Nonlinear Unsupervised Clustering of Hyperspectral Images with Applications to Anomaly Detection and Active Learning

The problem of unsupervised learning and segmentation of hyperspectral images is a significant challenge in remote sensing. The high dimensionality of hyperspectral data, presence of substantial noise, and overlap of classes all contribute to the difficulty of automatically clustering and segmenting hyperspectral images. In this article, we propose an unsupervised learning technique that combines a geometric estimation of class modes with a diffusion-inspired labeling that incorporates both spatial and spectral information. The mode estimation incorporates the geometry of the hyperspectral data by using diffusion distance to promote learning a unique mode from each class. These class modes are then used to label all points by a joint spatial-spectral nonlinear diffusion process. The proposed method, called spatial-spectral diffusion learning (DLSS), is shown to perform competitively against benchmark and state-of-the-art hyperspectral clustering methods on a variety of synthetic and real datasets. The proposed methods are shown to enjoy low computational complexity and fast empirical runtime. Two variations of the proposed method are also discussed. The first variation combines the proposed method of mode estimation with partial least squares regression (PLSR) to efficiently segment chemical plumes in hyperspectral images for anomaly detection. The second variation incorporates active learning to allow the user to request labels for a very small number of pixels, which can dramatically improve overall clustering results. Extensive experimental analysis demonstrate the efficacy of the proposed methods, and their robustness to choices of parameters.