Change Detection In Remote Sensing Images Based On Pixel Information And Deep Learning

Image change detection has become an important issue in the field of image processing due to its wide applications in remote sensing research,agricultural survey,disaster relief and so on.The techniques developed so far focus on the alleviation of the adverse factors(like noise corruption)so as to detect the real changed areas.Therefore,it is required to make a deep and profound research on the feature information in the images.Through the pixel information and deep learning technique,this thesis makes a systematic research on the three following sub-issues: the binary change detection for the synthetic aperture radar images,the ternary change detection for the information unbalanced images and the multiple change detection for the 1-D images.Specifically,the main works of the thesis can be summarized as the following five aspects.(1)A locally fitting and semi-expectation-maximization model for the threshold approach is proposed,improving the thresholding accuracy when segmenting the difference image.Traditional threshold approaches focus more on the development of diverse models to fit the whole gray-level axis,usually overlooking the inner feature of the histogram.The thesis first makes a deep research on fitting of the histogram,analyzing its morphological characteristics theoretically.According to such characteristics,the unchanged class and the changed class are estimated through the locally fitting model and the semi-expectation-maximization algorithm,respectively.Finally,the optimal threshold is acquired through the Bayes formula which involves the computation of the posterior probability.The experimental results demonstrate that the model performs better than several traditional approaches available,being able to determine the optimal threshold accurately.(2)A fuzzy clustering approach with a modified Markov random field(MRF)energy function is proposed.After a research on the MRF,an MRF energy function is added into the traditional fuzzy c-means algorithm with the information provided by the neighborhood pixels fully utilized.To reduce the impact of the speckle noise further and to adapt the problem of change detection better,a novel form of the MRF energy function with an additional term is established to modify the membership of each pixel,and the degree of modification is determined by the relationship of the neighborhood pixels.The specific form of the additional term is contingent upon different situations,and it is established finally by using the least-square method.In this way,the good information in the neighborhood is fully utilized and the bad information is avoided to a large extent.The experimental results demonstrate that the novel clustering approach is abler to reduce the impact brought by the noise than several other improved clustering algorithms,leading to a more accurate result.(3)A dynamic guided learning framework is proposed to update the difference image.The posterior probability by the threshold approach and the membership by the fuzzy clustering approach are used as the referential criteria for updating the difference image.The thesis first makes an analysis of the characteristics of the posterior probability and the membership,and then a dynamic learning strategy is developed by using such characteristics to update the difference image in an iterative manner.Finally,the level-set approach which has been proved to perform well in edge preservation is adopted to generate the final map.The experimental results demonstrate that it is robust to learn the difference image by using the dynamic guided learning framework in terms of anti-noise and anti-overlapping.In addition,the final results suggest the superiority of the framework to the other ones which are based on static difference image.(4)A novel deep learning and mapping framework oriented to the ternary change detection task for information unbalanced images is proposed.Different from the traditional intensitybased methods available,the framework is based on the operation of the features extracted from the two images,and two networks are involved here.First,the stacked denoising autoencoder is used on two images,serving as a feature extractor.Then the stacked mapping network is employed to establish the mapping functions which denote the relationship between the features for each class.Finally,a comparison between the features is made and the final ternary map is generated through the clustering of the comparison result.The experimental results demonstrate its robustness and effectiveness in terms of accuracy,avoiding the influence brought by the unbalanced information to a large extent.(5)A novel fuzzy autoencoder(FAE)is proposed to tackle the multiple change detection for the 1-D images.Different from the traditional approaches based on the pixel intensity,FAE includes a multilayer structure through self-reconstruction to extract the feature from an image.Due to the problems existing in the ordinary autoencoder,the fuzzy number is introduced to the autoencoder to suppress the noise and learn robust features.In this way,the information in the fuzzy domain is introduced into the input,and in practice the fuzzy domain is discretized to facilitate the calculation,obtaining the corresponding propagation formulae.The experimental results from the FAE and the other compared approaches indicate its effectiveness and robustness in terms of accuracy,demonstrating that the FAE can generate more robust features even than the denosing autoencoer.