Reproduce “SolarFormer: Multi-scale Transformer for Solar PV Profiling” on OpenResearch

About this paper

Adrian de Luis, Minh Tran, Taisei Hanyu, Anh Tran, Liao Haitao, Roy McCann, Alan Mantooth, Ying Huang and 1 more

SolarFormer introduces a multi-scale Transformer-based deep learning model for semantic segmentation of solar photovoltaic (PV) installations from aerial imagery, specifically designed to address unique complexities of PV images. The model achieved up to 4.39% higher IoU and 3.5% higher accuracy on the USGS California dataset compared to existing state-of-the-art methods.

Key insights

• A dedicated multi-scale Transformer architecture, specifically tailored for aerial imagery, can effectively address the unique challenges of solar PV segmentation, outperforming general-purpose deep learning models adapted for the task.
• Integrating multi-scale feature processing with a masked-attention Transformer decoder allows for robust handling of varying object scales, high intra-class heterogeneity, and accurate differentiation from inter-class homogeneous objects.
• Combining both high-level semantic query embeddings and detailed low-level per-pixel embeddings within the decoder enables precise localization and segmentation of even small and densely packed solar panels.

Summary by alphaXiv.