Machine Learning and Genetic Algorithms: An Intricate Relationship for Locating Methane in Satellite Images

Detecting methane in remotely-sensed hyperspectral images is a crucial task in environmental monitoring. It aids in identifying methane leaks and emissions, which is essential for mitigating their impact on the climate change. We tackle this challenge and propose an end-to-end and flexible approach for detecting and locating methane plumes in satellite hyperspectral images. It couples classic machine learning with a genetic algorithm to identify the most important image features that contribute to methane localization. We extract features from superpixels—locally-coherent image areas—and then perform feature selection to identify the most discriminative ones, pruning unnecessary (or even noisy) features. Our experiments performed over the large-scale methane dataset (STAR-COP) indicated that feature subsets optimized using a genetic algorithm allow us to build effective methane detectors that outperform supervised models trained over full feature sets. For reproducibility, we refer to our GitHub repository: https://github.com/smile-research/GeneticAlgorithms4MethaneDetection.git.