Ensemble-EA and HS-SEA: Attention-Based Efficient Architectures With Bayesian Hyperparameter Optimization for RGB and Hyperspectral Remote Sensing Image Classification
Abstract
Conventional deep learning models still face challenges from substantial intra-class variability, complex spatial structures, inter-class similarity, and high spectral dimensionality in hyperspectral imagery. To address these limitations, a unified remote sensing classification framework is proposed consisting of two complementary architectures: Ensemble-EA for RGB aerial scene classification and HS-SEA for hyperspectral image classification. The Ensemble-EA architecture integrates a truncated EfficientNet-B0 backbone, a self-attention mechanism, hierarchical feature learning, and ensemble decision refinement to capture local and global context in RGB imagery. For hyperspectral data, the proposed HS-SEA architecture introduces a Band Grouping Attention Module (BGAM) and a Spectral-to-Spatial Transfer Gate (SSTG) to exploit spectral-spatial dependencies while efficiently mitigating the Hughes phenomenon. Bayesian optimization is employed to optimize the downstream ensemble classifiers. The proposed framework is validated on nine benchmark datasets, including EuroSAT, NWPU-RESISC45, Bijie Landslide, Turkey Earthquake, MLRSNet, Indian Pines, Salinas, Pavia University, and the recently launched Geology-1 hyperspectral microsatellite dataset. Experimental results show superior classification performance across RGB and hyperspectral modalities, achieving overall accuracies of 6.75% on EuroSAT, 99.28% on NWPU, 93.31% on Bijie Landslide, 98.90% on Turkey Earthquake, 95.00% on MLRSNet, 99.15% on Indian Pines, 99.87% on Salinas, and 99.48% on Pavia University. On the labelscarce Geology-1 dataset, cross-site transfer learning using HS-SEA embeddings and a linear classifier trained on one site achieved 46% and 68% accuracy when transferred to two geographically distinct, unseen sites, demonstrating the framework's ability to generalize to novel, operationally relevant geological hyperspectral data. The results validate the framework's robustness, cross-modality generalization, and discriminative feature-learning effectiveness.