高分辨率遥感图像自适应语义分割方法研究

Semantic segmentation is an important approach for the interpretation of high resolution remote sensing images, which is highly significant for applications such as environment protection, disaster management and precision agriculture. For high resolution remote sensing images, feature extraction is challenging due to redundant details, complex spatial information of the objects is difficult to use and large variance of different object scales leads to unbalanced samples. Therefore, this project plans to research high performance feature extraction, accurate spatial information modeling and adaptive semantic segmentation framework for these challenges. 1) For homogeneous and heterogeneous regions, deep neural networks with different structures are constructed on object level and pixel level accordingly. Multi-task learning is used to optimize the network parameters to achieve high performance feature extraction. 2) The adaptive conditional random fields are proposed to model complex spatial information in high resolution remote sensing images. The structure of latent layer is adapted to image content by using local coherence and non-local similarity. The bi-penalty is designed by fusing multi-feature to penalize continuity and discontinuity adaptively. 3) The adaptive semantic segmentation framework is established on the basis of the above two researches. Different semantic segmentation methods are proposed for homogeneous and heterogeneous regions respectively. The optimization of the combined model will be proposed to coordinate deep neural network and conditional random field for better performance of semantic segmentation. The results of this project will promote the development of the interpretation of high resolution remote sensing images and its application in the related fields.

语义分割是实现高分辨率遥感图像解译的重要途径，对环境保护、防灾减灾和精准农业等应用领域具有重要意义。高分辨率遥感图像为语义分割提供丰富信息，但也造成了新的问题。高分辨率遥感图像信息冗余多造成特征提取困难，复杂地物空间信息难以利用，目标尺度差异大造成样本不平衡。因此，本项目拟从特征提取、空间信息建模和自适应框架三个方面开展语义分割研究。1）对同质和异质区域相应地设计对象和像素级特征提取深度网络，结合多任务学习实现高性能特征提取。2）从隐含层结构和二元势函数这两方面，构建遥感图像的自适应条件随机场模型，充分利用复杂空间信息提升语义分割性能。3）在前两项研究基础上构建自适应语义分割框架缓解样本不平衡性，分别对同质和异质区域提出相适应的语义分割方法，研究深度网络和条件随机场联合模型的协同优化算法，进一步提升语义分割性能。项目研究成果将促进高分辨率遥感图像解译技术的发展及其在相关领域的实际应用。

语义分割是实现高分辨率遥感图像解译的重要途径，对环境保护、防灾减灾和精准农业等应用领域具有重要意义。高分辨率遥感图像中地物具有更加精细的空间细节，为语义分割提供更丰富信息，但也带来了新的问题。丰富细节增大了地物的类内变化，复杂地物空间信息难以利用，造成地物语义特征提取困难；目标尺度差异大造成样本不平衡，跨尺度高精度语义分割难度大；目前图像分割依赖人工标注数据，而高分辨率遥感图像语义标注专业要求更高，难度更大，无标签数据的利用率严重不足。针对以上问题，本项目提出了一种基于曲率滤波器的图像金字塔特征提取方法，通过提取多尺度区域特征，减少了类内变化对语义分割的影响；提出了一种超像素条件随机场的语义分割方法，以结构自适应的超像素为基本标记单元，减少了样本分布的类别不均衡性，提升了分割精度；提出了一种遥感图像语义分割的半监督方法，在生成对抗网络框架下设计像素级伪标签生成和一致性学习算法，充分利用无标签数据进一步提高了语义分割精度。项目研究成果在城市规划、土地利用、灾害预警等领域进行了应用探索。