Automated Gas Identification in Hyperspectral Imagery via 3D-CNN and Autoencoders

Abstract

Gas emission monitoring plays a vital role in environmental safety and public health protection. This work presents an advanced hyperspectral image-based gas detection system utilizing deep learning techniques in the longwave infrared (LWIR) spectrum. Unlike traditional approaches, the proposed method considers radiance as a combination of background and gas signatures, improving detection reliability. The system initially converts radiance data into luminancetemperature representations for better feature interpretation. A hybrid architecture integrating a 3D Convolutional Neural Network with an auto encoder is employed to perform spectral unmixing and extract meaningful abundance and endmember features. These extracted features are further processed using a fully connected neural network for accurate pixel-level gas classification. The model is evaluated on datasets containing gases such as methane and sulphur dioxide. Experimental results demonstrate improved detection performance compared to conventional techniques. An extended model combining CNN, Bidirectional layers, and GRU further enhances accuracy. Comparative analysis confirms the superiority of the proposed framework. The approach eliminates the need for thresholdbased detection and adapts efficiently to multiple gas types. Overall, the system provides a robust and scalable solution for remote gas detection applications.