|
Abstract:
|
The use of hyperspectral imagery in the remote sensing of gas plumes has proven to be vital for many military applications. For example, hyperspectral images can be used to detect a gas cloud invisible to the human eye, and to identify its chemical structure. A hyperspectral image is a massive cube of data composed of thousands of pixels each with around 100 observations over a range of frequencies in the electromagnetic spectrum. Algorithms that use hypothesis testing and assume independence over pixels have shown success in detecting plumes, but fail in identifying chemical components. In this paper we explain that identification is a variable/model selection problem, which can be solved robustly by taking advantage of spatial information in the image. We develop a Bayesian spatial model selection algorithm, which uses mixtures of g-priors and incorporates a Markov Random Field prior to induce dependence among neighboring pixels. We apply our model to several partially synthesized hyperspectral images, and through the use of confusion matrices and multi-class classification metrics, we show that our method outperforms state-of-the-art algorithms such as the LASSO and fused LASSO.
|
