Research On Different Deep Neural Networks For Remote Sensing Image Change Detection

Due to the population growth,the surface of the earth is undergoing continuous changes.The use of remote sensing image change detection to analyze and detect surface changes is a hot spot.In the past ten years,change detection has attracted more and more researchers' attention.A variety of methods for detecting changes in remote sensing image data have sprung up,and these methods have also been applied to many fields,such as disasters assessment,land planning,video surveillance,and forest warnings.However,since the acquisition of remote sensing images is easily interfered by various clutter,resulting in blurred image contours,and with the continuous development of remote sensing technology,the amount of data acquired also increases exponentially,traditional change detection methods face enormous challenges in terms of detection performance and efficiency.In recent years,deep learning has been applied to computer vision and natural language processing with its excellent feature representation ability.It breaks through the constraints of traditional algorithms and injects fresh blood into remote sensing image processing.This paper focuses on the change detection of remote sensing images,and designs three different deep neural network models to flexibly implement problem-driven change detection tasks.The main works is given as follows: 1.A remote sensing image change detection method based on deep separable convolution is developed.Considering the joint distribution of images,the method stacks bi-temporal images into a two-channel image,each of which is an image block at different times.After such processing,image change detection can be regarded as an image segmentation problem,and then the traditional U-Net structure of the full convolutional network,which is common in image semantic segmentation,is used as the basic framework,and the traditional convolution operation is replaced by the deep separable convolution,which improves the convolution efficiency,greatly reduces the training parameters of the model,reduces the requirement of data quantity and the risk of model over-fitting.Then the whole network is trained in an end-to-end way to optimize the model parameters.When predicting,the normalized two-time images are stacked into a two-channel image,which is input into the trained segmentation network to output the change detection results directly.At the same time,in order to satisfy the different preferences of different tasks for accuracy and recall rate,a loss function with controllable preferences is designed.The method is verified in different data sets,and the ideal experimental results are obtained.2.An image change detection method based on information transfer and attention mechanism is proposed.This paper presents an image change detection method based on information transmission and attention mechanism.Convolutional neural network can extract rich spectral-spatial features of remote sensing images,and recursive neural network can effectively analyze the time dependence of bi-temporal images.Therefore,combining the two methods,the former is used to extract spectral-spatial features of original images,and the latter is used to analyze temporal characteristics,so that change detection tasks can be well handled.Based on this,this paper constructs a deep neural network model,and designs an image semantic information transmission module and introduces attention mechanism.By using the rich semantic information of bi-temporal images and attention mechanism,the recognition degree of classification vectors is improved,and the change information of images is effectively enhanced.By comparing the method with other existing methods,the experimental results demonstrate the effectiveness of the method.3.Finlly,we also put forward a deep evolution network structure optimization method for remote sensing image change detection.This method implements deep neural network structure optimization through evolutionary algorithm,which replaces time-consuming and laborious manual tuning.Specifically,the deep neural network structure is first encoded to form an initial population,and then an appropriate evolutionary strategy is selected to update the network individuals in the population in an iterative manner,and the optimal network individual is obtained from the last generation population.The method can adaptively design a deep neural network structure to perform image change detection according to a specific data set or a detection task,so that image change detection becomes more flexible.