Improving Multispectral Image Classification with Convolutional Networks and Hyperspectral Band

This research work proposes an efficient framework to improve multispectral image classification through convolutional neural networks (CNNs) with optimized hyperspectral band selection. Hyperspectral images contain extensive spectral information, but the high dimensionality introduces redundant data, impacting processing efficiency and model performance. The proposed approach employs a hybrid band selection technique that combines mutual information and clustering algorithms to identify the most discriminative spectral bands before feeding them into a CNN model. This process enhances classification accuracy while significantly reducing computational load. Experimental results on the Indian Pines and Salinas hyperspectral datasets demonstrate the effectiveness of the proposed method. With our optimized band selection, the CNN achieved an overall classification accuracy of 95.7% on the Indian Pines dataset, an increase of 6.4 percentage points over the full-band CNN baseline accuracy of 89.3%, and reduced training time by 43%. Similarly, on the Salinas dataset, classification accuracy improved from 90.8% to 96.1%, with a 47% reduction in computation time. These results highlight the potential of hyperspectral band optimization in CNN-based classification, providing a robust and efficient approach for high-accuracy, real-time multispectral imaging applications.