摘要
本文提出结合深度卷积神经网络与在线高分遥感影像的分类方法,用于GlobeLand30地表覆盖产品的质量优化。首先,通过对多源地表覆盖产品的一致性分析,构建深度学习训练所需的高分辨率遥感大样本(224万样本量);其次,基于该大规模样本集训练适用于GlobeLand30优化的深度卷积神经网络模型(GoogleNet Inception V3);最后,利用训练好的神经网络模型对在线高分影像进行分类,用以优化GlobeLand30产品的不可靠区域。经独立测试样本集验证,经过训练的神经网络分类总体精度为87.7%,Kappa系数为0.86,相比原始GlobeLand30的精度(总体精度75.1%、Kappa系数0.71)有了明显提升。在4个试验区的GlobeLand 30产品优化实验表明:该方法能够有效优化GlobeLand30产品的分类精度。
GlobeLand30,as one of the best Globe Land Cover(GLC)product at 30 m resolution,was developed by China based on the integration of pixel-and object-based methods with knowledge(POK-based approach),which combines multiple levels of classification techniques to achieve high-accuracy land cover mapping. In particular,a knowledge-based verification process was employed to refine and grantee the product quality of Globeland30 by manual interpretation of online high-resolution images. However,the manual intervention suffers from large labor consumptions and the subjectivity influence. Considering the great achievements of deep learning in image recognition and classification,classifying online high-resolution remote sensing images with Deep Convolutional Neural Network(DCNN)may improve the efficiency and performance of the refinement procedure for GlobeLand30. However,the training of DCNN relies on a large number of training samples;and the existing remote sensing sample sets cannot satisfy the training requirements in terms of sample size and category system. Therefore,a method for generating large sample set of high-resolution remote sensing imagery based on multi-source GLC was proposed;and a large sample set with 2.24 million samples is automatically generated by this method. The DCNN model(GoogleNet Inception V3)was trained from scratch with the proposed large sample set and then used to refine Globeland30 product. Verification with an independent test sample set shows that the proposed trained DCNN model can achieve higher classification accuracy(Overall accuracy:87.7%,Kappa:0.856)than that of original GlobeLand30 product(Overall Accuracy:75.1%,Kappa:0.71). Finally,four test areas were selected for evaluating the performance of proposed refinement procedure.The results show that the GoogleNet(InceptionV3)model trained by the proposed large sample set can effectively refine the product quality of GlobeLand30.
作者
刘天福
陈学泓
董琪
曹鑫
陈晋
Liu Tianfu;Chen Xuehong;Dong Qi;Cao Xin;Chen Jin(State Key Laboratory of Earth Surface Processes and Resource Ecology,Faculty of Geographical Science, Beijing Normal University,Beijing 100875,China;Beijing Engineering Research Center for Globe Land Remote Sensing Products,Faculty of Geographical Science,Beijing Normal University,Beijing 100875,China)
出处
《遥感技术与应用》
CSCD
北大核心
2019年第4期685-693,共9页
Remote Sensing Technology and Application
基金
国家自然科学基金项目(41871224)