Polarimetric SAR Image Classification Based On Scattering Mechanism And Multi-Channel CNN

Synthetic Aperture Radar(SAR)is an active imaging system that works in the microwave frequency band of electromagnetic waves.It can obtain highresolution,large-width microwave images all day long,all weather.Polarimetric SAR can obtain image data under electromagnetic waves in different polarization modes,and can obtain more than 100 kinds of polarization characteristic information of targets through polarization decomposition methods,which is of great significance to the image classification and other applications of target refined interpretation.The full utilization of polarization feature information is the key to the application of polarization SAR.Because deep learning has powerful feature analysis capabilities,the classification of polarization SAR images based on deep learning algorithms has become a research hotspot in recent years.The conventional deep learning-based polarimetric SAR classification method inputs various polarization features into the network without distinction for learning,which has problems such as feature redundancy,large amount of calculation and so on.Considering that polarization features are often associated with target scattering mechanisms(surface scattering,double-bounce scattering,volume scattering,and diffuse scattering),multi-channel convolutional neural networks(CNN)can extract features in different channels.This paper combines the two to study the polarimetric SAR image classification method based on the scattering mechanism multi-channel CNN.The main work and innovation are as follows:(1)A multi-channel CNN based on scattering mechanism for polarimetric SAR image classification is proposed.In view of the redundancy of polarization features and the easy interference between features,the polarization features are input into different channels of CNN according to different scattering mechanisms.For the problem that certain types of ground objects may contain multiple scattering mechanisms at the same time,the extracted features from different channels are fused.Aiming at the problem of a small number of labeled samples,a new loss function combining cross entropy and average cross entropy is constructed to reduce the risk of network overfitting.The actual data of AIRSAR Flevoland and GF-3 Hulunbuir are used to verify the effectiveness of the two improvement points one by one.When the training sample size is 0.5%of the total labeled sample,the classification accuracy is over 94.2%,and it has strong robustness.(2)A multi-channel Dilated-CNN based on scattering attention mechanism for polarimetric SAR image classification is studied.On the basis of the aforementioned research,in view of the limited receptive field of conventional CNN standard convolution,the standard convolution is replaced by dilated convolution to expand the receptive field and improve the learning ability of the network.For speckle noise and the error problem of the decomposition process,an attention model is added to every network channel,focusing on extracting the features corresponding to the scattering mechanism in each channel.The actual data of AIRSAR Flevoland and GF-3 Hulunbuir are used to verify the effectiveness of the two improvement points one by one.When the training sample size is 0.5%of the total labeled sample,the classification accuracy is over 95.3%.