This paper presents a novel approach that leverages satellite-based machine learning (ML) to detect methane, a major driver of climate change. We train an ML model using "unortho-corrected" data, which bypasses traditional image processing steps (geometric distortion correction, matched filters, etc.), and demonstrate that it achieves comparable performance to models trained on ortho-corrected data. Furthermore, we demonstrate that models trained on ortho-corrected data can outperform existing methods based on matched filters. We release two ML-ready datasets, consisting of ortho-corrected and unortho-corrected hyperspectral images from the EMIT sensor, along with model checkpoints.
