Hyper Spectral Image Clustering and Local Feature Selection using Gini Impurity

This study proposes a unique segmentation-based clustering algorithm that utilises k-means for segmentation, further uses a local feature selection technique to obtain the top bands for each cluster and deploys clustering on segmented hyperspectral imagery. The suggested methodology is a framework with several stages. k-means is initially utilized for image segmentation. From the obtained segments, significant segments are identified using Gini impurity. Finally, the cluster map is obtained by merging insignificant clusters with significant clusters. This step also makes use of novel local feature selection strategy. Three sets of hyperspectral images are used in investigations to evaluate the efficiency of the proposed methodology. For assessment, the criteria Normalized Mutual Information and Purity score are utilised. The investigation findings demonstrate that the proposed methodology outperforms the other segmentation methodologies that were compared. According to the results, using band selection and redundancy strategies significantly improves accuracy.