DBXAI : Unlocking the Spectral insights for Hyperspectral Landcover Images with CNN-XAI Synergy

Such vast spectral data embedded in HSI challenges model interpretation and its efficiency. Hence, finding minimum spectral information is difficult in order to improve the efficiency of DNN models. In this respect, the performance of three CNN models, namely SSRN, RESNET34, and HybridSN, has been evaluated in the Indian Pines dataset and achieved high accuracies of 99.80%, 97.56%, and 99.56%, correspondingly using all bands. Using the explainable AI methods DeepLIFT, LRP, and GradientSHAP, the 14 most important bands were selected. The strong performance of the model remained high at 98.0%, 98.26%, and 96.82%, respectively. When further aggregation of bands across the XAI methods was performed, higher accuracies of 99. 46%, 98. 24%, and 98. 97% were achieved. In this way, a multimodal dataset that combined CNN-extracted features with XAI-derived band importance enabled classical machine learning models such as SVC, RF, and KNN to achieve superlative accuracy, with RF reaching 99.99%. This work is an example of how the combination of CNNs and XAI could provide useful and interpretable HSI classification, highlighting a trade-off between accuracy, interpretability, and computational efficiency.