Research On Extraction Of Rice Planting Area Based On Deep Learning And Radar Remote Sensing Data

As an extremely important global food crop,rice ensures the food source of about 50% of the world's population.Timely,objective,and accurate extraction of rice planting area has important reference value for governments to formulate food policies.Remote sensing can reflect the characteristics of different growth states of plants,and has obvious advantages in the field of agricultural research.Rice is mainly grown in areas with sufficient rainfall and cloud cover,but traditional optical satellites are difficult to achieve continuous and high-quality imaging.Therefore,the Synthetic Aperture Radar?SAR?,which is not affected by weather and other factors,has become an important detection means of the ground object in cloudy and rainy environments.Full polarimetric SAR has unique advantages in extracting ground textures,and its backscattering coefficients and target polarimetric decomposition parameters are of great significance for crop monitoring.In recent years,the achievement of deep learning in the classification of natural scenes has attracted extensive attention from scholars in the field of remote sensing.This method has outstanding ability of expression,self-learning,and fault tolerance,and can extract representative and differentiated features from remote sensing images efficiently and accurately in a hierarchical way.However,due to the limitations of image quality and dataset size,deep learning is still in the mining stage for the extraction of crops such as rice from remote sensing images,especially SAR.Given the current situation of deep learning in rice area extraction,Meishan City,Sichuan Province was taken as the research area in this paper.The potential value of automatic extraction of rice planting area from SAR image using deep learning semantic segmentation model is explored by creating rice datasets with different SAR parameter combination information.The main research contents and conclusions are as follows:?1?Seven parameters,which involve quad-polarized backscattering coefficients?VH,VV,HH,HV?and Freeman-Durden three polarimetric decomposition parameters extracted from RADARSAT-2,were arranged and combined according to the commonly used three wavebands considering the content of image information and the effectiveness of model training.The field survey data and spectral characteristics of SPOT-6 were used as auxiliary information to demarcate the rice spots on the ground.The 35 different combinations were sorted and grouped using optimum index factor?OIF?,and seven different rice label datasets based on SAR images were created and expanded.?2?In order to realize the automatic extraction and the selection of the best model of rice planting area,through the theoretical exploration of deep learning and transfer learning,the first dataset Ds₁ was used as the input to optimize the super parameter and train the semantic segmentation model of MobileUNet,BiSeNet and GCN respectively.The result showed that the three models for optical image are also suitable for SAR research.And compared with other two models,MobileUNet,which is suitable for small sample,can get more detailed information of rice,and the Precision,Recall,and MIoU are 0.964,0.962 and 0.826 respectively.?3?The remaining 6 datasets were trained as input data of MobileUNet to explore the performance of rice area extraction from different datasets.The middle feature map of the model corresponding to the combined images with different parameters was visualized and the information difference was analyzed to explore the importance of different parameters and their optimal combination to the extraction of rice planting area.The result showed that the performance of area extraction was similar the OIF arrangement trend of the dataset used in the model,and the dataset Ds₁,which contains more information of polarimetric decomposition parameters,is the best for rice area extraction.The combination of Dbl,Surf and HH is more suitable for extracting rice planting area than the images with other parameters or directly using polarization decomposition parameters and backscattering coefficients.The Precision,Recall and MIoU of the test are 0.96,0.957 and 0.823 respectively,and the backscattering coefficient HH is of great importance to rice information,which is also consistent with the theory of polarimetric SAR echo.