CSPPNet: A Convolution and State-Space-Based Photovoltaic Panel Extraction Network Using Gaofen-2 High-Resolution Imagery

With the rapid development of photovoltaic (PV) industry, it is crucial to accurately identify PV panels using remote sensing data. However, the existing methods still face problems, such as difficulty in distinguishing PV panels from easily confused ground objects, such as dark buildings, roads, and black plastic, lacking the analysis of the directional characteristics, and inadequate capturing the global dependencies. To address these challenges, a convolution and state-space-based photovoltaic panel extraction network (CSPPNet) is proposed. Specifically, we construct a PV panel index based on visible and near-infrared bands to improve the separability of the PV panels from those easily confused ground objects. Second, considering the unique horizontal characteristics exhibited by the PV panels in remote sensing images, a south-facing orientation prior module is designed to enhance the horizontal features and improve our network in capturing horizontal objects. Finally, the encoder of our network adopts a parallel structure of depthwise separable convolution and state-space module to capture local detailed features and global semantic features of PV panels layer by layer. Furthermore, we build a high-quality dataset [named Photovoltaic Panels in the Eastern and Western Regions of China (PPEWRC)] containing four bands ranging from visible to near-infrared wavelength based on Gaofen-2 satellite images. The experimental results show that our proposed CSPPNet effectively reduces the misjudgment of PV panels, with high completeness and clear edges. The intersection over union, precision, recall, and F1-score can reach 77.37%, 87.42%, 85.11%, and 86.98%, respectively, on the PPEWRC dataset. Ablation experiments validate the effectiveness of each module and fusion process, providing insights into the sustainable utilization of renewable energy.