Unsupervised land-cover clustering of hyperspectral images on a superpixel graph whose structure is learned jointly with the clustering. A graph encoder with adaptive filters captures both low- and high-frequency graph signal components instead of only low-pass smoothing; a self-training decoder sharpens soft assignments with a KL-divergence objective to generate pseudo-labels; and homophily-enhanced structure learning re-estimates node connections by orientation correlation and sparsifies heterophilous edges so the graph increasingly links same-class superpixels. Joint optimization alternates network self-training with graph updates, and K-means on the latent features yields the final clusters.
AHSGC (Adaptive Homophily Structure Graph Clustering) is an unsupervised clustering method for hyperspectral images that works on a superpixel graph rather than raw pixels. Homogeneous region generat…