SAR Image Change Detection Based On Stacked Neural Network

The wide application of remote sensing image change detection technology in the fields of environmental detection,geography research,industrial production and military strikes has got full of scholars' attention.Synthetic aperture radar(SAR)has important practical value in dealing with environmental degradation,urban changes and other practical problems because of its all-weather monitoring,day&night imaging,wide coverage and high resolution.The direction of change detection technology development is strive to overcome external adverse conditions and effectively distinguish the changing regions in remote sensing images.The stacked neural network can extract the depth information of the image and make the change detection result achieve higher classification accuracy.And it has become a popular research object in recent years.Aiming at the two-class task of SAR image change detection,this paper mainly studies the change detection algorithm based on pixel information and stacked neural network,improving the two steps of generating,analyzing the difference map in the traditional change detection process and proposing a new change detection framework by using the stacked neural network.The research results show as follows:1)A change detection method based on improved neighborhood ratio and stacked autoencoder is proposed.The logarithmic constraint is added to the neighborhood ratio algorithm to construct a better quality difference map to reduce the interference of multiplicative noise on the image.Analyze the difference map with the stacked autoencoder to fix the inaccurate classification area in the previous step.The denoising effect of this method performs better,and the contour edge information of the change region is also well preserved.The change detection map generated on the two data sets has the highest consistency with the reference map.2)A change detection method based on the feature extraction and mapping of stacked neural network is proposed.The original image inputs into the neural network for training,which gets rid of the steps of preprocessing and constructs the difference map in the traditional change detection process.It is completely based on the image of detecting changes in depth feature matching,analysis,and classification.The accuracy of the change detection is improved by the reasonable parameter setting of the two neural networks.The use of stacked neural networks can extract and combine robust image features,effectively avoiding the destruction of images by speckle noise and unbalanced information.