Research On POLSAR Image Classification Based On Multitask Learning

Polarimetric Synthetic Aperture Radar(POLSAR)is one of the most advanced and important monitoring technologies,which has the characteristics of all-weather and all-time observation.POLSAR image classification is one of the important and difficult tasks in SAR image processing.Its purpose is to classify each pixel into a specific category.POLSAR image classification is a key step in understanding SAR data,and has been widely used in military,civil and other fields.However,due to the complexity and diversity of natural terrain,POLSAR image classification is still a challenging task.Polarization SAR image classification and image processing is an important research direction,which is a hot and difficult research topic at present.For POLSAR image classification,two algorithms are proposed in this paper.In the third chapter,a POLSAR image classification method based on multi-task sparse representation learning is proposed.In order to improve the computational efficiency and spatial consistency of classification.Firstly,the polarimetric SAR image is segmented into super-pixels,and then different types of features are extracted to form a feature dictionary of each class.Then the target polarimetric SAR super-pixel feature vectors are classified into specific classes by multi-task sparse representation learning method.For the multi-task sparse representation learning model adopted in this paper,an accelerated proximal gradient method is adopted for numerical solution.In the fourth chapter,a polarimetric SAR image classification method based on low-rank multitask learning is proposed.First,the image is segmented into super-pixels,the training sample and the test sample are segmented into many small pieces,and then the features extracted from the super pixels of different category tags are formed into the feature dictionary of each category,and the super-pixels of the test image are divided into specific classes by the multi-task learning model based on robust low-rank representation.Experiments show that the proposed multi-task learning method based on robust low-rank representation is superior to the classification method based on a single feature.