PV development and vegetation restoration activities to combat desertification (Liu et al., 2020).Remote sensing observations show that Chinese PV desert control projects have achieved positive greening benefits in desert areas (Xia, Li, Zhang, et al., 2022).Based on the significant environmental improvements, the compatibility of PV facilities with other dryland industries, such as cash-crop farming and grazing, is also being actively explored to deliver socioeconomic and environmental co-benefits (Xia et al., 2023).However, the physical characteristics of the site and the local climate have a big influence on the techno-ecological synergistic outcomes (Tanner et al., 2020).In some habitats with coarse-grained soil and low rainfall, the panel's rain shadow effect exacerbates moisture stress and negatively affects vegetation abundance (Tanner et al., 2020).Moreover, unreasonable site preparation practices also caused land degradation in some areas (Xia, Li, Zhang, et al., 2022).Therefore, developers need to project the potential impacts on vegetation in PV site planning to maximize beneficial technical-ecological synergistic outcomes and avoid negative environmental impacts.Previous works were oriented toward the conservation value of lands to evaluate environmental suitability for ground-mounted solar energy development, which lacks a quantitative evaluation of the potential impacts on dryland vegetation (Lovich & Ennen, 2011;Stoms et al., 2013).
To fill the knowledge gap, we selected China's drylands as the study area and used the Normalized Difference Vegetation Index (NDVI) to monitor the vegetation changes due to existing PV plants deployment based on largescale satellite-based observations.With the objective to understand which environmental factors impact vegetation changes due to PV plant deployment, we then built a Random Forest Regression (RFR) model to model the relationship between NDVI changes and environmental factors, accounting for the effects of key climatic variables, topography, human activity, pre-deployment vegetation condition, and soil as covariates.We further included the effect of site preparation before the deployment.Finally, the impact of deploying PV plants in different areas of the drylands on NDVI dynamics was predicted, revealing the potential of PV plants for technoecological synergies in the drylands.

Study Area
We selected all of China's drylands as the study area (Figure 1a).China is the country with the largest new and cumulative installation (Xia, Li, Chen, et al., 2022).China also plans to achieve a total installed generation capacity of more than 1,200 GW of wind and solar power by 2030 (Yu et al., 2023), which includes the establishment of more massive PV plants in drylands.
The dryland map was obtained from the United Nations Environment Programme World Conservation Monitoring Centre (UNEP-WCMC) data set (Sorensen, 2007), where drylands are defined as regions with aridity index (AI; The ratio of annual precipitation to annual potential evapotranspiration) below 0.65.China's drylands cover an area of 3.03 million km 2 and are further classified into four categories according to different aridity classes: sub-humid (0.5 < AI < 0.65), semi-arid (0.2 < AI ≤ 0.5), arid (0.05 < AI ≤ 0.2) and hyper-arid (AI ≤ 0.05).The 2020 China PV plant spatial data provided by Zhang et al. was used to map the distribution of dryland PV plants (Zhang et al., 2022).For each PV plant, we used a buffer zone of 1 km around it as the control area for its Earth's Future 10.1029/2024EF004427 XIA ET AL. ecological impact analysis (Figure 1b).For a PV cluster consisting of multiple PV plants (their distance from each other was smaller than 2 km), we used the buffer zone of 1 km around it as the control area for each PV plant within the clustered area (Figure 1c).

Data
Except for the China PV plant spatial data (Zhang et al., 2022), we also used the utility-scale PV solar energy facility footprints mapped by Kruitwagen et al. (2021) because this data set documents the deployment dates of global PV plants constructed between 2016 and 2018, which can help select the PV plants built in the same period.The land cover type of the PV plant and its control area is derived from the 30 m China land cover data set (CLCD), ranging from 1990 to 2019 (Yang & Huang, 2021).Landsat 8 OLI-derived NDVI product was used to analyze dryland vegetation dynamics.The maximum NDVI in the growing season was used to calculate the annual index.
Data on environmental factors affecting dryland vegetation dynamics are collected from various aspects, including climate, vegetation, human activity, topography, and soil.Climate data are obtained from interpolated data sets with a 1 × 1 km pixel size, based on monitoring data from more than 2,400 weather stations, provided by the Institute of Geographical Sciences and Natural Resources Research (https://www.resdc.cn/DOI/DOI.aspx?The dryland soil information was characterized by 11 relevant variables (Coarse fragments, bulk density, soil organic carbon, etc.) at six standard depths (0-5, 5-15, 15-30 cm, etc.) obtained from the SoilGrids250 m database in a 250 m spatial resolution (Hengl et al., 2017).The Digital Elevation Model (DEM) data was collected from the Shuttle Radar Topography Mission (SRTM) of 30 m spatial resolution for topographic data.The slope (SLP) was exported based on the DEM data.In addition, we also used the 2020 population density data to represent the intensity of human activity, which was obtained from the Gridded Population of the World (GPWv4) data set at 1 km resolution.

Methods
In this study, vegetation changes due to existing PV plant deployment were quantified by comparing the differences in NDVI changes before and after PV plant deployment between the PV plant and control area.Random forest regression model trained from the quantified results and multi-source geographic data was used to predict the impact of deploying PV plants in different areas of the drylands on NDVI dynamics.Additionally, agrivoltaics, water-surface PV and urban PV in drylands are out of the scope of this study because they have a low impact on vegetation or the impact is more dominated by agricultural management than others.We aggregated the land cover data set to the plant scale to determine the main land types of PV plants by taking the mode value of pixels in each PV plant.

Quantifying the Vegetation Changes due to PV Plant Deployment
The NDVI changes between two given periods (before and after PV plant deployment) were used to characterize the vegetation changes after PV plant deployment.To exclude the contribution of climatic factors to vegetation changes, a quantitative indicator of the vegetation changes due to PV plant deployment was established by comparing the relative NDVI changes between PV plants and control areas (Equation 1).
where ΔNDVI PV and ΔNDVI NPV represent the NDVI changes for PV plant and the control area in T1 and T2 periods, respectively.ΔNDVI represent the NDVI changes due to PV plant deployment.When the ΔNDVI exceeds 0, the NDVI within the PV plant increases relative to that of the control area, typically indicating a positive impact of PV deployment on vegetation growth.Whereas if it falls below 0, the NDVI within the PV plant decreases relative to that of the control area, suggesting a negative impact of the PV deployment on vegetation growth.Considering that most PV plants were deployed after 2013 and the PV data set used is mapped for 2020 (Zhang et al., 2022), T1 and T2 periods are chosen for 2013-2020, respectively.
This approach assumes that the NDVI changes in PV plants are affected by both climate change and PV plant deployment, while climate conditions mainly control the NDVI variation in the control areas (Qin et al., 2022).In light of the close distance, the contribution of large-scale climatic conditions to the NDVI changes in the PV plant and control area is assumed to be the same (Qin et al., 2022).Therefore, differentiating the NDVI changes of the PV plant from the control area could effectively eliminate the natural climate variability, and fully attribute the change to PV plant deployment.Meanwhile, we excluded land cover/use mainly influenced by anthropogenic activities (e.g., agricultural practice and irrigation, urbanization) to ensure that NDVI dynamics in the control area were climate-driven.Therefore, for NDVI in the control area, only land cover types that are natural lands (i.e., forest, shrub, grassland, barren) were included for analysis throughout the two periods.

Quantifying the Impact of Site Preparation on NDVI Dynamics
In site preparation before the deployment of some PV plants, the surface vegetation was removed by massive grading, which led to an abrupt drop in NDVI (Figure 2).As the PV plant became operational, the NDVI Earth's Future 10.1029/2024EF004427 XIA ET AL.
gradually recovered and became even higher than that of the control area.This restoration period has been observed to take at least 2 years in some sites (Liu et al., 2019).However, according to desertification theories, if such disturbance exceeds the critical threshold, it will cause land transition to a desertified state.At that point, it will be difficult to restore the vegetation cover to its initial state (D'Odorico et al., 2013).Therefore, site preparation was also included as one of the important factors affecting the impact of PV plants on vegetation changes in the study.We used the lowest annual NDVI (NDVI min ) values during the 3 years before and after PV plant deployment (Figure 2) to capture this disturbance.The impact of site formation on vegetation usually depends on the vegetation condition before PV plant deployment.Therefore, by constructing a linear regression model between NDVI min and the NDVI before PV plant deployment (NDVI 2013 ), we were able to interpolate the NDVI min across whole drylands, which reflects the average impact of current site preparation on NDVI dynamics.

