Spectral Channel Attention Network: A Method for Hyperspectral Semantic Segmentation of Cloud and Shadows (Papers Track)
Manuel Pérez-Carrasco (University of Concepción); Maya Nasr (Environmental Defense Fund, Harvard University); Sébastien Roche (Environmental Defense Fund, Harvard University); Chris Chan Miller (Environmental Defense Fund, Harvard University); Zhan Zhang (Harvard University); Core Francisco Park (Harvard University); Eleanor Walker (Harvard University); Cecilia Garraffo (Center for Astrophysics $|$ Harvard & Smithsonian); Douglas Finkbeiner (Harvard University); Ritesh Gautam (Environmental Defense Fund); Steven Wofsy (Harvard University)
Abstract
Accurate detection of clouds and cloud shadows is essential for reliable atmospheric methane retrievals, a critical component of global climate monitoring efforts. This work presents the Spectral Channel Attention Network (SCAN), a simple deep learning architecture that addresses the fundamental challenge of spectral band selection for hyperspectral cloud and shadow detection through channel-wise attention mechanisms. Unlike traditional approaches that treat all spectral bands equally, SCAN dynamically weights spectral channels based on their discriminative power for atmospheric artifact detection. We evaluate SCAN on MethaneSAT and MethaneAIR hyperspectral datasets, demonstrating superior performance on MethaneSAT (71.53\% F1-score vs U-Net's 68.56\%). Furthermore, we show that SCAN's spectral attention capabilities can be effectively combined with spatial processing through ensemble approaches, achieving state-of-the-art F1-scores of 78.50\% for MethaneAIR and 78.80\% for MethaneSAT. These improvements directly enhance satellite-based methane monitoring reliability, supporting global climate mitigation efforts. Our code will be available upon acceptance.