Cross-domain hyperspectral image classification

Cross-domain hyperspectral image (HSI) classification represents a significant and complex challenge within the realm of remote sensing. HSIs obtained from diverse domains, captured at different times and under varying environmental conditions, frequently display substantial discrepancies. These variations can lead to a marked deterioration in the efficacy of classification algorithms when they are applied to novel scenes, thereby constraining the practical implementation of existing technologies. This paper presents a thorough review of the challenges associated with cross-domain HSI classification, examining the issues from the perspectives of problem definition, methodology, datasets, and applications. Initially, we introduce an innovative framework for cross-domain hyperspectral image classification tasks, organizing current advancements into four distinct sub-tasks: label-limited, label-free, unknown-label, and sample-free. For each sub-task, we systematically summarize recent technical advancements and notable methodologies. Subsequently, we conduct a comparative analysis of state-of-the-art techniques for these tasks, utilizing several widely recognized datasets to underscore current technological progress and trends. Additionally, we adopt a novel approach to address the emerging challenge of cross-domain data heterogeneity, specifically focusing on spectral heterogeneity arising from differences in imaging devices as a case study, which provides valuable insights into this issue. Finally, we outline the existing challenges and propose promising future research directions in cross-domain HSI classification, with the aim of stimulating further investigation and enhancing the applicability of HSI classification techniques in practical scenarios.