Predicting the Impact of Deploying PV Plants on NDVI Dynamics Across the Dryland
We opted for the RFR on Google Earth Engine (GEE) platform to model the relationship between ΔNDVI and environmental factors.As an ensemble machine learning technique, the Random Forest (RF) algorithm uses decision tree and bagging methods to solve regression problems.It can process high-dimensional data and typically delivers superior prediction accuracy compared to other methods like maximum likelihood and singlelayer neural networks (Belgiu & Drăguţ, 2016;Li et al., 2023).In addition, it can rank the importance of input variables, compensating for the shortcomings of traditional methods (Meng et al., 2021).Further details regarding the calculation method can be found in the study by Genuer et al. (2010).The ΔNDVI was used as the dependent variable in the model.The PV plants used in the model are selected from the batch constructed in 2016 to 2018.Using PV plants with similar construction times offers a twofold advantage: Firstly, it minimizes the impact of inconsistent operating years on the prediction of ΔNDVI; Secondly, the evaluation results of these earlier constructed PV plants can reflect their stable ecological impacts on vegetation after a few years.Using ΔNDVI of these PV plants as input data can build a more stable model than using all the PV plants.
The set of explanatory variables includes information on meteorological conditions, topography, soil properties, the area of the PV plant, vegetation characteristics, the impact of site preparation (NDVI min ), and anthropogenic disturbance of each PV plant.In particular, the NDVI before PV plant deployment (NDVI 2013 ) was used to characterize the vegetation at the PV plant site, and population density was used to measure the disturbance intensity of natural vegetation by local human activities.Explanatory variables were ranked using RF importance scores, and variables with minor importance were removed.For soil properties, we removed most of the variables Earth's Future 10.1029/2024EF004427 XIA ET AL.
that did not improve the model's accuracy and retained only bulk density (BDOD) and total nitrogen (TN) for 30-60 cm.Finally, a total of 15 explanatory variables were used (Table S1).We specified the number of trees for the RFR model as 400 and left the remaining parameters at GEE's default.To evaluate model accuracy, we randomly selected 80% of samples as training (n = 586), and the remaining 20% was reserved for independent tests (n = 147).The root-mean-square-error (RMSE), mean absolute error (MAE), and R 2 were calculated as evaluation metrics using the validation data set.Finally, we applied the model to the whole dryland to predict the impact of deploying PV plants in different areas of the drylands on vegetation dynamics.
Moreover, we created partial dependence plots for important explanatory variables by predicting the ΔNDVI while keeping all explanatory variables at their average values except for the explanatory variable of interest (Bosmans et al., 2022).The average values of explanatory variables are derived from statistics for the entire region (Table S1).

The Distribution Characteristics and ΔNDVI of Existing PV Plants in Dryland
By 2020, the total area of PV plants in China's drylands is 1,832 km 2 .Of all the dryland subtypes, semi-arid drylands boast the most extensive PV development regarding the total area, covering a vast expanse of 706 km 2 and accounting for 39% of the total area of PV plants in drylands (Figure 3a).Sub-humid drylands come in second with 622 km 2 , followed by arid drylands with 403 km 2 , and hyper-arid drylands with the smallest area of 101 km 2 .Moreover, the average size of PV plants is larger in areas with higher aridity.Based on the 2013 land cover data before the deployment of PV plants, the main land use types of PV plants were analyzed.Throughout the drylands, the main land cover types of PV plants are mainly cultivated land, grassland, and bare land (Figure 3b).In sub-humid drylands, the land cover types of PV plants are mainly cultivated land and grassland, other types of land cover include water bodies, woodland and impervious surfaces has less PV deployment.However, the land cover types for other drylands are mainly bare land and grassland.Especially in hyper-arid areas, the land cover type of PV plants is mostly bare land.
When the impact of PV plants on vegetation dynamics was assessed, only PV plants located in natural lands were selected.The total area of the assessed PV plants is 1,441 km 2 , accounting for 79% of the total area of dryland PV plants.Figure 3c shows the spatial distribution of ΔNDVI of existing PV plants.Of these, 53% of the PV plants exhibit a ΔNDVI greater than 0, while ΔNDVI less than 0 were observed in 47% of the PV plants (Figure 3d).In sub-humid drylands, the deployment of most PV plants leads to a reduction in NDVI, with 69% of PV plants exhibiting a ΔNDVI less than 0. Conversely, the deployment of over half of the PV plants drives an increase in NDVI in other dryland subtypes, with ΔNDVI greater than 0 observed in 55%-63% of the PV plants.

