Research On The Remotely Sensed Images Change Detection Methods Based On Uncertainty Analysis

Change detection(CD)based on remotely sensed images,as a remote sensing application technology,has significant demands for reality in the fields of Land use/cover change monitoring,city development,environment monitoring and disaster assessment.Thus,CD has been a research hotspot in the world.However,the process and the results of CD have a variety of uncertainties problems resulting from the complication of the ground-surface conditions,the uncertainties in the process of remote sensing information collection and transmission and the defects of the change detection methods.Thus,in this dissertation,the uncertainty problems in CD were firstly discussed and the CD methods based on the uncertainty analysis were then studied directing at the spectral uncertainty,spatial information uncertainty and the scale uncertainty.This dissertation is aiming to 1)improve the basic theory of CD from the point of uncertainty.2)promote a series of more reliable CD methods to sovle the uncertainty problems.All the efforts are to give theory supports for CD applications.The main research works and conclusions are as follows.(1)The CD technology architecture and the sources of uncertainty in the CD process was both systematically and deeply discussed.Furtherly,the spectral uncertainty,spatial uncertainty and scale uncertainty were emphatically discussed.Specially,the relativity of the ‘true changes' and the non-isostasy of ‘changes' and ‘no-changes' in change detection were proposed.For scale uncertainty,four connotations of the ‘scale' concept in remote sensing were extracted from the point of data process model,including the inherent scale of geo-objects,observation scale of remotely sensed images,the scale of image analysis and the model scale in image analysis model.(2)For spectral feature,a strategy that introducing priori information in CD process was proposed to decrease the uncertainty.Specifically,a fuzziness guided adaptive threshold CD method was proposed.In this method,the degree of membership belonging to the changed or unchanged class is considered as the priori information to guide the threshold decision based on a typical global threshold.The theoretical analysis indicated that,the adaptive threshold has a stretch effect against the global threshold.The experimental results showed that the proposed method is robust to noise and has abilities to detect detailed change areas.The accuracy of the CD results was improved by the proposed method.(3)For spatial feature,a strategy that introducing the fuzziness to the spatial constrains was proposed to decrease the uncertainty.Two improved methods were proposed focusing on the defects of classical markov random field spatial constrains model.1)The information entropy was considered as the label uncertainty description of the neighbored pixels.The spatial constrains model was improved by introducing the information entropy in describing the spatial relationship between the neighbored pixels more reasonably.It is more robust to the uncertainties.2)An adaptive superpixel based markov random field CD method was developed.In this method,the spatial contrains is described with a region of adjacent map,which is a graph model with a unregular size.Furtherly,the label fuzziness of a superpixel and the dissimilarity with its adjacent superpixel were both introduced into the graph based spatial constrains.The experimental results of the two methods shows that introducing fuzziness into the spatial constrains model can improve the uncertainty robustness.It can also conquer the over smooth problem in the classical markov random field model and improve the accuracy of the CD results.(4)For scale character,a strategy that fusing the multi-scale information was proposed to decrease the uncertainy.For image observation scale,the scale effects in CD was firstly studied in which a scale effects analysis method and optimal resolution chosing method using both spectral and shape feature was proposed.The experimental results have indicated the effectiveness of the proposed method.For image analysis scale,a CD framework based on multi-scale saliency detection was developed,in which a pixel level weighted fusion method for multi-scale saliency map was desighed.The experimental results indicated that 1)the proposed method is noise immuned;2)It is not the facts that the tinier the scale is,the more accurate the CD results are;3)The proposed pixel level weighted fusion method for multi-scale saliency map can not only keep the regional character of the change areas but also has an ability to detect the detailed changes.