Classification Of Rice-cropping Systems Based On Pretrained Convolutional Neural Network

Rice is an important food crop in China.Due to the influence of national agricultural policy,regional economy and rural labor force,the rice cropping system in a region has changed,which in turn affects the comprehensive utilization intensity of agricultural resources,ecological environment and food security.Remote sensing has the characteristics of macro,dynamic and fast,and it is a convenient way to obtain the change of rice cropping systems at regional scale.We choose Zhuzhou,Hunan Province as the study area.Due to the cloudy and rainy weather in southern China,the lack of remote sensing data in the key growth period of rice are easy to occur,and the accuracy of traditional remote sensing monitoring in rice cropping systems is not high.In this study,a classification model of rice-cropping systems based on pretrained convolutional neural network(CNN)is presented.Combined with the idea of hierarchical classification,the information of rice-cropping systems was obtained by extracting the spatial and spectral trajectories from sentinel-2 data.The important work and conclusions are as follows:(1)Multitemporal Sentinel-2 satellite images were collected to obtain time series curves of texture(spatial trajectories)and vegetation index(spectral trajectories)as inputs of the hierarchical classification model.The first layer of the model is used for land cover classification,so as to obtain accurate crop range.The second layer mainly subdivides the crop pixel,to further extracts the accurate rice-cropping systems distribution.(2)CNN can obtain more abstract features than shallow models.The pre-training method is introduced to resolve the issue of sample shortage.Combined with the method of transfer learning,the model is pretrained by using MNIST data sets,and then the target data set composed of time series curves is used to fine-tune the parameters of the model,so as to complete the full training of the model.(3)The classification results of land cover and rice-cropping systems were acquired using the pretrained CNN,and the method achieved overall accuracy of 94.78% and 94.87%,respectively.The accuracy of rice-cropping systems using pretrained CNN was 7.11% higher than that of the SVM.The adaptability of the model was also investigated for the insufficient data in remote sensing time series images.The model obtained satisfactory performance with the overall accuracy of land cover types and rice-cropping systems of 93.52% and 93.23%,respectively.This study suggests that pretrained CNN can improve the accuracy of ricecropping system mapping by extracting deep features of spatial and spectral trajectories during rice growth cycle,and such method may provide a new idea for other remote sensing crop classification studies.