关键词: 土地利用/土地覆被分类, Sentinel-1卫星, 深度残差神经网络, 迁移学习, 深度学习

Abstract:

Land use/land cover classification of small sized datasets cannot be accurately and automatically classified by deep learning method.In order to solve this problem,a land ues/land cover automatic classification algorithm applying the reduced-modified deep residual neural network(Resnet-50) for the transfer learning is proposed.Firstly,use the original remote sensing data provided by Sentinel-1 satellite to create dataset.Secondly,compress convolution template numbers of per convolution layer in Resnet-50 and cascade the adaptive networks behind to the obtained reduced residual network.Finally,pre-training the network with ImageNet dataset and transferthe trained model to the Sentinel-1 dataset to fine-tune the parameter of the network.And the high-precision automatic classification of land use/land cover employing deep learning on small datasets is finally achieved.Experimental results show that the classification accuracy of this algorithm on SAR data set is 95.15%,which verifies the feasibility of the proposedalgorithm.

Key words: land use/land cover classification, Sentinel-1 satellite, deep residual neural network, transfer learning, deep learning