3D-CNN and Autoencoder-Based Gas Detection in Hyperspectral Images

Abstract

Gas detection in industrial and environmental monitoring scenarios is a critical safety task. Hyperspectral imaging captures hundreds of contiguous spectral bands, providing rich spectral signatures enabling identification of gas plumes invisible to conventional cameras. This paper proposes a hybrid 3D Convolutional Neural Network (3D-CNN) and Autoencoder-based framework for gas detection in airborne and ground-based hyperspectral imagery. The 3DCNN captures joint spatial-spectral features across the hypercube, while the autoencoder learns a compact normalstate spectral representation enabling anomaly detection through reconstruction error thresholding. The combined architecture achieves detection accuracy of 93.7% with a false alarm rate of 2.1% on the Hyper Gas benchmark dataset, outperforming traditional Reed-Xiaoli and spectral matched filter detectors as well as standard 2D-CNN baselines.