Extraction of Rice-planted Area based on MobileUnet Model and Radarsat-2 Data

Extracting rice-planted area based on Synthetic aperture radar (SAR) image can effectively solve the cloud pollution problems of optical images. Compared with the traditional image classification method, deep learning method based on semantic segmentation algorithm can effectively utilize the information that can represent the image features and backscatter difference between different ground objects in remote sensing data, which satisfies the need of rapid and accurate extraction of ground information in the field. In this paper, the extraction of rice-planted area was implemented using MobileUnet model and rice label datasets were created using Radarsat-2 quad-polarized backscattering coefficient images and Freeman-Durden decomposition images obtained during the rice growth period in 2016. The results of our experiment showed that the Precision, Recall and MIoU are 0.964, 0.962 and 0.826 respectively. We analyzed the results and proposed future research prospects finally.