CINet: Causal Inference inspired Network for Cross-Scene Hyperspectral Image Classification

Abstract Hyperspectral image (HSI) classification across varying geographical scenes demonstrates significant potential in mitigating data discrepancies caused by environmental diversity, particularly in operational domains requiring robust remote sensing interpretation. Current several approaches exhibit improved generalization performance in the domain of cross-scene hyperspectral image classification. However, object similarity, data heterogeneity, and imbalanced sample distribution remain critical challenges that significantly constrain the generalization capability of cross-scene hyperspectral classification. To address these problems, we propose a Causal Inference inspired Network for hyperspectral image cross-scene classification (CINet). The Causal Inference Block based on the Random Fourier Features is introduced in the discriminator, which eliminates spurious feature correlations and enhances focus on relevant feature-label relationships by sample weighting. To enhance the robustness and generalization of the cross-scene task, we incorporate supervised contrast learning and contrast-adversarial training in the discriminator, where the former optimizes the sample distance and the latter suppresses inter-domain differences by generating extended domains with significant differences. Finally, we conduct extensive experiments on two public datasets: Houston and Pavia. Compared with two Domain Adaptation methods and three Domain Generalization methods, the overall accuracy of our proposed CINet is more than 5% higher than the two DA methods, and 0.5%-2.5% higher than the three DG methods. The experimental results highlight the superior performance of CINet, confirming its effectiveness in cross-scene classification task.