Research On Accuracy Characterization In Land Cover Change Information

Categorical information of land cover and land cover change is premise and basis for geo-spatial information acquisition,resources survey and analysis,dynamic environment monitoring and globalization researches.It also supports geography and spatial analysis applications.Classification of remote sensing images which is the most direct method to obtain land cover distribution and change information belongs to scope of thematic mapping or categorical mapping.Land cover information includes static spatial distribution at some time in the past and dynamic changing features during some period of time.The latter is known as land cover change detection which is the processing progress of tracking the change of certain object or phenomenon,identifying and quantifying land cover change,representing in polygon features or raster datasets format generally.As one of the important subjects of land cover and its change,accuracy is an indispensable part of geo-spatial information research and analysis.Estimating accuracy of geographic category information such as land cover and its change provides not only evidence for thematic information quality assessment,selection of valuable information products and analysis reports,but also reference to optimization of both method and procedure,for better service for spatial analyses and applications.However intrinsic subjectivity,complexity and error of geographic category information lead to limited accuracy.Spatial and temporal correlation of categorical information from different time phase makes land cover change accuracy research more complex.Based on the local accuracy characterization in land cover change information,this thesis expand the research of geography category information in time dimension,considering spatial and temporal correlation of geographic category and its change information,this article constructs a systematic framework for quantification of local change information accuracy from the perspective of the user by integrated use of geostatistical methods,generalized linear model and other empirical models,etc.The main research work is summarized as the following:(1)Research goal,content,methodology and significance are discussed with current research on uncertainty of categorical information as starting point,combining description of uncertainty,theory of accuracy assessment,analysis method and sampling theory,analyzing current research status at home and abroad,summarizing existent problems in key area,comparing advantages and disadvantages of existing technologies and methods;(2)The theoretical foundation and related conceptions of geostatistic and generalized linear model are introduced and their application in land cover change accuracy assessment are explained.(3)A strategy well suited for the unmeasurable geographic categorical information and limited samples is proposed for location-specific characterization of accuracy in land cover change maps.The local accuracy characterization strategy is established based on local patterns of land cover change information and binary data representing pixel-specific(in)correct classification,which include exploring user domain covariates which significantly relate to regression modeling,logistic regression kriging method and discrete random field spatial variogram modeling.(4)Contraposing inaccessibility of different temporal samples at the same locations for land cover change information validation in practical studies,suggest one local change information quantification methods which effectively utilize samples at different locations.On the basis of direct logistic regression kriging,a two-step logistic regression kriging using non-collocated validation samples in addition to collocated ones is proposed and improve the characterization accuracy.Effectiveness and applicability of this method has been proved by experiments when sample data proportion is complex,especially there are not enough collocated samples.(5)Considering weaknesses of conservative sampling method in categorical information accuracy assessment,present an adaptive spatial sampling strategy based on uncertainty of prediction accuracy,such as variance or standard error.Use standard error of prediction accuracy as assessment criteria by which to recognize areas where classification accuracy need to be enhanced,progressively locate and adaptively collect the samples.Experiments have certified this sampling optimization improves efficiency of accuracy prediction-oriented reference data collecting.Main innovations of this thesis are summarized as followed:(1)Propose a local accuracy characterization method which integrating logistic regression and kriging while accommodating spatial correlation.Logistic regression is used to predict the probabilities of correct change categorization based on local patterns of map being analyzed,while kriging is performed to make corrections to regression predictions based on discrete random field spatial variogram modeling.It was found that this method leads to more accurate predictions of local accuracies.(2)An accuracy characterization method built on temporal correlation,data modeling and analyzing is presented,in order to predict local accuracies in land cover change information when samples collected for a specific map may not be at the same location as those collected for maps of the same areas but different dates,increase quantification accuracy for complexly configured validation samples,especially there are few or even no collocated samples.(3)Suggest a adaptive spatial sampling strategy based on prediction standard errors.Experiments indicate that this method is beneficial to improvement of sampling efficiency and prediction accuracy in land cover change categorization,compared with conservative spatial sampling method.Issues need to be further studied are listed below:(1)Spatial sampling method used for accuracy characterization when collocated samples are not available is one of the further research interests.(2)Local accuracy characterization of high resolution imagery and object-oriented land cover change maps is worthy to be discussed.(3)Utilization of the uncertainty of the reference data in analyzing uncertainty in land cover information and extensive landscape pattern modeling.