Landsat图像是获取大尺度地表覆盖信息的重要数据源，深度学习用于高分辨率图像分类效果良好，但对30m分辨率Landsat图像的大尺度地表覆盖分类精度不甚理想。项目将在现有深度学习模型基础上，可视化不同网络层次Landsat图像特征提取结果，分析其特征表达规律，构建有效突显Landsat图像特征的样本库；定义深入挖掘Landsat图像类内纹理结构特征的卷积操作，探究不同网络结构对地物类间差异特征的表达优势，建立面向Landsat图像特征的自适应卷积神经网络模型；考虑不同地理分区地表覆盖类型的差异性，研究不同分区训练样本的采样方式，结合迁移学习理论，提出采用少量新增样本将基准分区网络向目标分区迁移的方法，实现不同分区Landsat图像地表覆盖的精确分类，形成基于Landsat图像特征的深度卷积神经网络智能化地表覆盖分类方法，为获取大尺度、长时序地表覆盖分类信息提供科学可行的理论和技术支撑。

Landsat image is one of the most important data source for obtaining large scale land cover information. Accuracy of deep learning on high resolution images are high, while that on Landsat image with 30m resolution is unsatisfactory. Based on existing deep learning models, this project will visualize the feature extraction results of different layers from convolutional neural networks and analysis the variation of the characteristics. Then a sample library, in which the features of Landsat images are highlighted, is established. Considering the study ability of deep learning method on Landsat image, a convolutional operation only studying the characteristics of textures inside a class, is defined. Along with the expression advantage of deep learning on Landsat image, a self-adaptive convolutional neural network is established. As the difference between land cover types is the major problem hindering obtaining high resolution from large scale areas, this research divides a large area into small areas with similar land cover types and studies the sampling method on different areas. Combined with transfer learning theory, then the well-trained network in an area can be transferred to other areas with little samples. Finally, large scale land cover classification based on Landsat images can be achieved according to the above-mentioned methods. This research can be used in the production of large scale land cover map based on Landsat images.