Spatial Distribution of the Predicted Impact of Deploying PV Plants on NDVI Dynamics
The validation results show that the RFR model has high predictive capabilities, which predicted ΔNDVI with an RMSE of 0.05, an R 2 of 0.7 and an MAE of 0.03 on the validation data set (Figure 4b).Among all the explanatory variables, the top three factors with the highest importance scores were the NDVI before PV plant deployment, RH, and SD (Figure 4c).The NDVI before PV plant deployment is considered to be the most significant factor affecting the accuracy of model predictions, with a much higher importance score than other explanatory variables.According to the trained RFR model, we mapped the ΔNDVI for the whole drylands (Figure 4a).We counted the area of regions where PV plant deployment may result in either an increase or decrease in NDVI, considering only the natural lands in drylands with a total area of 2.85 million km 2 .
The PV plant deployment is expected to promote an increase in NDVI in 64% of the whole drylands, while the remaining regions will experience a decrease in NDVI.The impact of PV plants on vegetation dynamics exhibits spatial heterogeneity.Specifically, in sub-humid and semi-arid drylands, the PV plant deployment is expected to lead to a decrease in NDVI in the majority of areas.In sub-humid drylands, as much as 94% of the areas are anticipated to experience a decline in NDVI due to PV plant deployment.Conversely, the PV plant deployment is expected to promote an increase in NDVI in the majority of areas (≥90%), bringing ecological benefits to the local vegetation in arid and hyper-arid drylands.Significant ecological benefits are mainly observed in the central part of northern China (Figure 4a).
In addition, the fitting result between NDVI 2013 and NDVI min (R 2 = 0.86) (Figure 4e) confirmed that site preparation tends to reduce the NDVI (i.e., reflected on NDVI min ) within the deployed area in the short term, accounting for an approximate 20% reduction.

Determinants of the Impact of PV Plants on NDVI Dynamics
To investigate how different environmental factors determine the spatial distribution of impacts of PV plants on NDVI dynamics, we further plotted the dependence of the determinants with ΔNDVI (Figure 5).In drylands, the ΔNDVI declined with an increase in NDVI 2013 , RH, PR, and SLP, but increased with a rise in SD, POP, TN, and GHI.In arid and hyper-arid drylands, ΔNDVI increased with WS and EVP, while it decreased with BDOD.Conversely, in semi-arid and sub-humid drylands, an opposite trend was observed.Furthermore, the ΔNDVI of semi-humid drylands considerably reduced with increasing DEM.With increasing LT, ΔNDVI increased in arid and semi-arid drylands.In addition, ΔNDVI increases and then decreases with increasing TEMP in all areas except sub-humid drylands.In summary, the ΔNDVI of PV plants in regions characterized by sparse vegetation, low humidity, and long sunshine durations tends to be higher.This implies that deploying PV plants in such areas is more likely to yield significant ecological benefits.

Comprehensive Land Suitability of PV Plant Deployment
The RFR model results show that over 1.81 million km 2 of drylands are suitable for PV plant deployment based on anticipated impacts on vegetation dynamics, where a positive NDVI effect (ΔNDVI > 0) is observed.Nevertheless, the siting decisions are influenced by various land constraints, and the model can be utilized in combination with other constraint and opportunity models to enable developers to make well-informed and comprehensive siting decisions.Here, we overlapped the RFR model results with the distribution map of PV development potential index (DPI) (1 km resolution) developed by Oakleaf et al. (2019) to generate spatially explicit land suitability maps under the dual consideration of ecological suitability and development potential (Figure 6).The DPI map takes into account both development feasibility and resource potential, and removed Earth's Future 10.1029/2024EF004427 XIA ET AL.
areas with constraints based on the most commonly cited criteria and limitations for development (Oakleaf et al., 2019).
It is found that in areas with high development potential, deploying PV plants often brings positive ecological impacts on vegetation.For instance, in areas with DPI exceeding 0.7, covering an area of approximately 380,000 km 2 in arid areas of China, as much as 88% of these areas exhibit an ΔNDVI greater than 0. This suggests that there may be positive synergies between PV development potential and ecological impacts in drylands.
Deploying PV plants in areas with high PV development potential can achieve higher economic gains and ecological benefits.Such synergistic effect also strengthens and augments the beneficial ecological outcomes produced by PV plants.

