The noise pollution issue seriously obstructs subsequent interpretation and application of the hyperspectral image (HSI). In this work, differing from most existing HSI denoising methods ideally assumed that noise in different bands denotes independent & identically distributed (i.i.d.), we propose a novel HSI denoising approach focusing on non-i.i.d. noise removal. The presented framework collaboratively models the non-i.i.d. noise embedding within HSI and removals them under a deep spatio-spectral Bayesian posterior (DSSBP) structure. Specifically, the non-i.i.d. noise estimation, distribution and removal procedure are both executed with the model-driven based strategy and data-driven based strategy. Through blending the Bayesian variational posterior and deep convolutional neural network, the proposed method both inherits the reliability of traditional model-driven based methods for HSI noise modeling and the high efficiency of data-driven based methods for parameters learning. Simulated and real experiments in different HSIs and non-i.i.d. noise scenarios testify that the proposed DSSBP approach outperforms other existing methods for non-i.i.d. noise removal, in terms of evaluation indexes and executive efficiency.

噪声污染问题严重阻碍了高光谱图像（HSI）的后续解释和应用。在这项工作中，与大多数现有的HSI去噪方法不同，理想地假设不同频带中的噪声表示独立且均匀分布（iid），我们提出了一种专注于非iid噪声去除的新颖HSI去噪方法。提出的框架协同建模非iid噪声嵌入HSI，并在深时空光谱贝叶斯后验（DSSBP）结构下将其去除。具体地，非空噪声估计，分配和去除过程都通过基于模型驱动的策略和基于数据驱动的策略来执行。通过混合贝叶斯变分后验和深度卷积神经网络，该方法既继承了传统的基于模型驱动的HSI噪声建模方法的可靠性，又继承了基于数据驱动的参数学习方法的高效率。在不同的HSI和非iid噪声场景中进行的模拟和真实实验证明，在评估指标和执行效率方面，所提出的DSSBP方法优于其他现有的非iid噪声消除方法。