Abstract:
Objectives Convolutional neural network (CNN) has been used in spaceborne global navigation satellite system-reflectometry（GNSS-R） sea ice detection, which has the advantages of simple data preprocessing and maximum retention of reflector information. However, the data sets used in previous studies of the GNSS-R CNN sea ice detection method have a small span in time and limited representativeness, and the influence of the delay-Doppler map (DDM) ratio of seawater and sea ice in the training set on the generalization ability of the method is not considered.
Methods To solve these problems, a method of screening out malformed DDM is proposed.The appropriate CNN structure and parameters are designed, and the dataset selection strategy is optimized through comparative tests of small samples. A large sample dataset from 2018 is used to evaluate the validity and reliability of the improved method in the case of large data volume and large time.
Results and Conclusions The results show that the proposed method can screen the wrong data effectively. The CNN model has high accuracy and the best generalization ability when the DDM ratio of seawater and sea ice in the training set is 1∶1, and the improved method is still effective and reliable in large data volume and large time span. The improved method improves the generalization ability and reliability of the CNN sea ice detection method by strengthening data quality control and optimizing the dataset selection strategy, to make it more applicable to practical application scenarios and provide a reference for studies on sea ice melting.
下载:
