ONE-SHOT LEARNING FOR METHANE LEAK DETECTION IN REMOTE AREAS USING HYPERSPECTRAL DATA

Methane leaks pose a significant threat to the environment, public health, and safety. The Aliso Canyon gas leak incident in California in 2015, which was the largest methane leak in United State's history, highlighted the need for better leak detection methods. Traditional methods includng sensors can be time-consuming and costly, making it difficult to detect leaks before they become catastrophic. In recent years, remote sensing technologies using hyperspectral imaging have shown promise in detecting methane leaks. In this study, we propose a deep learning approach for methane leak detection using hyperspectral data. We use meta-learning, specifically one-shot learning, to compare unknown signatures with signatures of methane. We evaluated our approach on AVIRIS data of Aliso canyon methane leaks and found that it outperforms traditional methods such as matched filters in terms of both accuracy and speed. Our approach has the potential to significantly improve methane leak detection and reduce the impact of these leaks on the environment and public health.