Pointwise Mutual Information-Based Band Selection for Hyperspectral Image Classification

This paper introduces a novel hyperspectral image classification method using Pointwise mutual information-based band selection (PMIBS). The proposed method aims to achieve higher classification accuracy while significantly reducing spectral redundancy between bands. The key innovation lies in adaptive weighting through PMIBS, which optimizes the trade-off between computational efficiency and redundancy reduction. The method was evaluated on the datasets TG1HRSSC and Washington-DC; it was, therefore, shown to outperform conventional methods, including PCA, mRMR, and Band Prioritization, by quantitative assessment of overall and average accuracy, and kappa coefficient (κ). The PMIBS method achieved a very impressive reduction of 88.45% redundancy for the Washington-DC dataset and 85.24% for the TG1HRSSC dataset, at a great improvement over other technique. The results shows that PMIBS is an efficient real-time method for hyperspectral image classification implementations, particularly large datasets. Future work is on integrating by means of the PMIBS framework into deep learning architectures, improving the accuracy of classification as well as steps towards system scalability concerning more complicated hyperspectral data.