当前，语义分割是遥感图像处理领域的热点及难点问题。虽然强监督学习在遥感影像语义分割任务上取得了巨大成功，但是，大量领域专家知识和海量精确标注导致数据标注成本急剧升高。因此，有学者将目光投向弱监督语义分割。然而，弱监督学习大多数采用图像级标注，存在标注信息不精确的问题；同时，遥感影像自身又常常会受到雾霾、云层等因素的干扰。两者叠加将对遥感影像语义分割效果产生重大影响。本文针对上述问题提出两种新的含雾遥感图像弱监督语义分割算法：

1）基于注意力金字塔特征融合网络的含雾遥感图像语义分割算法。该算法分为去雾阶段和分割阶段。去雾阶段：首先通过设计不同尺度的密集卷积群，构建注意力金字塔特征提取结构，然后通过多尺度通道-空间注意力模块增强多尺度融合特征，最后经过特征重构和全局跳跃连接得到去雾结果。分割阶段：采用基于渐进监督学习的语义分割算法，首先训练分类网络，然后通过GradCAM算法生成伪标签，最后使用伪标签训练反馈显著性分析网络。去雾后的图像经过该网络后输出即为对应的分割结果。

2）基于显著性感知对齐策略的含雾遥感图像语义分割算法。该算法受一致性正则化启发，主要由对齐网络和目标网络构成，能够有效优化伪标签生成阶段的输出结果。流程如下：首先在使用清晰遥感图像训练对齐网络后，再用含雾遥感图像训练目标网络，通过计算两个网络之间的一致性损失，使目标网络对含雾图像具有鲁棒性。然后使用目标网络生成的伪标签训练反馈显著性分析网络。含雾图像经过该网络处理后得到对应的分割结果。

本文将提出的两种方法与多种图像去雾算法和基于去雾预处理的含雾图像语义分割方法进行了比较，实验结果表明，本文提出的方法可以更好地针对含雾遥感图像实现语义分割。

Currently, semantic segmentation is a hot and difficult problem in the field of remote sensing image processing. Although supervised learning has achieved great success in remote sensing image semantic segmentation tasks, the cost of data annotation has risen sharply due to a large amount of domain expert knowledge and massive accurate annotations. Therefore, some researchers have turned their attention to weakly supervised semantic segmentation. However, weakly supervised learning mostly uses image-level annotations and suffers from inaccurate annotation information. At the same time, remote sensing images are often interfered by factors such as haze and clouds. The combination of the two will greatly affect the effect of semantic segmentation. This paper proposes two novel algorithms for semantic segmentation of hazy remote sensing images in response to the above two research difficulties:

1) A hazy remote sensing image semantic segmentation based on an attention pyramid feature fusion network is proposed. The algorithm is divided into a dehazing stage and a segmentation stage. Dehazing stage: First, an attention pyramid feature extraction structure is constructed by designing dense convolution groups of different scales. Then, multi-scale fusion features are enhanced through multi-scale channel-space attention modules. Finally, post-dehazing processed images are obtained through feature reconstruction and global skip connections. Segmentation stage: A semantic segmentation algorithm based on progressive supervised learning is adopted. First, the classification network is trained, then pseudo-labels are generated through the GradCAM algorithm, and finally the feedback saliency analysis network is trained using pseudo-labels. The post-dehazing processed image is passed through this network to obtain the corresponding segmentation result.

2) A hazy remote sensing image semantic segmentation based on saliency analysis is proposed. Inspired by consistency regularization, this paper proposes a model framework based on saliency-aware alignment strategy, mainly composed of an alignment network and a target network, used to optimize the pseudo-label results during the pseudo-label generation stage. The main process is as follows: After training the alignment network using clear remote sensing images, the target network is trained using hazy remote sensing images. By calculating the consistency loss between the two networks, the target network has robustness to hazy images. Then use the pseudo-labels generated by the target network to train the feedback saliency analysis network. The hazy image is passed through this network to obtain the corresponding segmentation result.

This paper compares the two methods proposed with various image dehazing algorithms and hazy image semantic segmentation methods based on dehazing preprocessing. The experimental results show that the methods proposed in this paper can better achieve semantic segmentation for hazy remote sensing images.