Research On PolSAR Landcover Classification Based On Deep Learning

With the successive use of high-resolution Polarimetric Synthetic Aperture Radar(PolSAR)satellites,PolSAR technology has gradually become one of the research hotspots in the field of remote sensing.In particular,the launch of GF-3 is to make up for the gap in the field of high-resolution radar in China.PolSAR land classification is an important technique for SAR image interpretation.It has important research and application value in military,agriculture,forestry,environmental protection,hydrological monitoring,urban planning and geological exploration.PolSAR feature classification algorithms based on traditional image classification methods usually rely on the selection of specific features.And their process usually relies on strict mathematical calculations.Therefore,the determination of parameters such as classification thresholds depends on certain prior knowledge.At the same time,traditional classification methods usually use pixels as the basic unit,which makes them vulnerable to speckle noise.Deep learning methods can abstract higher-dimensional features from raw data and make full use of two-dimensional spatial information.Therefore,in this paper,we have carried out research on the classification of polarimetric SAR features based on deep learning.The main research contents and innovations are as follows:1.A PolSAR land features classification method that combines multi-scale segmentation and RBF is explored.The application of RBF in the field of PolSAR land classification is studied.In order to eliminate the inherent boundary features of SAR images and the irregular boundary conditions of real ground objects,the pixel-level classification based on RBF combined segmentation objects is explored.This method effectively reduces the generation of broken isolated points and effectively improves the classification accuracy.2.A PolSAR feature classification method based on multi-scale GoogLe Net convolutional neural network is proposed.Considering that the pixel-level classification process does not take into account the two-dimensional spatial information and texture features of SAR images,we have introduced multi-scale CNN for land classification after analyzing different scales of different ground types in real cover.The method considers both the real object boundary and the two-dimensional spatial information of the polarimetric SAR image,which effectively improves the integrity and classification accuracy of the classification result.3.In order to avoid the inconsistent regional classification accuracy caused by multi-scale segmentation,a sample set is constructed using the superpixel segmentation method.Then train Alex Net and classify the experimental area.This method takes the superpixel segmentation object as a sample,and chooses a slightly shallower Alex Net convolutional neural network to adapt to the small-scale sample,which improves the classification accuracy,simplifies the classification process,and increases the universality.