The Role of Different Site Preparation Practices in the Impact of PV Plants on Vegetation Dynamics
In model predictions, the impact of site preparation on vegetation is assumed to be at the current average level.However, in reality, the level of impact from site preparation is uneven, depending on the different ground management practices.During PV construction, extreme ground management practices are considered to threaten local vegetation communities and delay vegetation recovery after the construction (Kim et al., 2021).To quantify  S1), except for the determinant of interest.Spatial random effects were not considered while making predictions.
the role of different ground management practices in the impact of PV plants on vegetation dynamics, we conducted a more in-depth scenario analysis.
Three scenarios were proposed for evaluating the dryland vegetation changes caused by site preparation practices.The first scenario represents the average impact of current site management conditions on vegetation (baseline scenario).The second scenario assumes minimal impact of site preparation practices on vegetation (ecological protection scenario).For instance, when the PV panels are installed, the PV posts are drilled into the ground to minimize disturbance to the surface vegetation (Graham et al., 2021).After the installation was completed, the site was prepared for restoration with stockpiled topsoil and native plants (Ott et al., 2021).In contrast, the third scenario assumes that site preparation practices would have a significant negative impact on vegetation (development priority scenario).In this scenario, the construction of PV facilities leads to significant vegetation degradation, which may result from excessive site grading, heavy use of herbicides, and the movement of construction machinery and transport vehicles.After construction, the site space reseeded is rarely for ecological conservation.
Based on assumptions of two extreme scenarios (ecological protection and development priority scenarios), the difference between two scenarios hinges on whether there is a severe or minor decline in short-term NDVI in lands with the same pre-deployment NDVI level.We used a quantile regression method to estimate the magnitude of NDVI decline induced by the two extreme scenarios in the short term (Hao et al., 2019).We categorized PV plant samples into 73 groups based on the pre-deployment NDVI (NDVI 2013 ), with each group consisting of 10 samples.Each group represents PV plant samples with the same pre-deployment NDVI level.We selected the 10th-90th percentiles of the NDVI min sequence for each group of PV plant samples as representatives of severe and minor declines in short-term NDVI, respectively.The selected PV plant samples correspond to those with severe ecological disruption or effective ecological protection during construction.The two sets of data obtained are then re-fitted to derive new regression relationships for NDVI min and NDVI 2013 , reflecting short-term NDVI changes in various regions under extreme scenarios (Figure 7a).The simulated results were reintegrated into the previously trained model as explanatory variables to predict the influence of severe/minor short-term NDVI changes induced by extreme scenarios on future vegetation dynamics (ΔNDVI).The spatial distribution of the predicted impact of PV plant deployment on NDVI dynamics under different scenarios was remapped by updating the explanatory variable.To compare the differences, the area distribution curve of ΔNDVI for the whole drylands under different scenarios was calculated (Figure 7b).For sites with negative impacts from future PV development (ΔNDVI < 0), the negative ecological impact is worsened in the development priority scenario, which is shown in the figure as a leftward shift of the area distribution curve.In the three scenarios, the difference in the areas affected by negative impacts is less than 1%.However, the areas with more severe negative impacts (ΔNDVI < 0.05) are 16% larger than the baseline scenario.Conversely, in the ecological conservation scenario, the areas with more severe negative are reduced by 12% compared to the baseline scenario.This implies that conducting development-oriented site preparation in ecologically vulnerable areas susceptible to negative impacts in future PV massive expansion will lead to further exacerbation of ecological damage.Active ecological conservation practices during construction can help mitigate this negative impact and reduce land degradation risk.

