Answers are grounded in this paper and cite their sources.
Hyperspectral images (HSIs), with rich spectral information and tightly coupled spatial-spectral structures, are widely used in environmental monitoring, precision agriculture, geological exploration, and defense. However, complex noise introduced by sensors, atmosphere, and processing severely degrades HSI quality and downstream analysis. This paper presents a comprehensive and critical survey of HSI denoising. We first build a unified taxonomy of realistic HSI noise (including mixed and deadline noise), prior constraints, model-based methods, and deep architectures, and summarize public datasets and evaluation metrics for benchmarking. Beyond cataloging methods, we analyze how different priors and network designs, such as CNNs, Transformers, state-space networks, and diffusion models, exploit spatial–spectral correlations and cope with complex noise. To bridge the gap between methodology and practice, we conduct comparative experiments on Indian Pines and Pavia University with synthetic mixed noise and on HYDICE Urban with real sensor noise, covering classical low-rank tensor and recent deep models. The results show that tensor low-rank approaches form strong baselines, while T3SC achieves the best overall performance and generalization to real noise. Finally, we identify key challenges and future directions, including physics-informed and self-/unsupervised learning, realistic real-noise benchmarks, and lightweight denoising models for resource-constrained applications.
No takes yet. Share an insight, caveat, or question.
Zhao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: