Self-supervised spatial-spectral transformer with Extreme Learning Machine for Hyperspectral Image Classification

While deep learning has significantly advanced hyperspectral image (HSI) classification, capturing complex spatial-spectral features efficiently and achieving robust performance with limited dynamic masking within masked image sparse attention mechanism enhances computational efficiency by selectively attending to key spectral-spatial patches, effectively reducing complexity, without compromising classification performance. An Extreme Learning Machine (ELM) is integrated as the final classification layer, leveraging SST-extracted features for efficient and lightweight classification. The resulting hybrid SST-ELM model achieves 99.93% overall accuracy on the Salinas dataset, improving upon the baseline by 1.12% with a 5 reduction in training time. Similar improvements are observed on Pavia University (99.06%), Longkou (99.83%), Hanchuan (96.04%), and Honghu (97.10%) datasets.