What question did this study set out to answer?

January 16, 2026Open Access

Mapping the Spatial Distribution of Photovoltaic Power Plants in Northwest China Using Remote Sensing and Machine Learning

Key Points

The research aims to create an accurate spatial inventory of photovoltaic power plants in Northwest China using advanced remote sensing and machine learning techniques.
Developed a solar facility inventory framework on the Google Earth Engine platform.
Integrated Sentinel-1 and Sentinel-2 data with interpretable machine learning techniques.
Applied a Random Forest classifier for creating a distribution map with 10 m resolution.
Utilized SHAP to elucidate model behavior and identify key predictors of PV deployment.
Incorporated MaxEnt for assessing spatial drivers of PV installations.
Achieved an overall accuracy of 98.4% and a Kappa coefficient of 0.96 for the model.
Mapped approximately 2560 km2 of PV installations in Northern Ningxia, central Shaanxi, and parts of Xinjiang and Gansu.
Identified the Enhanced Photovoltaic Index and Normalized Difference Built-up Index as strong predictors for PV classification.
Revealed that land-use constraints and grid accessibility are dominant drivers of PV clustering, surpassing solar irradiance effects.

Abstract

Photovoltaic (PV) power generation is essential for achieving carbon neutrality and advancing renewable energy development. In Northwest China, the rapid expansion of PV installations requires accurate and timely spatial data to support effective monitoring and planning. Addressing the limitations of existing datasets in spatiotemporal resolution and driver analysis, this study develops a scalable solar facility inventory framework on the Google Earth Engine (GEE) platform. The framework integrates Sentinel-1 SAR, Sentinel-2 multispectral imagery, and interpretable machine learning. Feature redundancy is first assessed using correlation-based metrics, after which a Random Forest classifier is applied to generate a 10 m resolution distribution map of utility-scale photovoltaic power plants as of December 2023. To elucidate model behavior, SHAP (SHapley Additive exPlanations) is used to identify key predictors, and MaxEnt is incorporated to provide a preliminary quantitative assessment of spatial drivers of PV deployment. The RFECV-optimized model, retaining 44 key features, achieves an overall accuracy of 98.4% and a Kappa coefficient of 0.96. The study region contains approximately 2560 km2 of PV installations, with pronounced clusters in northern Ningxia, central Shaanxi, and parts of Xinjiang and Gansu. SHAP analysis highlights the Enhanced Photovoltaic Index (EPVI), the Normalized Difference Built-up Index (NDBI), Sentinel-2 Band 8A, and related texture metrics as primary contributors to model predictions. High EPVI, NDBI, and Sentinel-2 Band 8A values contribute positively to PV classification, whereas vegetation-related indices (e.g., NDVI) exhibit predominantly negative contributions; these results indicate that PV mapping relies on the integrated discrimination of multiple spectral and texture features rather than on a single dominant variable. MaxEnt results indicate that grid accessibility and land-use constraints (e.g., nighttime light intensity reflecting human activity) are dominant drivers of PV clustering, often exerting more influence than solar irradiance alone. This framework provides robust technical support for PV monitoring and offers high-resolution spatial distribution data and driver insights to inform sustainable energy management and regional renewable-energy planning.

Ask AI

Helpful

Bookmark

View Full Paper