Extracting Photovoltaic Panels From Heterogeneous Remote Sensing Images With Spatial and Spectral Differences

Abstract

The accurate extraction of the installation area of the photovoltaic power station is an important basis for the management of the photovoltaic power generation system. Deep learning has proven to be a powerful tool for rapidly detecting the distribution of photovoltaic panels in remote sensing images. The wealth of information from various remote sensing sensors aids in distinguishing photovoltaic pixels within complex backgrounds. However, the distinct imaging characteristics of different sensors present challenges for deep learning models. In this paper, we propose a deep learning extraction method for photovoltaic panels that effectively improves the spatial and spectral differences inherent in remote sensing images. Considering the characteristics of different sensors, two attention modules and a feature fusion module are applied to suppress the inconsistency of spatial resolution and spectral resolution. Based on the Unet model, we implement the photovoltaic power station identification method and compare it with several commonly used semantic segmentation models. Qualitative and quantitative accuracy assessments show that PV-Unet method can effectively overcome the spatial and spectral differences of remote sensing images. It achieves 98.04% F1 score and 96.15% IoU on the test dataset, verifying the superiority of this method. PV-UNet method is potential for identifying photovoltaic panels from multi-source remote sensing data.