An innovative method for extracting useful spectral-spatial characteristics from hyper-spectral imaging for remote sensing is presented in this research. The procedure comprises merging three feature sub-cubes produced by separate feature extraction methods into a single feature sub-cube. The resulting multi-feature extraction sub-cube is then enhanced using an intrinsic structure-preserving recursive filter to improve its spatial features. Then, the produced sub-cube is trained for classification using a SVM that has an RBF kernel. The Indian Pines and Salinas datasets are used to evaluate the performance of the proposed architecture. The results show that the overall accuracy achieved with this model is 96.28% for Indian Pines and 94.23% for Salinas, surpassing the performance of existing methods. This framework highlights its adaptability and relevance across a range of domains, including geological surveying and environmental monitoring.