Transfer Learning For Polarimetric SAR Time Series: Methodology Research And Application

As an important remote sensing technique,polarimetric synthetic aperture radar(Pol SAR)features all-time and all-weather capabilities and can obtain the dynamic change information of earth surface,which has thus found great values in applications of natural disaster monitoring,resource investigation,urban and rural planning,and military investigation.As China enters the era of remote sensing big data,the scarcity of training samples directly limits the ability to extract information from massive Pol SAR time-series images.To address this problem,considering the characteristics of Pol SAR time-series data,two novel transfer learning algorithms were proposed,which can transfer the knowledge contained in the historical sample data.Based on the proposed novel algorithm,two information extraction flowcharts were proposed,which is more efficient and reliable than conventional methods.The main research contents and conclusions are as follows:(1)A novel relational-based transductive transfer learning method is proposed for Pol SAR time-series images,based on the relational knowledge of ground objects in images.This method uses a special time-series clustering method to find out the invariant samples on the whole time series,so that the class labels of the source domain samples can be directly transferred to target domain images across the time series.Experimental results show that the proposed method has strong robustness and effectiveness in the label transfer between Pol SAR time-series images with different imaging direction,from different sensors,and with complex object classes.(2)A novel active transfer learning(DFATL)method is proposed for Pol SAR images,which combines active learning and the deep forest model.In this method,the training set and feature space were adjusted simultaneously,to gradually reduce the marginal distribution difference between the training set and the test set,to improve the predictive performance of the model on the test set.The experimental results demonstrated that the proposed method was able to perform well with different numbers of target domain labeled samples and has good knowledge transfer capabilities in different groups of Pol SAR images from different sensors and with various distribution differences.(3)The effectiveness of water body extraction from Pol SAR images via machine learning classifiers was first evaluated,then a rapid water body extraction process for Pol SAR time-series images based on transfer learning technique was proposed.In the proposed process,the DFATL were used to transfer the class labels between images for expanding the training set.Then,the expanded training set were used to train the machine learning classifier for water body extraction.By this process,the artificial cost of collecting training samples can be reduced,while maintaining high water body extraction accuracies.This study can improve the application value of Pol SAR time-series images in flood disaster monitoring and water distribution investigation.(4)Aiming at the problem of insufficient reliability of conventional methods in the fine classification of crops from airborne Pol SAR time-series images,the characteristics of UAVSAR time-series images were first analysis,and then two novel methods for fine classification of crops were proposed.The first one is to extract more effective time-series features,then use the DFATL model as a classifier for fine crop mapping,with the incidence angle information of samples to aid the model training.The second one is using a new active pairwise constraint learning method to generate informative instance-level constraints,to constrain time-series clustering,to improve the accuracy of crop mapping.Experimental results show that the two methods can significantly improve the accuracy of fine classification of crops in UAVSAR time-series images and provide NEW solutions for crop mapping from airborne Pol SAR images.This paper not only studies new transfer learning methods for Pol SAR time-series images,but also explores new processes for the information extraction from Pol SAR time-series images,which provides theoretical and methodological support for broadening the application of Pol SAR images and improving the application level of remote sensing earth observation technology.