Causal-Intervened Contrastive Learning With Collaborative Domain Alignment for Cross-Domain Few-Shot Hyperspectral Image Classification

Cross-domain few-shot hyperspectral image (HSI) classification aims to classify land cover categories in a target domain (TD) with scarce labels by transferring knowledge from a source domain (SD). However, the inherent spectral variability among different scenes often introduces spurious correlations, decreasing the generalization ability of few-shot models. In this article, we propose a novel causal-intervened contrastive learning with collaborative domain alignment (CCCDA) framework that mitigates these spurious correlations and enhances cross-domain feature alignment. Specifically, we propose a causal intervention-guided contrastive learning (CICL) module, which presents a random mask causal-intervened (RMCI) strategy to generate counterfactual samples by preserving class-related central regions while disturbing noncausal backgrounds. Subsequently, an intervention-invariant contrastive learning process is designed to minimize the feature distance between query samples and their counterfactual counterparts, enforcing the model to learn causally invariant features. Furthermore, to overcome the domain gap, we propose the collaborative domain alignment (CDA) module to integrate a causal meta-learning branch, which applies prototype matching on support and mask sets, with a conditional adversarial training branch that aligns domain-invariant features between the SD and TD. Extensive experiments on four publicly available target HSI datasets demonstrate that the proposed CCCDA framework is superior to other state-of-the-art (SOTA) methods. The code is available athttps://github.com/Chirsycy/CCCDA