Hyperspectral Image Segmentation of Methane Plumes Using Large Machine Learning Model

Abstract This paper presents a hybrid model for methane plume detection, merging traditional filters with a large computer vision model for efficient monitoring of large-scale emissions. It swiftly identifies methane emission plumes, overcoming false alarms and manual inspection issues. Compared to the previous methods, results show that the new method enhances scalable methane detection while minimizing data acquisition and training efforts. The new approach integrates a swift-matched filter and a robust computer vision model. The initial matched filter outcomes, derived from image processing, are utilized as customized prompts for the Segment Anything Model (SAM), a potent segmentation system. SAM, renowned for its zero-shot generalization capabilities, efficiently identifies methane plume objects and images without requiring additional training. This eliminates the necessity for extra training efforts, streamlining the process of scalable methane emission detection for enhanced efficiency in monitoring large-scale emissions through hyperspectral imagery. The proposed hybrid model undergoes comparison with the baseline matched-filter model published previously. Experimental studies are conducted using the NASA JPL Benchmark for Methane Source Detection from Imaging Spectrometer Data, involving 54 spectrometer images taken from a 3km high altitude with 161 CH4 plume sites. The new model's accuracy and detection of false alarms are assessed against matched filters. A significant discovery is the large machine learning model's ability to zero-shot generalize for segmenting methane plume objects with minimal false alarms, requiring no additional training on methane data. This is crucial for automating large-scale processing, avoiding manual intervention necessitated by previous high false alarm rates. Additionally, the absence of training requirements is noteworthy due to the computational cost and data imbalance issues associated with traditional anomaly detection machine learning procedures, where methane data is a small fraction amid predominantly background data. The novelty of the new model lies in its ability to solve methane plume segmentation efficiently with high performance, achieved through prompt engineering and the utilization of a large machine-learning model.