PolSAR Terrain Classification Based On K-shot Learning

Synthetic Aperture Radar(SAR)is a radar technology with all-weather,all-weather working mode and multi-channel,multi-parameter working characteristics.Polarographic SAR feature classification is an important task in radar image interpretation.With the rapid development of polarized SAR technology,a large amount of polarized SAR data is emerging,and the demand for rapid interpretation and classification of polarized SAR images is also increasing.K-shot learning is an important application of meta-learning.It solves the problem of extremely few samples in the test sample set.By sampling the samples in the training sample set for meta-learning and training,it can be compared on the test sample set with untrained categories.Good classification results.Using the K-shot learning method to solve the polar SAR ground feature classification can meet the needs of small samples in the rapid interpretation of polar SAR image scenes.This paper focuses on the rapid classification of polarized SAR radar images and the need to reduce the dependence of labeled samples,and proposes a K-shot learning method for polarized SAR feature classification.And proposed a feature classification method suitable for polarimetric SAR data processing.The research results of the paper are as follows:In order to solve the problem of rapid interpretation of polarized SAR data and fewer labeled samples,this paper proposes a solution strategy using K-shot learning.The traditional unsupervised polar SAR feature classification method has poor classification accuracy,and supervised algorithms generally rely on large-scale labeled data sets.Although the semisupervised method takes into account both,it requires a long time on the existing data set.training.The K-shot learning method can be trained using historical data sets,and a small amount of annotations can be directly tested on the newly acquired data set.This method can reduce labor costs and calculation time costs.Based on the above thoughts,the thesis makes an in-depth analysis of the classification of polar SAR data based on the DN4 model.The DN4 model is a K-shot learning model based on metric learning.Using the DN4 network model to train directly on the Flevoland I data set,the test results on the Flevoland II data set and the San Francisco data set prove the feasibility of using the K-shot learning method to classify the polarized SAR features.Aiming at the problem of low accuracy of polar SAR feature classification using DN4 model directly,this paper proposes a polar SAR feature classification method based on improved DN4 model.In order to solve the characteristics of high homogeneity in homogeneous regions that are common in polarized SAR data,but large differences between classes in non-homogeneous regions,this paper proposes to use a cross-entropy loss function and a Hinge Loss loss function in combination.Linear combination loss function for the number of iterations.The experimental results prove that the linear combination loss function can improve the classification accuracy on the polarized SAR data set,even exceeding the advanced polarized SAR ground feature classification method directly trained on the test set.Aiming at the problem of sensitive sample selection in the polarized SAR feature classification method based on DN4 model,this paper proposes an improved algorithm based on neighborhood minimum spanning tree sample selection.The sample selection algorithm based on the minimum spanning tree in the neighborhood can select more samples from the center of the sample category,improve the representativeness of the labeled samples in the support set,and thus improve the classification ability of the model.