Design of normalized matched filter for gas detection with hyperspectral images

Abstract

A hyperspectral imaging system (HIS) with a Fourier transform infrared (FTIR) spectrometer is an emerging technology for the remote detection and identification of chemical gas clouds. The NMF, which is uniformly most powerful invariant test, is widely used to chemical gas detection. However, it is vulnerable to a background contamination problem and a low signal-to-ratio (SNR) problem. In this thesis, we propose an hard expectation-maximization normalized matched filter (EM-NMF) to address the background contamination problem. We also propose a cooperative NMF based on the optimal cooperation scheme to address the low SNR problem.

To design the NMF, background statistics calculated from a background training dataset are required. However, in practice, because the training dataset is likely to contain gas-on background pixels, the performance of the NMF is severely degraded. This problem is referred to as a background contamination problem. To address this issue, we propose an algorithm that estimates the posterior probability of each pixel belonging to either the background or the gas class. The optimal posterior probabilities are obtained by maximizing the log-likelihood of a contaminated dataset using the EM algorithm. Based on the posterior probability, we extract gas-free background pixels from the contaminated dataset and design an hard EM-NMF with extracted gas-free background pixels. We demonstrate that the proposed algorithm is an effective solution for the background contamination problem, via experimental results conducted with actual chemical gas data measured by a Bruker HI--90 instrument in an outdoor setting as well as synthetic chemical gas data.

Given the small light received at each pixel in the hyperspectral image (HSI), the spectrum of each pixel has a low SNR and the detection performance of the NMF is limited. Therefore, we propose a linear cooperation scheme that allocates cooperation coefficients to the spectra of the neighbouring pixels. The optimal cooperation coefficients, which removes noise signatures while minimizing the distortion of gas signatures, are acquired by finding the maximum likelihood estimator (MLE) of the cooperation coefficients and determined according to the corresponding spectral data. Finally, we design a cooperative NMF with the optimal cooperation scheme. We demonstrate that the proposed cooperative NMF is capable of robust detection performance via outdoor experiments with actual chemical gas data.