Research On The Semantic Segmentation Of SAR Images Combined With Transfer Learning Strategies

In the field of synthetic aperture radar images,there are fewer open data sets,image coherent speckle noise and other factors,resulting in low accuracy of some land cover classification based on synthetic aperture radar images.The application of convolutional neural networks and transfer learning techniques in the field of semantic segmentation has achieved better results.Convolutional neural networks can extract high-level features of images,and transfer learning techniques use prior knowledge of the source domain to help the training of semantic segmentation models,providing new ideas for improving the accuracy of the land cover classification application based on synthetic aperture radar images.Aiming at the problem of the lack of synthetic aperture radar image semantic segmentation dataset,a scene of Terra SAR-X dual-polarization data in the research area was purchased,and the Terra SAR-X dual-polarization scene is processed to obtain pseudo-color image data,H/A/alpha-L1 image data and H/A/alpha-wishart image data.Annotations are made by entrusting professional surveying and mapping institutions.The synthetic aperture radar dataset GDUT-Nansha with sufficient data volume is prepared.Aiming at the shortcomings of traditional convolutional neural networks in model storage and prediction speed,the redundant structure and up-sampling method of ENet are improved,and the lightweight semantic segmentation model RWL-ENet is proposed,in order to reduce the existence of synthetic aperture radar images,using the weighted cross-entropy loss function to enhance the model's segmentation effect for small-sample categories;the effect of coherent speckle noise on the semantic segmentation model is studied through experiments;it is compared with the classic semantic segmentation model,it is verified that RWL-ENet reduces model storage and prediction time under the premise of ensuring model accuracy.Based on RWL-ENet,the transfer learning technique is introduced,using the CCF optical remote sensing data set as the source domain,and the synthetic aperture radar image as the target domain,experiments are carried out from two aspects of fine-tuning strategy and hyperparameters,and it was found that the model training process oscillated less,the training time was reduced more,and the model accuracy was improved slightly with an appropriate migration learning strategy compared with the results without any transfer learning strategy.Under the premise of using appropriate fine-tuning strategies and hyperparameters,by combining the three sub-data sets of synthetic aperture radar images and the CCF optical remote sensing data set,five different transfer learning strategies are obtained to investigate the effect of the correlation between the source and target domains on the transfer learning effect.It was found that for SAR data,transfer learning between the three seed datasets in the dataset yields better results,the groups with stronger correlation between the source and target domains outperformed the other groups in terms of training time and accuracy.Finally,the transfer learning strategy with better prediction accuracy evaluation is selected and combined by model integration.Compared with the model accuracy trained by any single transfer learning strategy before integration,the prediction accuracy after integration has been greatly improved.