UniConEt: A novel Hybrid Deep architecture for fine grained Land Use and Land Cover Classification Using Hyperspectral Imagery

In the rapidly developing field of Remote Sensing (RS), Land Use Land Cover (LULC) classification of Hyperspectral Images (HSI) is a fundamental yet challenging task. HSI is also challenging because of its high spectral dimensionality. Therefore, this research develops a comprehensive spectral and spatial classification methodology, UniConEt, using the Indian Pines and Salinas datasets. The workflow starts with demanding image pre-processing stages that consist of three approaches for performing three functions: flat-field correction for radiometric calibration, the empirical line method (ELM) for atmospheric correction, and Counterlet (CT to improve the HSI’s spectral reliability, followed by dimensionality reduction. To improve the discriminative capability, spectral and spatial features were extracted and optimized using the mRMR and ReliefF selection algorithms. After that, an intelligent hybrid Deep Learning (DL) framework is proposed to integrate the spectral and spatial indicators for enhanced LULC classification. Also, the model further improves transparency via Explainable AI (XAI) strategies such as Grad-CAM, SHAP, and saliency maps, which are combined to identify significant spatial regions and spectral bands contributing to improved LULC classification results. The developed methodology provides interpretability visualizations, uncertainty evaluations, and higher-resolution LULC maps. Additionally, experimental analysis using stratified cross-validation shows that the accuracy is 0.99, the precision is 0.99, the recall is 0.99, the F1 Score is 0.99, the Kappa is 0.989, and the SSIM is 0.992 for the Salinas dataset. Similarly, the Indian Pines dataset has an accuracy of 0.9765, a precision of 0.976, a recall of 0.9765, an F1 Score of 0.976, a Kappa of 0.96, and an SSIM of 0.976.