Change Detection For Hetergeneous Remote Sensing Images Based On Deep Neural Network

Image change detection is a process that identifies the differences between the two remote sensing images of the same area acquired at different times.With the development of application requirement,the technique for change detection based on heterogeneous remote sensing images is paid more attention.However,the detection of the differences between two heterogeneous images is very challenging,for the images shot by equipment installing different sensors have different data representation for the same true area,which results in that generating a difference image is impossible by directly comparing.In this thesis,the deep neural network is utilized to extract features from the two heterogeneous images and the changed areas can be acquired by analyzing extracted features.This dissertation primarily includes the following three aspects:1)Change detection based on local information reconstruction for heterogeneous images is proposed.One of the two images is input the deep neural network to reconstruct the local information of the other image,and then reconstruction error is used to generate the corresponding difference image.The result map can be acquired by classifying the difference image into two categories.2)The method based on discriminative feature learning for change detection is designed for heterogeneous images.In this method,through an approximately symmetrical deep neural network the two remote sensing images are transformed into feature space,in which the difference image is generated by comparing the paired feature vectors.The changed and unchanged areas can be obtained by segmenting the difference image.The whole process does not need the reference image when the deep network is being fine-tuned and the samples are selected automatically.It is completely unsupervised.3)The simulated labels learning is analysed for change detection based on deep neural network.The analysis of difference in the feature space is improved.In the updating of parameters,in order to enlarge the disparity between changed areas and unchanged areas the simulated labels are suitably revised.And the difference image is segmented by an improved fuzzy clustering algorithm.These three methods gradually improve the accuracy of the change detection results for heterogeneous remote sensing images.The method based on local neighborhood information reconstruction select all the pixels for samples when training network,and the accuracy of the result is much better than the traditional postclassification comparison algorithm.The design of discriminative feature learning selects the samples from two classes,the changed and the unchanged,and sets two different type of labels for them during the fine-tuning of parameters.The performance of the difference image and the result map is improved significantly.The method based on simulated labels learning mends the value criterion of the simulated labels,and it optimizes the performance of the change detection.The experiments results demonstrate that the proposed methods achieve quite high accuracy for heterogeneous remote images.