Uncovering the location of photovoltaic power plants using heterogeneous remote sensing imagery

Abstract

Accurate monitoring of photovoltaic (PV) spatial distribution using remote sensing imagery is critical for understanding energy production dynamics. The integration of spatial and spectral features facilitates precise identification of diverse PV installation scenarios. However, existing methods primarily depend on single-source multispectral or high-resolution imagery, limiting their ability to balance spatial detail and spectral richness. To address this, this paper proposes a spatial-spectral differential semantic fusion network named FusionPV to comprehensively map PV locations within complex geographical environments. First, a spatial-spectral differential semantic aware module (SDAM) is proposed to extract spatial and spectral features related to PV discrimination from multimodal images. Subsequently, a dual-domain adaptive cross-fusion module (DAFM) is designed to deeply aggregate and cross-focus multimodal information using a cross-attention mechanism. Furthermore, a local-global semantic aggregation module (LGAM) is introduced to construct global descriptors by locally encoding and aggregating images, thereby enhancing contextual comprehension of intricate scenes. We construct a multimodal PV dataset by integrating GF-2 and Sentinel-2 imagery, focusing on Hubei Province, China. Experimental results demonstrate that FusionPV outperforms five state-of-the-art methods, achieving Kappa coefficient improvements ranging from 3.78 % to 7.23 %. Additionally, a comparison with four existing PV products indicates that FusionPV is a superior solution for acquiring a high-quality, extensive database of PV locations.

Keywords

• Photovoltaic
• Remote sensing
• Deep learning
• Multimodal fusion
• Hubei province