Methane Gas Emission Detection using Deep Learning and Hyperspectral Imagery

Scientists from leading institutions have warned that global warming is the next potential world ending catastrophe, which will result in increasing natural disasters and social disruptions. Methane gas (CH4) is up to 80 times more effective in trapping heat than carbon dioxide, and it is considered to be the largest contributor to global climate change. Methane gas is invisible, odorless and colorless, which makes it impossible to detect, much less combat. Previous machine learning methods require expert intervention and are often largely inaccurate. Hyperspectral mask-RCNN (H-MCRNN) is a recently developed method among the deep learning-based approaches, which allows for accurate automatic detection due to its incorporation of hyperspectral imagery in conjunction with deep learning neural networks. While powerful, the original H-MRCNN model lacks optimization in terms of the hyperparameters, which directly affects the efficiency and accuracy of the model, resulting in multiple unfeasible detections. This project focuses on the optimization of hyperparameters to reach the full potential of the H-MRCNN model to realize true automatic methane detection without expert intervention. Results have shown that higher epoch numbers and lower training rate can produce clearer and more accurate masks, with a minimum of 24 epochs and a training rate of 1e-6. By using the optimized hyperparameters, 87% success rate can be achieved in the detection of methane leaks.