Enhancing Hyperspectral Image Classification Through Reinforcement Learning Guided Active Learning

Hyperspectral image (HSI) classification is vital for environmental monitoring, land cover mapping, and precision agriculture, but its effectiveness is often constrained by the scarcity of labeled samples and high spectral similarity among classes. To address these challenges, we propose CrossCapsViT, a hybrid classification framework that integrates capsule networks (CapsNets) and vision transformers (ViTs) through a cross-attention fusion mechanism and a cross-layer adaptive fusion module, enabling richer and more discriminative spectral-spatial feature learning. To further improve efficiency in data scarce scenarios, we embed an actor-critic reinforcement learning-based active learning (RAL) strategy that jointly leverages accuracy, uncertainty, and diversity in the reward structure, guiding the selection of the most informative samples while reducing labeling effort. Experiments conducted on four benchmark datasets (Kennedy Space Center, Pavia University, Houston University 2013, and Salinas) and a custom UAV-based saltmarsh dataset (Derrymore, collected with a Pika-L sensor) demonstrate that CrossCapsViT with RAL consistently outperforms CapsViT and other baseline models in terms of classification accuracy, robustness, and generalizability. The proposed framework achieves up to 25% improvement in class-level accuracy on challenging vegetation classes, while reducing dependence on large annotated datasets, highlighting its potential for practical deployment in real-world ecological monitoring and remote sensing applications.