Driving Mechanisms for Techno-Ecological Synergies of Dryland PV Plants
The findings show that in most drylands of China, many PV plants contribute to the healthy development of local ecosystems and desertification control by promoting increasing vegetation cover.Previous research by Xu et al. (2024) also highlighted the positive vegetation impact of large-scale PV facilities in the arid area of China, using Moderate Resolution Imaging Spectroradiometer Enhanced Vegetation Index (EVI) products in a 250 m spatial resolution (Xu et al., 2024).Expanding on this finding, we further elucidated the spatial variations in the impact of PV plants on vegetation and the driving factors at a finer spatial scale using higher resolution remote sensing imagery and machine learning models.The ecological benefits brought by PV plant are particularly pronounced in regions characterized by sparse vegetation, low humidity, and extended periods of sunlight.This finding is consistent with the main inferences of previous studies that PV plants can provide better moisture conditions for dryland plant growth through shading and humidifying effects (Liu et al., 2019;Yue et al., 2021).In densely vegetated areas, site leveling before constructing PV plants and the ongoing vegetation management practices during their operation (e.g., weeding) can lead to a reduction in vegetation, thereby inducing negative ecological consequences (Xia, Li, Chen, et al., 2022).
It was also found that PV plants offer greater ecological benefits in densely populated areas.This can be attributed to the regulatory and exclusionary fencing around PV plants, which limits the adverse impact of human disturbance on the vegetation (selective logging, free-range livestock grazing, etc.), although this may impede wildlife migration.The environmental factors have different effects on the ecological benefits of PV plants in regions of the drylands, possibly due to differences in the main driving forces that affect vegetation.For instance, in arid and hyper-arid drylands, wind erosion is considered the most important driver of land degradation (Li et al., 2021).Areas with high wind speed and evaporation have shown better ecological benefits due to the wind resistance effect of PV panels that reduce soil moisture loss.Conversely, in semi-arid drylands, PV plants are more beneficial in areas with higher surface temperatures, where the cooling effect of PV panels reduces the hightemperature stress caused by excessive high soil temperatures (Yue et al., 2021).This knowledge contributes to advancing the management of PV plants to further harness techno-ecological synergies.For instance, deploying sand barriers within PV facilities in arid and hyper-arid drylands to further reduce wind speed and promote ecological restoration.
By deploying PV plant to drive ecological restoration, the provision of ecosystem services (soil stability and carbon sequestration, etc.) can be enhanced without additional water and fertilizer inputs (Hernandez et al., 2014).
In turn, the transpirational cooling from the vegetation reduces the temperature on the underside of the panels, leading to an increase in PV efficiency (Barron-Gafford et al., 2019).However, plant growth can block solar radiation to reduce power generation and increase fire hazards.Hence, site maintenance, including seasonal mowing and herbicide application is required to mitigate these issues.Although clearing vegetation incurs labor costs, these costs can be offset by establishing agrivoltaics, which can balance PV plant operation and ecosystem management (Luo et al., 2023).

Limitations and Potential Improvements
Satellite observation hinders the ability to observe vegetation below PV panels, leading to potential underestimation of NDVI in PV plants.As a result, the actual ecological benefits of PV plants in drylands may exceed our estimates.More field investigations are needed to supplement knowledge of the growth state and drivers of vegetation in areas which shaded by PV panels.
In this work, relative changes in NDVI were used to quantify the ecological impact of PV plants on vegetation, but the assumption has some limitations.Besides altering the vegetation abundance, the microclimates resulting from the shading effect of PV panels may also affect the plant physiology and morphology, such as the abundance and timing of floral blooms (Graham et al., 2021).In addition, site preparation practices may adversely affect native plant species and facilitate the invasion of exotic species because of soil disturbances (Grodsky & Hernandez, 2020).Alterations in the land cover resulting from the PV facilities construction can also drive habitat loss and fragmentation, having cascading effects on the movement and migration patterns of wildlife (Hernandez et al., 2015).Considering that these factors could potentially alter local ecosystem services (biodiversity, pollination services, etc.), it is also imperative to incorporate them into the PV siting decisions.
Additionally, site-level random effects driven by other uncaptured factors can also lead to some uncertainties in our model prediction results.Other factors, such as the operating years of PV plants, the mounting methods below PV panels and vegetation management practices are all known to affect the ecological impact of PV plants on vegetation, and almost certainly contribute to the variation observed among different sites.These possible explanatory factors need further exploration due to the lack of site-level data.Future additions to this data will allow for a more accurate assessment of the impact of PV plants on vegetation dynamics and help translate the results into specific recommendations for restoration practice.For example, choosing the right design of PV panels (e.g., rotating or stationary) to minimize the negative impact on vegetation based on the site conditions, or specifying a reasonable seeding method and seeded species to drive the restoration seeding within the PV plants.
Finally, the installation of large-scale PV panels also results in changes to albedo, which can affect regional climate and subsequently alter vegetation dynamics (Li et al., 2018).The albedo impact depends on both the background albedo and the PV panels efficiency (Xu et al., 2024).Future related projections and corresponding site plans should incorporate these factors.

