Cross-Sensor Hyperspectral Image Classification via Multiscale Spectral–Spatial Representation Learning and HyperLoRA-Based Adaptation

Abstract

Hyperspectral image classification has significantly benefited from recent advances in deep learning. However, its performance remains constrained by limited labeled samples, high spectral dimensionality, and substantial cross-sensor variability, which hinder the generalization ability of models trained on limited or heterogeneous datasets. To address these challenges, this article proposes a unified framework for cross-sensor hyperspectral image classification that integrates representation learning from scratch with lightweight parameter-efficient adaptation. In the first stage, a multiscale branch fusion network is developed to enhance spectral–spatial representation learning by aggregating multiscale contextual information centered on each pixel. In the second stage, a HyperLoRA-based adaptation framework with adversarial robustness enhancement is introduced to improve cross-sensor robustness while updating only a small number of parameters, enabling efficient and scalable model transfer across heterogeneous sensors. Extensive experiments on five benchmark hyperspectral datasets, including Pavia University, Indian Pines, Houston2013, Botswana, and WHU-Hi-LongKou, demonstrate that the proposed framework achieves superior or highly competitive classification performance with favorable efficiency, highlighting its effectiveness for robust hyperspectral image classification under cross-sensor scenarios.