基于深度特征提取和匹配的多源时空遥感影像序列变化检测研究

Image change detection has become an important issue in the field of image processing due to its wide applications in remote sensing research, agricultural survey, disaster relief and so on. Images from multi-source space-time remote sensing image sequences are quite different in their properties due to imaging approaches and environment, and hence their change detection is of more practical significance and more challenging. This project plans to employ deep neural networks as basic tools, exploiting a flow chart of "feature extraction and feature mapping" to detect multiple changes from image sequences. In the process of feature extraction, the super-pixel segmentation and similarity measurement are first employed, thus reconstructing super-pixels by utilizing local and non-local spatial information as the inputs of the networks. Then the fuzzy number is added in the traditional autoencoder to ameliorate its performance in deep learning, thus extracting robust features. In the process of feature mapping, the hierarchical strategy is adopted to obtain representative samples. Then the deep convolutional neural network with orthorhombic structure is employed to map the features, constructing the corresponding mapping function with speed and flexibility. The previous work and initial experimental results have demonstrated the feasibility and effectiveness of the techniques in the project.

影像的变化检测技术在遥感研究、农业调查、抗震救灾等方面有着广泛的应用，目前已成为遥感领域重要课题之一。多源时空遥感影像序列因成像方式和成像客观条件的不同使影像间的诸多性质存在差异，其变化检测也因此更具有现实意义和挑战性。本项目拟使用深度神经网络作为基本工具，开发 “特征提取-特征匹配”这一框架来检测出影像序列中存在的诸多种类的变化。特征提取方面，本项目先采用超像素分割和相关性度量，利用局部和非局部的空域信息重构超像素作为深度网络的输入；然后利用模糊数改进传统的自编码器，使得深度学习中网络性能得以优化，从而提取出更加鲁棒的特征。特征匹配方面，本项目首先使用分层分类策略，以无监督的分层方式在匀质区域内获取具有代表性的样本特征；然后使用输入输出正交结构的深度卷积神经网络匹配特征，快速、灵活地建立诸类别的网络映射函数。先前的工作基础和初步的实验结果表明了该项目中诸方案的可行性和有效性。

多源时空遥感影像序列的变化检测中，深度神经网络起着关键的作用，最近流行的集成神经网络能够有效检测出诸类变化的区域，有利于农业土地调查、城市规划发展等现今存在的问题。本项目基于深度神经网络进行该课题研究，通过对现有语义分割深度网络的学习和理解，研发出带有曲折传播的网络，有效地将变化区域和非变化区域分开；同时为了更精确地检测出相关变化，还进一步在多尺度的网络上增加了注意力机制模块，提高了变化检测在指定区域的精度。与此同时为了进一步检测出多种不同的变化，从传统的“先特征后联合”与“先联合后特征”两种框架出发，构建了一种融合的网络框架结构，并在网络中使用了性能优良的特征提取和特征分类模块，形成一套较为完整的变化检测体系，能够较为快速、有效地在遥感影像中检测出诸类别的变化。