Conclusions
As solar projects have been operating at a large scale in drylands, the impact of PV plants on dryland vegetation dynamics remains unclear.To address this issue, the study quantitatively assessed the impacts of PV plant deployment on vegetation dynamics through satellite data.By building a spatial quantitative analysis model, we analyzed the driving mechanisms behind these impacts and predicted spatial distribution of impact of deploying PV plants on dryland vegetation changes.Pre-deployment vegetation condition, relative humidity, and sunshine Earth's Future 10.1029/2024EF004427 duration are the most important determinants.In arid and hyper-arid drylands, PV plants exhibit predominantly positive impacts on vegetation dynamics, contributing to techno-ecological synergies and promoting environmental restoration.Conversely, in sub-humid and semi-arid drylands, the effects of PV plants on vegetation dynamics are mainly negative.Deploying PV plants in these areas requires a careful balance between energy development and ecological protection.The negative impact can be mitigated by prioritizing ecological protection measures during the site preparation phase.The model results demonstrate that 64% of the dryland natural lands have the potential to achieve ecological benefits from future PV deployment, indicating that PV deployment is a win-win solution for energy and the environment in most drylands of China.Through proper site planning, future solar energy adoption can be encouraged to bring positive technical-ecological synergistic outcomes.

Figure 1 .
Figure1.The distribution of China's drylands and the sectorial diagram illustrates the proportions of area across the four aridity levels classes (a), land cover changes before and after PV plant deployment for two example sites, with a 1 km wide buffer zone around the PV plant after its construction defined as the control area (b, c).

Figure 2 .
Figure 2. NDVI dynamics of control areas and PV plants in example region from 2013 to 2020 (a).Google high-resolution images of the example region depict vegetation changes in different periods (b): before the PV plant deployment (2013), the site underwent leveling and experienced a gradual process of re-vegetation over the following years (2016-2019).

Figure 3 .
Figure 3.The total and average area of PV plants in different dryland subtypes (a).Land cover types of PV plants in different dryland subtypes (b).Spatial distribution of ΔNDVI of existing PV plants (c).The average value and area distribution of ΔNDVI of PV plants in different dryland subtypes (d).

Figure 4 .
Figure 4. Spatial distribution of the predicted impact of deploying PV plants on Normalized Difference Vegetation Index dynamics across the drylands (a).The average value and area distribution of ΔNDVI in different dryland subtypes (b).Model accuracy and validation results (c).Importance score of explanatory variables (d).The regression linear outcomes for NDVI 2013 and NDVI min (e).

Figure 5 .
Figure 5. Partial dependency plots of determinants.Partial dependence was calculated by predicting ΔNDVI using the developed model while holding all determinants constant at their mean values (more details in TableS1), except for the determinant of interest.Spatial random effects were not considered while making predictions.

Figure 6 .
Figure 6.The land suitability map produced by the intersection of the development potential index (DPI) map and Random Forest Regression model results.Yellow and red demonstrate the negative impact of PV plant deployment on vegetation dynamics, while yellow stands for low development potential and red for high development potential.Green and purple colors indicate positive impacts of PV plant deployment on vegetation dynamics.Green stands for low development potential, and purple for high development potential.The inset displays the area distribution at a given DPI, with red color representing areas where PV deployment has a negative impact on vegetation dynamics and green color representing areas where PV deployment has a positive impact on vegetation dynamics.

Figure 7 .
Figure 7.The scatter plot of NDVI 2013 and NDVI min (a): the NDVI 2013 values are grouped into 73 sets of 10 values each, with blue and red points representing the 90th-10th percentiles of the NDVI min sequence for each set, respectively.The area distribution of ΔNDVI in whole drylands in three scenarios (b).