DFDR-NLNet: A dual-frequency differentiated representation non-local network for photovoltaic panel segmentation

Abstract

Photovoltaic (PV) technology plays a crucial role in expanding renewable energy globally, yet achieving precise PV panel segmentation to optimize resource allocation and guide installation policy remains a challenge across urban, rural, and industrial environments. To address data diversity limitations, we propose a data augmentation method using a denoising diffusion probabilistic model (DDPM) to generate joint data distributions, enhancing model robustness. Building on this, we introduce a dual-frequency differentiated representation non-local network (DFDR-NLNet) for realistic PV panel segmentation. To enhance the efficiency of global contextual feature extraction in the Transformer branch, we propose a low-frequency representation Transformer that strengthens large-scale semantic modeling through frequency decomposition and preserves crucial positional cues using original phase information. Additionally, a cross-scale alignment module (CSAM) is proposed to facilitate semantic alignment and collaborative learning across different feature levels. To enhance the contribution of edge information in the segmentation process, we design an edge feature awareness module (EFAM) that focuses on high-frequency information. Finally, the correspondence between edge features and decoder representations is modeled to facilitate segmentation in ambiguous regions, via a multi-directional cross attention (MDCA). DFDR-NLNET achieves mIoUs of 83.39 %, 66.14 %, and 91.48 % on PVP-Dataset, BDAPPV, and PV01, outperforming other methods in PV panel localization and edge refinement. Furthermore, the method is used to evaluate the power generation capacity of the Kael Solar Power Plant in Senegal, where the array area is calculated to be 0.25 km², the system size is 38.13 MW, and the annual output power is 63.71 GWh.