随着传感器与遥感技术的不断发展，成像分辨率快速提高，数据规模的激增促使人们对于遥感影像自动解译的需求更为迫切。近年来，显著性分析理论与深度学习技术已在计算机视觉相关领域取得了突破性进展。显著性分析理论可以大幅提升模型对于复杂场景的处理效率，深度学习技术拥有十分强大的场景分析能力，二者结合能够大幅提高目标提取的精度，从而有效缓解遥感影像高速获取与低速解译间的矛盾。考虑到遥感影像自身特点，目前，基于深度学习的遥感影像显著性分析方法仍有如下问题亟待解决：

1）强监督学习方法依赖于大型的、具有像素级标注的训练数据集，需要大量的专家知识，人力成本消耗较高。

2）弱监督学习方法虽然可以有效降低模型对于标签精度的依赖。但是受标签精度的制约，模型精度与强监督方法仍有较大差距，繁琐的后处理步骤对于时间成本消耗较大。

针对上述问题，本文提出一种由弱到强的研究思路：首先，提出了一种基于分层弱监督学习的显著性分析方法，有效弥补了单一尺度特征的局限性，提升了弱监督显著性分析方法的检测精度；然后，提出了一种基于多视点学习与注意机制的显著性分析方法，结合深度特征对背景的抑制能力与视觉注意机制对于数据分布敏感性较低的优势，降低了数据分布对于模型精度的影响；再次，提出一种联合强、弱监督学习的显著性分析方法，结合强监督学习方法的强大表达能力与快速前传特性，在轻量模型下进一步提高了模型的提取精度与计算效率；最后，将弱监督显著性分析方法引入到遥感影像超分辨率重建中，实现了兼顾空间质量与视觉清晰度的遥感影像超分辨率重建。

本文的主要工作如下：

1）提出了一种基于分层弱监督学习的遥感影像显著性分析及目标提取方法，有效解决了手工提取特征泛化能力不足的问题，弥补了单一尺度特征在弱监督学习方法中的局限性。以分类卷积神经网络作为图像级标注和像素级显著图之间的媒介，将颜色相似性作为约束条件，利用低秩恢复算法融合不同尺度下输出层神经元对于中间卷积层的梯度特征响应，通过计算类间显著图的残差图谱，完成对于显著区域的增强与非显著区域的抑制，从而在低标注成本下，较为准确地实现了遥感影像显著性分析及目标提取。

2）提出了一种不均衡数据集下基于多视点学习与注意机制的显著性分析及目标提取方法，结合了深度特征对背景的抑制能力，与注意机制对数据分布敏感度低的特点，有效降低了数据分布失衡对于模型精度的影响。在训练阶段，利用加权的交叉熵损失函数完成网络参数的训练，以提升模型对小类目标的检测能力。在测试阶段，首先利用多视点学习对于复杂背景的抑制能力，完成初始显著图计算；其次，结合自底向上的注意机制对于训练集分布依赖较低、边界刻画准确的特性，对于初始显著区域进行颜色纹理特征提取，并完成初始显著图的修正，有效实现了不均衡数据集下遥感影像的显著性分析及目标提取。

3）提出了一种联合强、弱监督学习的遥感影像显著性分析及目标提取方法，将伪标签显著图作为强、弱监督之间的纽带，在轻量模型下大幅提升了测试阶段的计算效率与提取精度。首先，在弱监督机制下，提出一种无标记数据的伪标签显著图生成方法，以有效提升后续网络的训练数据规模；然后，构建了基于强监督学习的自纠错反馈卷积神经网络，引入反馈连接及侧向输出结构，提升轻量模型的场景分析能力；最后，构建混淆热图以衡量各像素的置信度，结合课程学习策略与降噪损失函数完成对于伪标签显著图的降噪纠错。在低标注成本下，利用轻量模型实现了快速、准确的遥感影像显著性分析及目标提取。

4）提出了一种基于显著性判别生成对抗网络的遥感影像超分辨率重建方法，利用显著性分析理论指导深度学习框架中遥感影像不同区域的差异化重建，有效抑制了重建结果的空间失真与伪纹理。首先，设计了显著性驱动的残差生成网络，引入显著图以衡量区域纹理结构的复杂性；然后，设计了一种显著性判别的对偶判别网络，利用弱监督方法生成的显著图作为门控限制，完成显著区域与非显著区域的分解，通过两个参数不同的判别器，完成对于两种区域的差异化度量；最后，通过生成器与判别器的对抗学习，实现了遥感影像的差异化重建，提升了重建结果的视觉清晰度，同时抑制了对抗学习中伪纹理的产生。

上述研究工作能为遥感影像的显著性分析及目标提取提供新思路，同时，也将为植被估计、农作物监测、城市规划等相关领域提供重要的理论支撑与技术保障。

With the development of sensors and remote sensing techniques, resolutions of remote sensing imagery have been significantly improved. The rapid growth in the scale of remote sensing images has raised urgent demands for automatic interpretation of remote sensing images. In recent years, saliency analysis theory and deep learning techniques have demonstrated excellent capabilities in image understanding in the computer vision community. Saliency analysis theory can greatly improve the efficiency of the model in processing complex scenes, while deep learning is good at discovering intricate patterns of complicated scenarios. The combination of the two techniques can effectively improve the accuracy of target extraction, thereby alleviating the contradiction between high-speed acquisition and low-speed interpretation of remote sensing images. Considering the characteristics of remote sensing images, the deep learning-based saliency analysis for remote sensing images still has the following problems:

1) Fully supervised saliency analysis methods rely heavily on accurate pixel-wise annotations and large training sets, which need expert knowledge and are incredibly human-force intensive.

