Research On Crop Classification Method Based On Optimal Learning Of Time-Series Polarimetric SAR Features

Polarimetric Synthetic Aperture Radar(PolSAR)is widely used for crop classification because it can provide more information on target scattering.The time-series polarimetric SAR images have the feature of reflecting the changes of various scattering characteristics of crops in different growth periods,and it can distinguish different crops by finding the scattering regularities of similar crops in different periods.However,there are still many problems in the current time-series polarimetric SAR crop classification process.At the data level,human labeling will inevitably produce noisy samples due to mislabeling;at the feature level,the temporal polarimetric features need to stack the polarimetric features of each time phase,which will lead to the problem of feature redundancy and large computation;at the classifier level,the existing classifiers do not consider temporal and polarimetric features in an integrated manner.To address the above problems,this paper conducts a research on SAR crop classification method based on optimal learning of temporal polarimetric features:(1)The proposed probability difference-based time-series polarimetric SAR noise sample classification method.For the case of noisy samples that often exist in feature classification,a classification method with anti-noise labeling capability is designed,and different Softmax-Losses are designed for two cases of known and unknown noise ratios,and the accuracy is significantly improved in the case of noisy samples in classification.(2)A redundant polarization feature optimization method based on the similarity metric is proposed.Based on the existing similarity metric,the method designs a measure applicable to the optimal selection of polarization-dimensional features,and introduces the spectral similarity metric commonly used in hyperspectral for the optimization of timedimensional features,and carries out the joint optimal selection of features from two perspectives of time-dimensional and polarization-dimensional,respectively,to achieve good classification results with fewer and effective features.(3)A SAR crop classification method based on 3D attention learning with temporal polarimetric features is designed.To address the problem of small input size of SAR data samples in the classification network,which is unfavorable to feature extraction,Vision Transformer(ViT)is introduced into the classification of time-series polarimetric SAR data for the first time;in order to broaden the time-series polarimetric SAR data extraction capability of the existing methods on time-series polarimetric SAR data,a classification network based on ViT is designed,and the traditional attention module is improved to design a three-dimensional convolutional learning method applicable to time-series polarimetric data.In order to broaden the capability of existing methods for time-series polarimetric SAR data,we designed a classification network based on Vision Transformer,and improved the traditional attention module and designed a 3D convolutional attention module for temporal data.