结合nDSM的高分辨率遥感影像深度学习分类方法  

High-resolution remote sensing image classification by combining deep learning with nDSM

作　　者：许慧敏 齐华 南轲 陈敏 XU Huimin;QI Hua;NAN Ke;CHEN Min(Faculty of Geosciences and Environmental Engineering,Southwest Jiaotong University,Chengdu 611756,China)

机构地区：[1]西南交通大学地球科学与环境工程学院

出　　处：《测绘通报》2019年第8期63-67,共5页Bulletin of Surveying and Mapping

基　　金：四川省科技厅重点研发项目(2017SZ0027)

摘　　要：针对高分辨率遥感影像因其地物类内差异大、光谱信息相对欠缺导致现有影像分类方法存在错分现象较多、地物边界残缺不完整等问题,本文提出了一种归一化数字表面模型(nDSM)约束的高分辨率遥感影像深度学习分类方法。首先,将nDSM数据作为附加波段叠加在遥感影像上并获取训练样本;然后,利用优化的U-Net网络进行模型训练得到最优模型;最后,利用最优模型对附加了nDSM波段的遥感影像进行地物分类。试验结果表明,本文方法引入nDSM数据用于U-Net模型训练和分类,可有效提高影像分类精度,得到更加真实可靠的分类结果。Although many classification methods for high-resolution remote sensing images have been proposed in recent years, there are still some problems (e.g. misclassification and incompleteness of object boundary) due to high intra-class variance and limitation of spectral information of high-resolution remote sensing images. In this paper, a high resolution remote sensing image classification method is proposed by combining nDSM (normalized digital surface model) data and deep learning framework. Firstly, nDSM data is combined with remote sensing image as an additional band to generate new imagery and produce training samples. Then, the optimized U-Net model is trained on the basis of training samples to obtain the optimal model. Finally, remote sensing images are combined with nDSM data to be the input data, and the trained optimal model is performed to get classification results. Experimental results demonstrate that the proposed method can effectively improve the classification performance in terms of classification accuracy.

关 键 词：影像分类 归一化数字表面模型 深度学习 U-Net 

分 类 号：P237[天文地球—摄影测量与遥感]