基于双层组合神经网络的土地利用自动分类方法  

An Automatic Land Use Classification Method Based on Double-Layer Combinatorial Neural Network

作　　者：曾鹏 喻宏伟 余雅滢 刘明勇 郑果[3] 黄亮 ZENG Peng;YU Hongwei;YU Yaying;LIU Mingyong;ZHENG Guo;HUANG Liang(Hunan Institute of Land and Resources Planning,Changsha 410007,China;Faculty of Artificial Intelligence in Education,Central China Normal University,Wuhan 430079,China;Hunan First Normal University,Changsha 410205,China;China Academy of Civil Aviation Science and Technology,Beijing 100028,China)

机构地区：[1]湖南省国土资源规划院,湖南长沙410007 [2]华中师范大学人工智能教育学部,湖北武汉430079 [3]湖南第一师范学院,湖南长沙410205 [4]中国民航科学技术研究院,北京100028

出　　处：《测绘地理信息》2023年第5期98-103,共6页Journal of Geomatics

基　　金：国家自然科学基金(62177017,41671377);中国民航科学技术研究院基础科研业务费项目。

摘　　要：当前土地利用分类样本标签的时效性和样本内部特征对象复合性、分类间界限模糊性和土地利用类型的不平衡性使得国土资源调查遥感影像土地利用分类自动化难度大、精度低。因此,提出一种双层组合神经网络用于土地利用自动分类。首先,利用城市道路交通网络矢量数据裁切影像,生成并标记具有明确土地利用类别边界的标签样本,构建基于第三次全国国土调查土地利用标准的城市遥感影像数据集;然后,组合U-Net的跳跃连接和DenseASPP的密集空洞空间金字塔池化结构,对多尺度上下文信息和抽象空间信息进行聚合,消除类不平衡现象,提升自动分类精度。在自建数据集和ISPRS Vaihingen公开数据集上进行实验,将所提方法与U-Net、DenseASPP和Deeplabv3三种典型深度卷积神经网络方法进行比较。在两种数据集上,所提方法的分类总体准确率分别为75.90%和89.63%;F1值分别为74.68%和86.56%,平均交并比(mean intersection over union,mIoU)分别为72.21%和82.23%,其语义分类结果皆优于所比较的方法。总的来说,所提方法提升了土地利用机器学习分类精度。At present,the timeliness of the labels of land use classification samples,the complexity of feature objects within samples,the fuzziness of classification boundaries and the imbalance of land use types lead to the difficulty and low accuracy of automation of land use classification of remote sensing images from national land resource survey.So,we propose a double-layer combinatorial neural network for automatic land use classification.First,we use the vector data of urban road network to crop images,generate and tag the labels with a clear boundary for land-use classes,thus construct an urban remote sensing image dataset under the land-use standard of the 3rd National Land Resource Survey.Then,we combine the skip connections of U-Net and the dense atrous spatial pyramid pooling structure of DenseASPP to aggregate the multi-scale contextual information and the abstract spatial information,in order to eliminate the imbalance of classes and improve the accuracy of automatic classification.The experiments are conducted on the self-constructed dataset and the public dataset named ISPRS Vaihingen,and the proposed method is compared with 3 typical deep convolutional neural networks(U-Net,Deeplabv3,and DenseASPP).On the two data sets,the overall classification accuracies,F1 values,and mean intersection over union(mIoU)values of the proposed method are 75.90%and 89.63%,74.68%and 86.56%,72.21%and 82.23%,respectively,which confirms its better semantic categorization results.In general,the proposed method improves the accuracy of land use classification with machine learning.

关 键 词：城市土地利用分类 语义分割 遥感 深度卷积神经网络 

分 类 号：P237[天文地球—摄影测量与遥感]