近些年来，合成孔径雷达(Synthetic Aperture Radar，SAR)技术蓬勃发展，多颗SAR遥感卫星成功发射。GF-3作为我国首颗C波段SAR卫星，在地球科学、气候变化研究、森林资源调查等领域发挥重要作用，使我国的遥感应用进入新的阶段。然而，如何从海量遥感数据中快速准确地提取地物信息，是遥感工作者亟待解决的问题。目前，在大数据的支持下，以深度学习为代表的人工智能技术掀起了巨大的浪潮，其在多个应用中展示的能力远远超过了传统的图像处理算法。本文将深度学习应用在GF-3极化SAR地物类型分类的工作中。采用具有代表性的深度卷积Highway Unit神经网络。该网络通过交替的卷积层和池化层从SAR图像中自动学习多层高级特征表达，能够很好的抵抗噪声的影响，充分挖影像在时间和空间上的规律，有效提高地物分类精度。 本文的研究内容包括以下三个方面： (1)本文采用Schmidhuber教授提出的深度卷积Highway Unit神经网络，通过可反复堆叠的Highway网络学习单元，使其优化方法基本与网络的深度独立，某种程度解决了网络信息损耗的问题，可以用较少的样本训练任意深度的网络。 (2)为了评价深度卷积Highway Unit神经网络的分类效果，本文以依根农林交错区为研究区，采用GF-3全极化SAR影像进行实验。该模型分类结果的平均精度达到88.12%，与传统的机器学习（如支持向量机和随机森林）相比，精度分别提高了29.09%和11.30%。其次，将预训练好的模型对不同时间、不同入射角、不同传感器获取的影像进行分类，以说明该模型在不同条件下的泛化学习能力，为第四章的大区域制图提供新的思路和技术支持。 (3)基于深度卷积Highway Unit神经网络的方法，充分发挥深度学习处理海量数据的能力，采用GF-3双极化SAR数据实现呼伦贝尔市大区域制图，为处理海量SAR数据提供良好的范例。

In recent years, with the rapid development of synthetic aperture radar (SAR) technology and satellites were launched, SAR data have ushered in the era of big data. As Chinese first C-band SAR sensor, GF-3 plays an important role in the earth science, climate change research, forest resources survey and other fields. SAR images interpretation under the era of big SAR data is a great challenge for scientific applications. At present, big data-based intelligent methods such as computer vision technology have achieved great success. Deep learning such as deep convolutional Highway Unit network has revolutionized the computer vision area. Deep learning-based algorithms have surpassed conventional algorithms in terms of performance by a significant margin. In this paper, deep learning will be applied to the work of GF-3 quad-pol SAR imagery classification. A deep convolutional Highway Unit network is employed to automatically extract a hierarchic feature representation from the data, based on which the land cover type classification can be conducted. The deep convolutional Highway Unit network by the SAR images, the proposed approach in the paper can reduce speckle, fully excavate the regularity of SAR images in time and space and effectively improve the accuracy of classification. Our experimental results, conducted on a commonly used GF-3 polarimetric SAR data, show that the proposed method provides classification results that are among the state-of-the-art. The research contents of this paper include the following three aspects： (1)The traditional convolutional neural network has become more difficult to train with the increase of network depth. In view of this, this paper improves the network structure. The construction of a deep convolution Highway Unit network by using the Highway learning Unit that can be repeatedly stacked，a way to solve the problems of the network information loss, which can train any deep network with fewer samples. (2)In order to evaluate the classification effect of the deep convolutional Highway Unit network, this paper uses the GF-3 quad-pol SAR image to conduct the experiment in the Yigen study area. Compared with traditional machine learning(Support vector machine and Random Forest), the average accuracy of the model classification results is 88.12%, and the classification accuracy of forest, grass and farm is over 90%. In addition, the pre-trained model is used to classify different time, different incidence angle and different sensors, so as to show that the generalization learning ability of the model under different conditions, providing new ideas and technical support for the fourth chapter of large-scale mapping. (3)A method based on deep convolution Highway Unit network to give full play to the ability of deep learning to deal with large data.This paper uses GF-3 dual-pol SAR data to realize large-scale mapping in Hulunbeier, and provides a good example for dealing with massive SAR data.