AstroAI Lunch Talk - July 13, 2026 - Manuel Pérez-Carrasco

Speaker: Manuel Pérez-Carrasco (CfA) Title: Deep Learning for Clouds, Cloud Shadow and Plume Segmentation in Methane Satellite and Airborne Imaging Spectroscopy Abstract: Accurate methane emission quantification from imaging spectroscopy depends on two segmentation problems: screening out clouds and cloud shadows that bias trace gas retrievals, and detecting the methane plumes themselves. In this talk, I present deep learning solutions to both challenges for MethaneSAT and its airborne companion, MethaneAIR. For cloud and shadow screening, I compare conventional classifiers against U-Net and a Spectral Channel Attention Network (SCAN), showing that a CNN-based ensemble of the two achieves the best performance across both sensors while remaining fast enough for operational deployment. For plume detection, I present an instance segmentation framework based on Mask R-CNN that addresses MethaneSAT label scarcity through cross-sensor transfer learning from MethaneAIR, combined with physics-informed postprocessing that yields two operational modes: high-sensitivity screening (recall 0.94) and high-precision source attribution (precision 0.92). Together, these results demonstrate how deep learning, cross-sensor transfer, and physical constraints can be combined into robust, scalable pipelines that enhance the emission quantification capacity of current and next-generation methane monitoring missions.