Optimization of Hyperspectral Image Classification Using Monte Carlo-Based Band Reduction

Hyperspectral images are the concatenation of images taken at different wavelengths with a very narrow bandwidth, from ultraviolet to mid-infrared. These different images are combined into a hypercube structure of very large size, which contains much more information than classical RGB images. One of the drawbacks of using a hyperspectral camera is the time it takes to acquire a hypercube, up to few minutes. However, it's easy to see that to detect a particular feature in a hypercube, only a part of the wavelengths are needed, allowing the use of multispectral cameras i.e. only a reduced set of wavelengths, acquired at the same speed as RGB cameras. The choice of wavelengths to capture depends on the detection required, in our case the segmentation of the image. Therefore, this paper focuses on the use of the Monte Carlo method to find the best wavelengths needed, and the use of a form of unsupervised classification through the clustering method k-means to evaluate the results. Multiple simulations were performed using the Monte Carlo method to select the wavelengths and several metrics were defined to compare the results between using all the wavelengths and using only a few of them.