Polarimetric SAR Image Classification Based On Feature Selection And Deep Learning

Polarimetric synthetic aperture radar has the imaging characteristics of all-weather,multi-band,etc.,which can obtain abundant polarization scattering information by interpreting polarimetric SAR images.Polarized SAR image classification is an important research content of polarized SAR image interpretation,and has important applications in observing crop growth and monitoring natural disasters.The features required for polarimetric SAR image classification are extracted manually,which makes the classification result vulnerable to the quality of extracted features.In terms of data processing,deep learning can automatically extract the abstract intrinsic features of the target data,and these advanced features are often more capable of expressing image information than features obtained by target decomposition;In terms of image classification,deep learning acquires the deep structure information of the image through abstract learning of input samples,there by constructing a stable classifier model.Based on this,the thesis starts from the perspective of combining feature selection and deep learning,and aims to improve the classification accuracy of polarized SAR images for research,and analyzes from the following two aspects:1.The traditional feature selection algorithm generally selects some of the features with high importance as the expression features of image information.These features only emphasize the correlation between features and categories and ignore the effects of redundancy between features,resulting in a relatively high quality of the acquired features.In order to give full play to the characterization ability of different types of polarization feature combinations for image information,a polarization SAR image classification algorithm based on optimal features and convolutional neural network is given.First,the multidimensional original feature set is obtained by polarimetric SAR image data and target decomposition.Then,a fast filtering feature algorithm is used to screen the feature set for the first time,and then particle swarm optimization is used as the search strategy,and the loss function of support vector machine is used as the evaluation criterion to screen the filtered feature set for the second time to get the optimal feature set.Finally,the optimal feature set is input to CNN for training to complete the classification.Experimental results show that this method performs well in obtaining the optimal feature set,and effectively improves the accuracy of image classification.2.Because traditional feature selection algorithms have more complex and inefficient when dealing with large-scale data,and deep learning can effectively process large-scale data and mine deeper information from data,a polarimetric SAR image classification method based on depth feature is given.First,use the multi-dimensional original feature set to pre-train the Restricted Boltzmann Machine,and combine the data reconstruction characteristics of the RBM network to calculate the reconstruction error of each input feature,which is used to eliminate the original feature set.Redundant features,then initialize the RBM network with the remaining features and corresponding weights and train the deep Boltzmann machine,and finally get a set of abstract features.Finally,input the features into different classification models to get the classification results.Experimental results show that the depth feature selection method is more efficient when processing large-scale data,and at the same time improves the accuracy of image classification.The above two methods are verified by two sets of measured data collected by American AIRSAR airborne system.The results show that the two methods proposed in this thesis can improve the accuracy of image classification.