基于深度学习的遥感影像目标分割是遥感影像智能解译领域的一项重要任务，在诸多领域有着广泛应用。近年来，以“深度学习+海量遥感数据+强标注”为核心的目标分割模式已取得了突破性进展。

然而，遥感影像地物信息丰富、纹理特征繁杂，因而精细标注严重依赖领域专家知识，导致标注难度高、成本大；此外，在某些特殊遥感场景还存在标注数据稀缺的问题。因此，在标注受限条件下开展遥感影像目标分割研究，即利用有限精度和规模的标注资源，完成遥感影像目标的准确、高效分割，具有重要的理论意义与实际应用价值。该研究不仅能为遥感影像智能解译提供全新视角，也将有效缓解遥感数据高速获取与低速解译间的矛盾。

尽管当前已有学者在相关领域开展研究，但由于起步较晚，仍存在诸多问题亟待解决：1）从表达能力角度看，低标注精度严重限制了目标分割模型的表达能力。2）从训练效率角度看，目前常用的多阶段训练模式的训练效率低下。3）从研究对象角度看，多类别遥感影像的类内方差大且类间可分性低，进一步限制了弱标注遥感影像目标的精确分割。4）从标注成本角度看，海量多类别遥感数据的标注成本仍需进一步降低。

针对上述问题，本文围绕弱标注和极弱标注这两种典型的标注受限情形开展遥感影像目标分割研究，目的是在标注精度较低且规模有限的情况下，有效提升模型在遥感场景下的分割性能和训练效率。论文的主要研究工作和贡献如下：

1）针对标注受限条件下算法表达能力下降的问题，提出一种基于不确定性感知与自修正学习的弱标注遥感影像目标分割方法。该方法从提升标注精度的角度入手，通过逐级优化标签质量，促进模型性能提升。首先，提出基于变分自注意力生成对抗模型的样本扩张策略，有效提升初始图像级弱标注的类间分布均衡性。其次，设计置信度加权的互补擦除机制生成语义完整的伪标签，实现标注精度从图像级到像素级的有效过渡。最后，提出不确定性感知的联合优化策略，逐步减轻含噪伪标签对分割模型训练带来的不利影响。实验结果表明，该方法能够实现标注成本和模型性能之间的有效平衡。

2）针对标注受限条件下算法训练效率低下的问题，提出一种基于联合互补学习机制的单阶段弱标注遥感影像目标分割方法。该方法从互补网络的联合优化角度入手，将多阶段算法中的目标定位和分割网络整合为统一框架，在保证算法分割精度的条件下有效提升训练效率。首先，设计了一种置信度引导的伪标签修正模块，将互补子网络学习到的目标位置线索和目标边界轮廓信息充分结合，引导互补子网络进行交互学习，生成目标定位准确、边界清晰的伪标签。其次，提出一种双重自监督多尺度一致性损失，缓解网络对伪标签噪声的过拟合，并有效提升网络在分割多尺度遥感目标时的鲁棒性。实验结果表明，该方法能够仅通过端到端的训练高效实现弱标注条件下遥感目标的准确分割。

3）针对多类别遥感影像类内方差大且类间可分性低，导致弱标注条件下目标分割性能严重受限的问题，提出一种基于多层级对比学习的弱标注遥感影像多类别目标分割方法。该方法从表征对比学习的角度入手，分别从图像间、区域间以及像素间充分挖掘层次化语义关联信息，从而在标注受限情况下充分提升模型对多类别地物目标的识别能力。首先，提出动态掩码引导的多原型对比学习策略，缓解区域级对比学习中含噪伪标签导致的嵌入空间混乱问题，同时有效捕获遥感影像的类内差异。其次，分别设计置信度重加权策略和基于特征相似度的困难负样本采样策略，提高像素级和图像级表征的鲁棒性和鉴别能力。实验结果表明，该方法能够在弱标注条件下实现对多类别遥感影像重要地物目标的准确分割。

4）针对多类别遥感影像标注精度与标注规模同时受限的极弱标注情况，提出一种基于伪监督再学习的极弱标注遥感影像多类别目标分割方法。该方法通过伪监督再学习充分挖掘极弱标注中大量无标注样本中的地物语义信息，并与少量真实类别信息结合，从而促进模型分割精度的有效提升。首先，设计类感知的数据集重划分策略，从基于少量弱标注生成的含噪伪监督中自适应筛选高质量伪标签，缓解标注数量有限带来的不利影响。其次，构造动态阈值引导的再学习网络，在进一步挖掘伪监督语义信息的同时，减少无标注样本的伪标签带来的置信偏差影响。实验结果表明，该方法能够在极弱标注条件下，取得与基于充足弱标注的方法相当的优越性能，实现多类别遥感影像中地物目标的准确分割。

本文开展的上述研究工作不仅能够为遥感影像目标分割开辟新的研究思路，推动高分辨率遥感影像的智能解译应用，而且促进了深度学习技术在遥感领域的应用和发展，可为城市规划建设、土地利用、自然资源监测等领域提供重要的技术支持。

As a significant task in the field of intelligent interpretation of remote sensing images (RSIs), deep learning-based RSI object segmentation has been widely applied in various fields. At present, the object segmentation mode with the core of “deep learning + massive remote sensing data + strong annotation” has made breakthrough progress.

