基于全上下文及深度特征分析的高分辨率遥感图像分类研究

Classification is one of the core tasks of very high resolution (VHR) remote sensing image processing. Semantic segmentation models for natural images can be used for the task. However, there are two problems that impede the further improvement of the precision of the models: 1) the VHR remote sensing images have much larger sizes than the natural images and contain much stronger contextual information; 2) there is a ubiquitously serious class imbalance problem in VHR remote sensing images. To address these two problems, this project provides large receptive fields by deep stacks and information abstraction, and captures global/local semantic/spatial contextual information of VHR images by modules including efficient memorial units. The redundant/irrelevant neuron inputs or weights can be screened out by quantitatively measuring the characteristics of deep features. The operation can replace the random mechanism of Dropout or DropConnect, and improve the robustness to the class imbalance problem. In the loss function, all types of contextual information are fused with adaptive weighting functions of imbalanced classes, in order to address the two problems. This study will further increase the representation and generalization ability of the semantic segmentation network, approach the good adaptability of the human brain to the two problems mentioned above, and improve the classification accuracy. The proposed methods could also be used in the semantic segmentation of other types of remote sensing images.

分类是高空间分辨率（高分辨率）遥感图像处理的核心任务之一。面向自然图像的语义分割模型可以用于该任务，但下面两个问题阻碍了模型精度的提高：1)高分辨率图像比自然图像图幅大得多，且上下文关系强得多；2)高分辨率图像广泛存在更严重的类别不平衡现象。针对这两个问题，本项目拟通过深度堆栈、信息抽象形成大感受野，再结合高效的双向记忆单元等模块，捕捉全局/局部的空间/语义上下文；通过度量深度特征性质，筛除卷积过程中冗余/不相关的神经元或权重，取代Dropout或DropConnect的随机性机制，提高对类别不平衡问题的鲁棒性；在损失函数中进一步捕捉上下文，同时融合面向类别不平衡问题的自适应加权函数，综合处理这两个问题。本研究将提高语义分割网络的描述能力和泛化能力，逼近人脑对前述两个问题所具有的良好的适应性，提升语义分割网络的分类精度，对其他类型的遥感图像的语义分割方法也有一定的借鉴意义。

基于深度学习的高分遥感图像处理和地理要素预测属于地理学的重要任务。对此，本项目开展了三个方面研究内容：基于精细化特征提取的语义分割、大幅面高分辨率遥感图像目标检测和基于地学知识增强的时序分析与预测：1) 基于高效空洞卷积，启发式或者自动调节有效感受野，有效提取多样化目标特征；再提取高层语义边界，提出超密集尺度网络，增强捕捉上下文能力。2) 建立新的大幅面图像目标检测框架，在全局局部分支高效地双向融合信息，通用性和检测效果都比对比方法更好。在遥感图像流的检测中提出要防止灾难性遗忘。另外还发现存储器中存在的对灾难性遗忘的影响巨大的新的不平衡类型，并提出针对性的方法。3) 在深度学习模型中引入并建模地学领域知识，包括近期历史数据的影响、强时序依赖性、不同气候条件下的温度降雨等的先验信息及其对径流的影响，并针对地学过程预测中的非线性问题引入机器学习领域的非线性处理技巧，融入深度学习模型，提高预测精度。