ABSTRACT
With the past explosion in sensor technology, and recent advancement in high resolution imaging with Unmanned Aerial Systems, hyperspectral images with high spectral and spatial resolution are becoming more prevalent. These hyperspectral images with high spatial resolution have created data mining problems because of their large size, and spatial and spectral dependencies. Many of the multispectral data processing methods are not effective for hyperspectral data because of their inability to handle big data. Many of the deep learning methods are not appropriate because of their inability to handle spatial and spectral dependencies. This chapter explores data mining processes and methods that are well suited for hyperspectral image data. Data mining of hyperspectral data typically has two major steps; the first is feature extraction and selection, and the second is information extraction. In the feature extraction/selection step, the high-dimensional hyperspectral data are transformed to a fewer number of features that are ideally uncorrelated to each other, and best at estimating or divulging the information of interest. In the information extraction step, the features selected in the previous steps are used to estimate the information of interest using a supervised or unsupervised classification method. Various methods used for both feature extraction/selection and information extraction are described in this chapter. The progress in classification methods started from pixel-based methods to methods incorporating spatial contextual information along with pixel values to true end-to-end spectral-spatial methods. Overall, the trend in classification is fast moving towards end-to-end classification (where feature section/extraction and classification all happen in one process) methods with deep learning structure that utilizes spectral-spatial features. Convoluted neural networks and support vector machines are some of the popular methods that are shown to provide high classification accuracies.
