Classification Of Polarimetric SAR Images Based On Multilayer Network Model

Synthetic Aperture Radar Polarimetric is an advanced measurement system, which can obtain the 4 SAR images in different polarization channels simultaneously. Therefore, through the full polarimetric SAR image, abundant information of objects can be obtained. Image classification is an important research content of polarimetric SAR system, which is widly used in the field of agriculture and forestry planning, environmental protection and so on, so the the research on full polarimetric SAR image classification is very important for improving the application level of the SAR remote sensing.Recently, the classification methods in optical field develop rapidly, and many new models and new concepts spring up, such as Bag Of Words, Spatial Pyramid Matching, feature representation, and so on. The excellent algorithms and concepts of the optical image classification are adopted to the field of PolSAR by the researchers, and many new feature representations and classification algorithms for PolSAR are promoted accordingly and achieve good results. Hinton, et al first puted forward the concept of deep learning in 2006, which opened the research of feature learning and can automatically learn the essential features from the original image by construct a multi-layers network, so as to facilitate the classification research. Since then, the related research on deep learning was in full swing, and create many miracles in optical image classification field. This paper hopes to introduce the idea of deep learning of optical field into the PolSAR image classification, but encounters several problems:(1) The imaging mechanism of PolSAR is different from the optical image. Gray information is achieved by visible light sensor for optical image, while the echo information is recorded by microwave sensor in binary comlex model for PolSAR. In addition, speckle noise is the inherent noise of SAR due to te special imaging mechanism, which seriously affect the interpretation and subsequent application of PolSAR;(2) Polarimetric SAR can obtain four SAR images in different channels, in the case of reciprocal medium, the same object also corresponds to the 3 SAR images. However, the original deep learning models are established on the single channel data, which can not fully use the rich information of objects in the polarization SAR image;(3) Massive data is need for traing the multilayer network, while there is no enough data to training multilayer network for the target PolSAR data.In order to make foll use of the advantages of deep learning for PolSAR image classification, a bridge is need to build between the deep learning and PolSAR image. Meanwhile, multilayer network suitable for PolSAR image is need to constructed for feature learning and classification. The statistical distribution models and polarimetric decomposition theory is discribed firstly. To solve the problems when the idea of deep learning is introduced into the PolSAR, seval research works are done as follows:(1) Considering that the imaging mechanism of PolSAR image is different from optical image, deep learning cannot be used for PolSAR image classification directly. Therefore, this paper constructs a network unit based on statistical distribution, which becomes the bridge between PolSAR image and deep learning;(2) In order to make foll use of the rich object information contained by PolSAR image, two multilayer network model is constructed in different angle for feature extraction and classification. In the first multilayer network, several features are extracted and fused, and SPM is used to realize the feature representation. Then a bilayer SVM model is constructed for classification. For the second multilayer network, the multilayer deconvolutional network is introduced and improved suitable for PolSAR image classification. Meanwhile a soft probability pool method is introduced in the the deconvolutional network;(3) Because that the target PolSAR does not have enough data for the training the multilayer network, a transfer learning idea is introduced in this paper, and the data similar to the target data is used to train the multilayer network. The features achieved by multilayer network is taked as mid-representation and transfer learned by the target PolSAR image, which can classify the target PolSAR data more precisely.Based on solving the above theory and technology problems, several multilayer network models are used for the classification research of PolSAR. The experiments are conducted based on the PolSAR data of Lingshui county achieved by the 38th institute. The experiment results show that multilayer network model has great potentials in the field of PolSAR image classification.