Fully Unsupervised Binary Change Detection for Hyperspectral Images Using Laplacian Eigenmaps and Clustering

Due to the high dimensionality and complexity of hyperspectral images, change detection has proven to be a challenging study field in multi-temporal remote sensing. More sophisticated techniques are required to exploit the rich information and reduce the redundancy of spectral bands and, as a result, enhance the quality of change maps. This paper proposes a manifold-based approach for binary change detection in multitemporal hyperspectral images using Laplacian Eigenmaps. The multitemporal difference is represented in the eigenspace of the Laplacian matrix, and the resulting latent vectors are utilized to cluster the changed vs. unchanged regions using k-means clustering. The clusters obtained from the first two latent vectors are combined to obtain the binary change detection map. The proposed method is fully supervised and no thresholding is required. The proposed approach is validated on two real bitemporal hyperspectral datasets.