High Resolution SAR Image Classification Based On Statistical Analysis And Improved CNN

Synthetic aperture radar(SAR)can observe ground information all-time and all-weather,which has important research value and very broad application prospects.Because of its special advantages,SAR is widely used in national defense fields such as military reconnaissance.At the same time,it also has important application value in the fields of national economy,such as map mapping,natural disaster prevention,crop classification,urban planning and so on.SAR image classification is to assign a label to each pixel in the scene through the corresponding algorithm.As a key part of SAR image interpretation,it has important practical value.However,with the development of technology,the resolution of SAR images is getting higher and higher,which poses new challenges for the classification of SAR images.For high-resolution SAR systems,the continuous optimization of spatial resolution can obtain more realistic and accurate image information,which also makes it difficult to solve the classification problem of high-resolution SAR images.Compared with low-and medium-resolution SAR images,the statistical characteristics of high-resolution SAR images have changed.Many classification methods used in low-and medium-resolution SAR images are no longer suitable for high-resolution SAR image classification.It is necessary to find a new method suitable for high resolution SAR image classification.The basic process of SAR image classification is "preprocessing,feature extraction,classification and post-processing".The quality of feature extraction and classification algorithm directly determines the effect of classification.This paper studies the classification of high resolution SAR images.The main work of this paper includes the following aspects:(1)The inherent randomness of SAR signal makes statistical model an effective tool for SAR image analysis.The parameters of statistical model provide valuable information for SAR image description.However,it is still a challenge for CNN to fit these parameters.According to the statistical characteristics of SAR signal,a statistical feature extraction module is designed to learn the basic statistical features.As an extension of the standard convolution layer,the statistical features of SAR image are fitted to adapt to the SAR image.The texture features of the image are introduced as prior information,and the initial feature map is obtained by combining the above features,which is then sent to the convolution neural network for learning and classification.The experimental results of several groups of real SAR images show that the algorithm can achieve better classification results compared with standard CNN and other classification methods.(2)Deeper deep learning network has stronger expressive ability.The deep network can integrate the characteristics of low,middle and high levels,and enhance the representation ability of the network.Using the deep residual network as the classifier can better extract the abstract features of the SAR image,and avoiding the slow convergence speed and optimization difficulties caused by the network being too deep.Channel and spatial attention mechanism are introduced in the network to make the algorithm model ignore irrelevant information and extract features from key parts.Combined with the above statistical feature extraction module,the feature expression that is more consistent with the nature of samples can be obtained.Experiments and Analysis on several groups of measured images show that the improved residual network classification method has higher classification accuracy.