PV Segmenter: A frequency-guided edge-aware network for distributed photovoltaic segmentation in remote sensing imagery

Abstract

Accurate localization and sizing of distributed photovoltaic (PV) systems using remote sensing imagery are critical for assessing installed capacity and forecasting solar generation potential. However, existing PV extraction methods predominantly rely on spatial-domain learning strategies, which struggle to capture the complex boundaries and fine details of small-scale PV systems. In this paper, we propose PV Segmenter, a frequency-guided edge-aware network that employs frequency-domain learning to improve edge detection and pattern recognition in distributed PV systems. Specifically, a frequency-enhanced edge detection module is designed to leverage frequency-domain decoupling for the extraction of edge semantics related to PV boundaries. An edge-guided feature discrimination module subsequently injects edge cues into multi-level semantic features to refine structural semantic representation. Furthermore, a context-aware cross-layer fusion module is designed to preserve critical details of small PV panels, facilitating robust edge detection. Finally, we introduce an object-edge hybrid loss function with deep supervision that jointly optimizes PV object and edge features. Experimental results on two distributed PV datasets demonstrate that PV Segmenter improves the Intersection over Union (IoU) by 1.96 % to 9.61 % compared to nine benchmark methods. The proposed method shows promise for accurately identifying small-scale PV systems and effectively defining complex boundaries, offering a viable solution for renewable energy assessment and smart grid planning.