一种针对小样本的高分辨率遥感影像道路提取方法  

A road extraction method for high resolution remote sensing image with small samples

作　　者：李朝奎 方军[1,2] 吴馁 宋璟毓[3] 周倩 周青蓝 LI Chaokui;FANG Jun;WU Nei;SONG Jingyu;ZHOU Qian;ZHOU Qinglan(National-Local Joint Engineering Laboratory of Geo-Spatial Information Technology,Hunan University of Science and Technology,Xiangtan,Hunan 411201,China;Hunan Province Engineering Laboratory of Geospatial Information,Hunan University of Science and Technology,Xiangtan,Hunan 411201,China;China State Shipbuilding System Engineering Research Institute,Beijing 100094,China)

机构地区：[1]湖南科技大学地理空间信息技术国家地方联合工程实验室,湖南湘潭411201 [2]湖南科技大学地理空间信息湖南省工程实验室,湖南湘潭411201 [3]中国船舶工业系统工程研究院,北京100094

出　　处：《测绘科学》2020年第4期81-88,共8页Science of Surveying and Mapping

基　　金：国家重点研发计划项目课题(2017YFB0503802);国家自然科学基金项目(41571374)。

摘　　要：针对高分辨率遥感影像道路提取过程中,深度学习方法较传统提取方法虽可有效地提高地物提取的精度,但需要大量样本训练,消耗较多计算资源,且大量高质量训练样本难以获得等问题,该文提出了一种将支持向量机(SVM)与卷积神经网络(CNN)结合的适用于小样本的道路提取混合模型,该模型能够在小样本训练时保证道路提取的精度。采用数据增强、正则化等方法优化训练策略,丰富小样本道路特征库,设计结合SVM的深度卷积神经网络结构来提取道路在影像中的高维特征,降低模型计算量,减少计算时间。以谷歌高分辨率遥感影像作为实验数据,用不同训练样本量来训练模型并验证道路提取的精度;同时,将该文提出的方法与逻辑回归(LR)模型、光谱结合SVM模型以及VGG16深度学习模型进行了道路提取效果的对比分析。结果表明:该文倡导模型方法在小样本情况下可以提取较高精度道路;与其他3种方法比较,该文方法能够快速构建与训练模型,在满足精度要求的同时,极大地提高了高分辨率遥感影像道路提取的效率,为道路数据的快速更新与变化检测提供了新的技术支持。Aiming at the problems of road extraction from high-resolution remote sensing images,the deep learning method can effectively improve the accuracy of ground object extraction compared with the traditional extraction method,but it requires a lot of training samples,consumes more computing resources,and a large number of high-quality training samples are difficult to obtain. A road extraction hybrid model suitable for small samples combining support vector machine(SVM) and convolutional neural network(CNN) had been proposed,which can ensure the accuracy of road extraction with limited training samples. The data augmentation and regularization methods were used to optimize the training strategy,and enrich the small sample road feature database. Besides,a deep convolutional neural network structure combined with SVM to extract the high-dimensional features of the road in the image was designed,and it reduced the calculation amount of the model and the calculation time. Using Google’s high-resolution remote sensing image as experimental data,the model was trained with different training sample sizes and the accuracy of road extraction was verified. At the same time,the proposed method was compared with some road extraction methods for comparative analysis of road extraction effects,including logistic regression(LR) model,spectral combined with SVM model and VGG16 deep learning model. The results showed that the proposed model method can extract higher precision roads under small sample conditions.Compared with the other three methods,this method can quickly construct and train the model,and greatly improve the high-resolution remote sensing while satisfying the accuracy requirements.The efficiency of image road extraction provides technical support for rapid update and change detection of road data.

关 键 词：深度卷积神经网络 SVM 小样本 高分辨率遥感影像 道路提取 

分 类 号：P237[天文地球—摄影测量与遥感]