Efficient capsule network with attention mechanism for hyperspectral remote sensing classification

The capsule network (Caps) has emerged as a promising neural network architecture for hyperspectral remote sensing classification, yet it faces critical challenges: gradient vanishing and computational inefficiency. To address these limitations, we propose an Efficient Caps (ECaps) that integrates a non-iterative self-attention routing mechanism for hyperspectral classification. ECaps was validated using three widely adopted public datasets: Pavia University (PU), Salinas-A (SA), and Indian Pines (IP). Its performance was evaluated and comparing it against benchmark models, including Random Forest, Support Vector Machine, Convolutional Neural Network, Caps, Caps integrated with coordinate attention module, Caps integrated with convolutional block attention module, VGG16, and DBANet. The overall accuracies of the ECaps achieved 98.72%, 97.96%, and 94.98% on the PU, SA, and IP datasets, respectively. These results outperformed the benchmark models. Notably, ECaps exhibited higher computational efficiency than the Caps. Collectively, these findings validate ECaps as a practical solution for hyperspectral remote sensing classification tasks.