Hyperspectral chemical plume detection algorithms based on multidimensional iterative filtering decomposition

Abstract

Chemicals released in the air can be extremely dangerous for human beings and the environment. Hyperspectral images can be used to identify chemical plumes, however the task can be extremely challenging. Assuming we know a priori that some chemical plume, with a known frequency spectrum, has been photographed using a hyperspectral sensor, we can use standard techniques such as the so-called matched filter or adaptive cosine estimator, plus a properly chosen threshold value, to identify the position of the chemical plume. However, due to noise and inadequate sensing, the accurate identification of chemical pixels is not easy even in this apparently simple situation. In this paper, we present a post-processing tool that, in a completely adaptive and data-driven fashion, allows us to improve the performance of any classification methods in identifying the boundaries of a plume. This is done using the multidimensional iterative filtering (MIF) algorithm (Cicone et al. 2014 (http://arxiv.org/abs/1411.6051); Cicone & Zhou 2015 (http://arxiv.org/abs/1507.07173)), which is a non-stationary signal decomposition method like the pioneering empirical mode decomposition method (Huang et al. 1998 Proc. R. Soc. Lond. A 454, 903. (doi:10.1098/rspa.1998.0193)). Moreover, based on the MIF technique, we propose also a pre-processing method that allows us to decorrelate and mean-centre a hyperspectral dataset. The cosine similarity measure, which often fails in practice, appears to become a successful and outperforming classifier when equipped with such a pre-processing method. We show some examples of the proposed methods when applied to real-life problems.

Keywords: empirical mode decomposition; iterative filtering; threat detection.

Figure 1.

(a) Contrast-enhanced spectral-mean image using the raw data from the dataset Location 1 released r134a and (b) groundtruth classification. Black, plume pixels; orange, pixels close to the boundaries of the plume. (Online version in colour.)

Figure 2.

(a) Detection map of the ACE classification of the raw data and (b) detection map after applying the PostP method to the ACE classification of the raw data. (Online version in colour.)

Figure 3.

(a) ROC curves for the ACE classification of the hypercube Location 1 released r134a with and without post-processing and (b) ROC curves in log scale along the horizontal axis. (Online version in colour.)

Figure 4.

(a) Detection map of the wavelets post-processed ACE classification of the raw data and (b) detection map of the MEEMD post-processed ACE classification of the raw data. (Online version in colour.)

Figure 5.

Detection map of the reversed COS classification of the raw data and (b) detection map of the COS classification of the pre-processed data. (Online version in colour.)

Figure 6.

(a)ROC curves for the COS classification of the hypercube Location 1 released r134a and (b) ROC curves comparison for the hypercube Location 1 released r134a. (Online version in colour.)

Figure 7.

(a)Detection map of the PostP post-processed COS classification of the pre-processed data and (b) detection map of the wavelets post-processed COS classification of the pre-processed data. (Online version in colour.)

Figure 8.

(a)Raw spectral signature of a pixel inside the plume and the chemical r134 signature and (b) pre-processed spectral signature of a pixel inside the plume. (Online version in colour.)

Figure 9.

(a) Contrast-enhanced spectral-mean image using the raw data Location 2 released sf6 blind and (b) detection map of the ACE classification of the raw data. (Online version in colour.)

Figure 10.

(a) Detection map of the MIF post-processed ACE classification of the raw data and (b) detection map of the wavelets post-processed ACE classification of the raw data. (Online version in colour.)

Figure 11.

(a) Detection map of the MF classification of the raw data and (b) detection map of the MIF post-processed MF classification of the raw data. (Online version in colour.)

Figure 12.

(a) Detection map of the wavelets post-processed MF classification of the raw data and (b) detection map of the reversed COS classification of the raw data. (Online version in colour.)

Figure 13.

(a) Detection map of the COS classification of the pre-processed data and (b) detection map of the post-processed COS classification of the pre-processed data. (Online version in colour.)