Chemicals released in the air can be extremely dangerous for human beings andthe environment. Hyperspectral images can be used to identify chemical plumes,however the task can be extremely challenging. Assuming we know a priori thatsome chemical plume, with a known frequency spectrum, has been photographedusing a hyperspectral sensor, we can use standard techniques like the so calledmatched filter or adaptive cosine estimator, plus a properly chosen thresholdvalue, to identify the position of the chemical plume. However, due to noiseand sensors fault, the accurate identification of chemical pixels is not easyeven in this apparently simple situation. In this paper we present apost-processing tool that, in a completely adaptive and data driven fashion,allows to improve the performance of any classification methods in identifyingthe boundaries of a plume. This is done using the Multidimensional IterativeFiltering (MIF) algorithm (arXiv:1411.6051, arXiv:1507.07173), which is anon-stationary signal decomposition method like the pioneering Empirical ModeDecomposition (EMD) method. Moreover, based on the MIF technique, we proposealso a pre-processing method that allows to decorrelate and mean-center ahyperspectral dataset. The Cosine Similarity measure, which often fails inpractice, appears to become a successful and outperforming classifier whenequipped with such pre-processing method. We show some examples of the proposedmethods when applied to real life problems.
展开▼