2) Weakly supervised saliency analysis methods can effectively reduce the dependency on label accuracy, thereby reducing the cost of human labor work. However, there is still a large gap between the performance of weakly and fully supervised methods. Moreover, most post-processing methods adopted to refine the results are usually time-consuming.

To address the issue mentioned above, we propose a weakly to fully supervised learning strategy: first, to cope with limitations of weakly supervised learning, we propose a saliency analysis method based on hierarchical weakly supervised learning, which can effectively address the problem of low accuracy brought by single scale features; second, we propose a saliency analysis method based on multiview learning and visual attention mechanism, which combines the advantages of high-level feature extraction in deep learning, and the low sensitivity of the attention mechanism to the distribution of datasets, to improve the capability in handling unbalanced dataset; then, we propose a weakly and then fully supervised saliency analysis method for target extraction, which takes advantages of the strong representation ability and fast-forwarding characteristics of the fully supervised learning method to develop a lightweight model with improved accuracy and test efficiency. Also, a weakly supervised saliency analysis method is introduced into the remote sensing image super-resolution reconstruction, which helps improve the super-resolution reconstruction of remote sensing images by taking into account the spatial quality and visual clarity.

The main work of the paper is concluded as follows:

1) We propose a saliency analysis and target extraction method for remote sensing images based on hierarchical weakly supervised learning (HWSL). The benefits of HWSL include effectively solving the problem of the insufficient generalization ability of manual feature extraction and making up for the limitations of single-scale features employment. The classification convolutional neural network (CNN) is regarded as the medium between the image-wise annotation and the pixel-wise saliency prediction. Color similarity is used as the constraint to fuse multi-scale gradient features, which are generated by computing the gradient to the middle layers of the classification CNN. By calculating the residual map of saliency maps between different classes, the salient regions get enhanced and non-salient ones suppressed.

2) We propose a saliency analysis method for target extraction based on multiview learning and attention mechanism (MLAM) for unbalanced datasets. The proposed method combines the advantages of high-level feature extraction in deep learning, and the low sensitivity of the attention mechanism to the distribution of datasets, to effectively reduce the impacts brought by unbalanced datasets. In the training phase, the weighted cross-entropy loss function is used to improve the decision-making ability for targets belonging to the minor class. In the testing phase, a multiview learning strategy is first proposed to suppress backgrounds to a great extent by computing initial saliency maps. Then, as the bottom-up attention mechanism is less influenced by the distribution of the training dataset and has stable boundary maintenance, we introduce the attention mechanism for remote sensing images, where the initial saliency maps are modified by the correction factor calculated by the color feature analysis.

3) We propose a saliency analysis method for target extraction in remote sensing images by combining fully and weakly supervised learning (CFWSL) , where the pseudo-label saliency map is used as the medium between fully and weakly supervised learning. The proposed method develops a lightweight model with improved test efficiency and accuracy. First, a weakly supervised saliency map generation method is proposed to automatically annotate unlabeled data and increase the number of training samples for the subsequent network. Then, a self-correcting feedback CNN based on fully supervised learning is constructed, where feedback connections and the side-output structure are introduced to improve the performance of the lightweight model. Finally, we evaluate the confidence of every pixel with heat maps and then combine the curriculum learning strategy and denoising loss function to reduce misjudgments of the pseudo-label saliency map. The proposed method achieves fast and accurate saliency analysis for remote sensing images with low annotation costs in a lightweight architecture.

4) We propose a super-resolution reconstruction method based on the saliency-discriminated generative adversarial network (SD-GAN) for remote sensing images. The saliency analysis theory is used to guide the unequal reconstruction of different regions with the deep learning framework to effectively suppress the spatial distortion and pseudo texture. Firstly, a saliency-driven residual generative network is designed, where the saliency map is regarded as a gate control structure and introduced to measure the texture complexity of different regions. Then, SD-GAN employs a new saliency-discriminated paired-discriminator to measures the distance between generated images and high-resolution images in salient and non-salient areas, respectively. By considering different texture complexities reflected by saliency maps, the saliency-discriminated paired-discriminator helps enhance the perceptual quality as well as preserving structural information. After adversarial learning between the generative network and discriminative network, the proposal can generate high-spatial-resolution remote sensing images with low spatial distortion and suppress the pseudo texture brought by adversarial learning.

Our work reports new ideas for saliency analysis and target extraction in remote sensing images and will provide important theoretical and technical support for vegetation estimation, crop monitoring, urban planning, and other relevant fields.