Answers are grounded in this paper and cite their sources.
Accurate segmentation of photovoltaic (PV) arrays from remote sensing imagery is essential for renewable energy monitoring, yet existing methods struggle to preserve the rectangular geometry characteristic of solar panel installations. Convolutional neural networks suffer from limited receptive fields that impede long-range dependency modeling, while transformer-based approaches incur prohibitive quadratic computational complexity for high-resolution imagery. This paper presents SR-Mamba, a structure-retentive state-space network that achieves geometry-aligned PV segmentation with linear computational efficiency. The proposed framework introduces three synergistic modules: (1) a Rectangular-Prior Tokenizer (RPT) that embeds edge and shape cues into input tokens via feature-wise linear modulation, providing early geometric bias; (2) a Grid-Aware State-Space Block (GSSB) that performs multi-directional state propagation along horizontal, vertical, and diagonal orientations with geometry-guided gating to prevent feature bleeding across panel boundaries; and (3) a Boundary-Preserving Decoder (BPD) that employs dual-stream decoding with cross-attention fusion to reconstruct sharp, coherent segmentation masks. Comprehensive experiments on the HRPVS and Multi-Resolution PV datasets demonstrate that SR-Mamba achieves state-of-the-art performance, attaining 92.24% IoU and 88.93% boundary-mIoU on HRPVS, and 90.87% IoU on Multi-Resolution dataset, consistently outperforming CNN, transformer, and existing SSM baselines. With only 12.8 M parameters and 16.1 ms inference time, SR-Mamba offers an effective balance between segmentation accuracy and computational efficiency, making it well-suited for large-scale PV mapping applications.
No takes yet. Share an insight, caveat, or question.
Nazir et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: