Edge-enhanced SAM for extracting photovoltaic power plants from remote sensing imagery

Abstract

Accurate geospatial extent data of photovoltaic (PV) power plants is essential for assessing their socioeconomic benefits and environmental impacts. However, existing semantic segmentation models often result in over-smoothed edges and the inaccurate delineation of PV power plants. To address these limitations, we proposed a novel edge-enhanced Segment Anything Model (ESAM) tailored for PV power plant extraction. It is designed as a multi-task network that integrates three key components: semantic segmentation, edge detection, and semantic and edge fusion. The semantic model leverages a modified Segment Anything Model (SAM) foundation model to extract semantic features of PV power plants. An edge module is developed to improve the edge delineation ability of ESAM. Additionally, a learning-based fusion module is designed to combine semantic and edge information to enhance PV power plant identifications. Validation demonstrates that ESAM achieved a high overall accuracy (OA = 98.46 %). Meanwhile, it outperformed three state-of-the-art models by providing higher accuracy metrics and more accurate edges of PV power plants. Thus, the proposed ESAM offers a robust tool for PV power plant extraction.

Keywords

• PV power plants
• Semantic segmentation
• Edge detection
• Learning-based Fusion
• SAM