However, due to the extensive ground object information and complex texture features of RSIs, pixel-level labeling is highly dependent on a considerable amount of domain expertise knowledge, which makes labeling challenging and expensive. In addition, in some special remote sensing scenarios, annotation data are scarce. Therefore, it is of great theoretical significance and practical application value to carry out the research on remote sensing image object segmentation under the conditions of limited annotations, i.e., to achieve accurate and efficient object segmentation for remote sensing images by utilizing the annotation resources with limited accuracy and scale. The research can not only provide a new perspective for intelligent interpretation of RSIs, but also effectively alleviate the contradiction between high-speed acquisition and low-speed interpretation of remote sensing data.

Although scholars have carried out research in related fields, due to the late start, there are still many problems to be solved: 1) From the perspective of representation ability, low labeling accuracy severely limits the performance of the object segmentation model. 2) From the perspective of training efficiency, the training efficiency of the commonly used multi-stage training mode is low. 3) From the perspective of study objects, the large intra-class variance and low inter-class separability of multi-class RSIs seriously affect the accurate segmentation of weakly labeled RSIs. 4) From the perspective of labeling cost, the labeling cost of massive and multi-class remote sensing data still needs to be further reduced.

To address the aforementioned issues, this work investigates object segmentation for RSIs under two typical scenarios of limited annotations, including weak labeling and extremely weak labeling. This study aims to effectively improve the segmentation performance and training efficiency of the model in remote sensing scenarios with limited and low labeling accuracy. The main research work and contributions of this dissertation are as follows:

1) Aiming at the problem that the expression ability of the algorithm decreases under the condition of limited annotations, we propose an RSI object segmentation algorithm based on an uncertainty-aware self-refinement learning framework. The algorithm starts from the perspective of labeling accuracy, and improves the model performance by optimizing the label quality step by step. Firstly, a data augmentation strategy based on variational autoencoder-self attention-wise generative adversarial network is proposed, to effectively improve the intraclass distribution balance of initial image-level category labels. Secondly, a confidence-weighted complementary erasing mechanism is designed to realize the effective label transition from the image level to the pixel level. Finally, we propose an uncertainty-aware joint optimization strategy to gradually alleviate the negative impact on the training of the segmentation model caused by noisy pseudo labels (PLs). Experimental results reveal that the proposed method can achieve a better tradeoff between labeling cost and segmentation accuracy.

2) Aiming at the problem of low training efficiency of the algorithm under the condition of limited annotations, we propose a single-stage RSI object segmentation algorithm based on confidence-guided joint complementary learning. The algorithm starts from the perspective of jointly optimizing the complementary networks, integrates the object localization network and object segmentation network from the multi-stage frameworks into a unified framework, and effectively improves the training efficiency under the condition of guaranteeing the segmentation accuracy of the algorithm. Firstly, we design a confidence-guided PL refinement module, to fully incorporate the object localization clues and object boundary details learned from each subnetwork together. Thus, it can guide the two complementary subnetworks to interactive learning, and generating high-quality PLs with accurate object locations and clear boundaries. Secondly, a dual self-supervised multiscale consistency loss is proposed to avoid overfitting on the incorrect pixels in noisy PLs and promote the model robustness on multiscale geographic objects. Experimental results demonstrate that the proposed method can effectively achieve the accurate segmentation of weakly labeled remote sensing objects only through end-to-end training.

3) Aiming at the problems of degraded segmentation performance caused by large intra-class variance and low inter-class separability of multiclass RSIs, we propose a multiclass object segmentation model based on multi-level contrastive learning for weakly labeled RSIs. The algorithm starts from the perspective of representation contrastive learning. Through fully excavating hierarchical semantic association information between images, regions, and pixels respectively, the model's recognition ability of multi-class geographic objects under the condition of limited annotations can be fully improved. Firstly, we propose a multi-prototype contrastive learning strategy guided by dynamic mask generation to alleviate the confusion of embedded feature space caused by pixel-level noisy PLs, and to facilitate the capture of intra-class discrepancy of multi-class RSIs. Secondly, a confidence reweighting strategy and a hard-negative samples sampling strategy based on feature similarity are designed respectively, to improve the robustness and discrimination of pixel-level and image-level representations. Experimental results validate that the proposed method can achieve accurate object segmentation from multi-class RSIs.

4) Aiming at the situation that the annotation accuracy and scale of multi-class RSIs are limited at the same time, an object segmentation algorithm based on pseudo supervision re-learning under the condition of extremely weak annotation is proposed. The proposed method fully mines the semantic information of geographic objects from a large number of unlabeled samples through the pseudo supervision re-learning mechanism, and combines it with a small amount of true category information to promote the progressive improvement of model segmentation accuracy. Firstly, we design a category-aware dataset re-split strategy to adaptively select high-quality PLs from noisy pseudo supervision, which can mitigate the adverse effects caused by the absence of large amounts of real category labels under extremely weak labeling conditions. Besides, we construct a re-learning network guided by a dynamic thresholding strategy, which helps further excavate the implicit semantic clues in pseudo-supervision, while also alleviating confirmation bias caused by the PLs. Experiments show that the proposed method achieves competitive performance under the condition of extremely weak supervision compared to methods that utilize complete image-level annotations, and can achieve accurate object segmentation from multi-class RSIs.

The above research work carried out in this dissertation can not only open up new research ideas for the object segmentation of RSIs, and promote the intelligent interpretation and application of high-resolution RSIs, but also promote the application and development of deep learning technology in the field of remote sensing. This dissertation can provide important technical support for urban planning and construction, land use, natural resources monitoring and other fields.