Change Detection In Remote Sensing Image Based On Evolutionary Optimization And Convolution Neural Network

Change detection is an important application of remote sensing images processing.It is a technique that uses two or more remote sensing images acquired at different time in the same geographical area to qualitatively or quantitatively detect the changed areas.With the development of high technology,the improvement of the acquisition of remote sensing data and the accumulation of remote sensing data,remote sensing images become an indispensable method for ground observation.The change detection method in remote sensing images can quickly obtain the change information of ground,which is more economical and convenient.It has a wide range of application prospects and urgent scientific research needs.Change detection consists of two steps: generating difference image and analyzing difference image to obtain change map,where obtaining change map is a binary classification.Genetic algorithm(GA)is a kind of evolutionary optimization method which obtains the optimal solution by random global search.It has the characteristics of high efficiency and parallelism.Therefore,GA can be used to analyze the difference image to obtain the category label for each pixel in the difference image.However,when the size of difference image is too large,the high computational complexity of GA becomes the bottleneck of its application.Considering the mechanism of the human visual system which pays attention to significant regions to obtain useful information and reduce the amount of computation.This thesis combined the saliency detection model with GA,and proposed two remote-sensing-image change detection methods based on evolutionary optimization.The aim is to reduce the search space of GA and improve the capability of noise suppression,so that the detection results obtained will be better.In addition,in the non-significant areas obtained by the saliency detection model,a large number of reliable unchanged-class samples with rich noise information can be obtained.Through pre-classification,highly reliable samples of changes can be automatically obtained.These samples are used as training samples,and a change detection method in remote sensing images based on convolutional neural network is proposed.The main work of this thesis is summarized as follows:(1)A method of change detection in remote sensing images based on the salient map guidance and an accelerated genetic algorithm(S-a GA)is proposed.The difference image is first generated by logarithm ratio operator based on the bi-temporal remote sensing images acquired in the same region.Then a saliency detection model is applied in the difference image to extract the salient regions containing the changed class pixels.The salient regions are further divided by fuzzy c-means(FCM)clustering algorithm into three categories: changed class(set of pixels with high gray values),unchanged class(set of pixels with low gray values)and undetermined class(set of pixels with middle gray value,which are difficult to classify).Finally,the proposed accelerated GA is applied to explore the reduced search space formed by the undetermined-class pixels according to an objective function considering neighborhood information.And an efficient mutation operator is designed by using the neighborhood information of undetermined-class pixels as the heuristic information to determine the mutation probability of each undetermined-class pixel adaptively,which accelerates the convergence of the GA significantly.The experimental results on two data sets demonstrate the efficiency of the proposed S-a GA.On the whole,Sa GA outperforms five other existing methods in terms of detection accuracy.(2)An accelerated genetic algorithm based on sub-block decomposition and fusion for change detection in remote sensing images is proposed.This method is an improvement of the first method.The main research motivation is to further improve the anti-noise performance and operating speed when GA applied on the change detection problem.In the proposed algorithm,the BM3 D algorithm is used to preprocess the remote sensing images,so as to enhance the useful information of the image and suppress the noise.Then,the preclassification method based on saliency detection and FCM is used to achieve automatic shrinking of the search space and the position of the indefinite class of pixels is obtained.Next,referring to the idea of divide-and-conquer strategy,the difference image containing only significant regions is globally sub-block decomposed,with each sub-block independently using an accelerated GA optimization.Finally,the optimization results of all the sub-blocks are remapped to their original difference image positions,and the results of sub-blocks are merged in the local neighborhood at the decision level to output the final change map.Through parameter analysis,the most suitable sub-block decomposition number 4×4 is obtained.The experimental results fully illustrate that the proposed method has good anti-noise performance and detection effect,as well as less running time.(3)A change detection method in remote sensing images based on saliency detection and convolutional neural network is proposed.First,the training samples are automatically acquired using pre-classification based on saliency detection and FCM in difference image.And the sample selection method as follows: The unchanged class pixels of Non-significant regions and the changed class pixels of salient regions are taken in the form of 9x9 neighborhood blocks centering on them.And a fixed number of training samples are randomly repeatedly sampled from them.Then,a convolutional neural network for change detection in remote sensing images is constructed and trained with selected reliable samples.Finally,using the trained CNN to classify the indeterminate pixels in the salient region,and obtain the final change map.In the process of sample selection,non-salient areas have large numbers of unchanged class samples with high reliability and rich noise information,so that the robustness of the convolutional network trained is strong,and satisfactory detection results can still be obtained in the remote sensing images with strong noise.