The change detection of remote sensing image is to detect the changes of two temporal phase remote sensing images in the same area.Synthetic Aperture Radar(SAR)uses active microwaves for imaging.Compared with other sensors such as optical and infrared sensors,SAR imaging is not affected by cloud,rain,and climate,and can realize all-day and all-weather detection of targets.With the continuous development of SAR imaging technology,more and more institutional radars with excellent detection performance have emerged,and higher requirements have also been placed on change detection methods.At present,SAR is very useful in flood relief,urban planning,agricultural management,and terrain explo-ration.This paper mainly studies the change detection method of SAR images,improves on the basis of the detection methods already proposed,and proposes more effective detec-tion methods combined with machine learning methods such as deep learning.The main research results of the paper are as follows:1.To improve the fitting degree of pixel distribution of remote sensing image in KI threshold segmentation algorithm,a kind of conditional distribution model based on G~0distribution is proposed.The analytic expression of G~0distribution is relatively simple,and the parameter estimation is very easy,and the applicability is very wide.It can accurately model single view and multi view areas with uniform,uneven and very uneven The detection effect is better.2.To solve the problem that the critical information of the fuzzy C-means(FLICM)algo-rithm is fuzzy in the changing and non changing regions of the image,this paper propose to extract the edge intensity information of the different image by using the bilateral filtering operator,and modify the pixel weight in the neighborhood window of the FLICM algorith-m,so that when the neighborhood window slides to the edge of the changing region,the tangential direction of the noise can be suppressed and the change can be well protected Details of the area.Combined with the idea of image fusion,the logarithmic ratio image and the neighborhood ratio image are fused by wavelet,and the advantages of the two different images are fused,which can improve the contrast between the changing and non changing regions as much as possible under the condition of background noise suppression,so as to facilitate the classification and detection of the difference image.3.Combined with deep belief network feature learning methods.Deep neural networks have many obvious advantages over earlier shallow neural networks.The training time is shorter.The layer-by-layer training method is used to avoid the problem of the weaker and weaker error propagation in traditional neural networks and prevent gradients.Diffusion occurs.The paper uses a two-dimensional Gabor filter to select appropriate training samples to train the deep confidence network,and use the trained neural network to perform classification detection to obtain the final change detection result. |