Exploring Gas Detection in Hyperspectral Imaging Through 3D-CNN and Autoencoder Models

The rapid industrial development has contributed to environmental pollution, particularly through gas emissions from industrial processes. Detecting and monitoring these gas leaks is crucial for mitigating environmental harm. This paper presents a novel approach for detecting industrial gas emissions using hyperspectral images and a 3D Convolutional Neural Network (3D-CNN)-based autoencoder-decoder model. Hyperspectral images, typically captured from drones or satellites, provide spectral data to detect gas emissions. The Spectral Angle Mapper (SAM) formula is applied to hyperspectral data, matching pixel distances with known gas signatures from a spectral database. A custom dataset of industrial gas leaks, along with publicly available methane and sulfur leak datasets, is used for training. The 3D-CNN model achieves 88% accuracy in detecting gas leaks. Additionally, a hybrid ensemble model combining CNN, RNN, and GRU networks outperforms the 3D-CNN model with 93% accuracy. This work provides a framework for accurate gas leak detection in industrial settings using hyperspectral imaging.