Gas Detection Using 3D CNN and Auto Encoder in Hyper Spectral Images

One of the most important issues for both human health and the environment is the monitoring of gas emission levels. With its capacity to detect gases at safe distances, hyper spectral image analysis has significantadvantages over conventional gas detection systems. In this article, we propose a deep learning-based hyperspectral gas detection method that combines unfixing and classification. It is observed that current hyperspectral gas detection methods in the thermal range ignore the fact that the captured radiance in the long wave infrared (LWIR) spectrum is better modeled as a mixture of the radiance of background and target gases. The radiance data is first converted to luminance-temperature data using the suggested approach. The obtained data is then subjected to a 3-D convolutional neural network (CNN) and auto encoder-based network, which is specifically made for unfixing, in order to obtain end members and abundances for every pixel. Ultimately, a three-layer fully connected network detects the target gases at each pixel by using the abundance values and extracted end member spectra. Using LWIR hyper-spectral images of methane and sulfur dioxide gases, the suggested method's better performance over traditional hyper spectral gas detection techniques utilizing spectral angle mapper and adaptive cosine estimator is confirmed. the contribution of the suggested system has been demonstrated by the ablation investigation concerning various combinations of the suggested structure, including direct classification and unfixing techniques.