Landforms are an important element of natural geographical environment, and textures are the research basis for the spatial differentiation, evolution features, and analysis rules of the landform. Using the regional difference of texture to describe the spatial distribution pattern of macro landform features is helpful to the landform classification and identification. Digital elevation model (DEM) image texture, which gives full expression to texture difference, is key data source to reflect the surface features and landform classification. Following the texture analysis, landform features analysis is assistant to different landforms classification, even in landform boundary. With the increasing accuracy requirement of landform information acquisition in geomorphic thematic mapping, hierarchical landform classification has become the focus and difficulty in research. Recently, the pattern recognition represented by Convolutional Neural Network has made great achievements in landform research, whose multichannel feature fusion structure satisfies the network structure of different landform classification. In this paper, DEM image texture was taken as the data source, and gray level co-occurrence matrix was applied to extract texture measures. Owing to the similarity of similar landform and the difference of different landform in a certain scale, a comprehensive texture factor reflecting landform features was proposed, and the spatial distribution pattern of landform features was systematically analyzed. On this basis, the coupling relationship between texture and landform type was explored. Thus, the deep learning method of Convolutional Neural Network is used to train the texture features, and the second-class landform classification is carried out through softmax. The classification results in small relief and mid-relief low mountains, overall accuracy are 84.35% and 69.95% respectively, while kappa coefficient are 0.72 and 0.40 respectively, were compared to that of traditional unsupervised landform classification results, and the superiority of Convolutional Neural Network classification was verified, it approximately improved 6% in overall accuracy and 0.4 in kappa coefficient.

地貌是自然地理环境的重要组成部分，纹理是地貌空间分异、演化特征和分析规律的研究基础。利用纹理的区域差异来描述宏观地貌特征的空间分布格局，有助于地貌分类识别。数字高程模型(DEM)影像纹理充分表达了纹理差异，是反映地表特征和地貌分类的关键数据源。在纹理分析之后，地貌特征分析有助于不同地貌分类，甚至在地貌边界。随着地貌专题制图对地形信息获取精度要求的提高，分层地貌分类成为研究的重点和难点。近年来，以卷积神经网络为代表的模式识别在地貌研究方面取得了巨大的成就，其多通道特征融合结构满足不同地貌分类的网络结构。本文以DEM图像纹理为数据源，应用灰度共生矩阵提取纹理测度。针对相似地貌的相似性和不同地貌在一定尺度上的差异性，提出了反映地貌特征的综合纹理因子，系统地分析了地貌特征的空间分布格局。在此基础上，探讨了纹理与地貌类型的耦合关系。因此，采用卷积神经网络的深度学习方法训练纹理特征，通过softmax进行二类地形分类。小起伏和中起伏低山的分类结果，总体准确率分别为84.35%和69.95%，kappa系数分别为0.72和0.40，与传统无监督地貌分类结果相比，卷积神经网络的优越性分类得到验证，总体准确率提高了约 6%，kappa 系数提高了 0.4。