Modeling Local Land Cover Change Information Accuracies With Complexly Configured Reference Samples

Accuracy is a fundamental property of land cover and land cover change information.Characteristics of land cover change accuracies are more complicated than single-date accuracy characteristics.Research and analyses about accuracy are indispensable components of land cover data production and applications.Designbased inference estimates global accuracy parameters by the error matrix.However,distributions of accuracy parameters are not uniform in the space.Recent studies have extended accuracy measures to per-pixel level.This dissertation focuses on local land cover change accuracies.Accuracy assessment and analysis is confined to a finite sample in the study area.Accuracies of bi-temporal classification are generally assessed by collecting sample units whose locations remain unchanged across multiple single-date thematic maps.Conventional design-based inference method is notably limited when collocated sample is sparse or even unavailable.Although the absence of change information,noncollocated sample units are more convenient to collect and of sufficient quantities.Therefore,the vital question of the dissertation is how to efficiently reuse existing noncollocated single-date reference sample units for local land cover change categorization accuracy predictions.This dissertation,which concentrates on estimation and analysis of land cover and land cover change accuracies,applies and compares several methods of land cover change accuracy estimation by employing related theories and methodologies including survey sampling,generalized linear models and geostatistical random field theory.Major content of the research is summarized:1.On the condition of complexly configured reference samples,propose a sampling protocol which exploits both collocated and non-collocated sample units.Proposed sampling design satisfies accuracy assessment requirements and improves sampling efficiency.Accuracies of land cover change categorization with complexly configured samples are assessed in design-based inference framework.2.Design-based inference method assesses accuracies by the error matrix.However,accuracy measures derived from the error matrix provide no information about error spatial distribution and local accuracy diversity.Based on different sampling schemes,build direct and indirect local accuracy models with complex sample configurations in model-based inference framework.Exploiting non-collocated sample units,achieve extended modeling methods which predict land cover change accuracies by single-date accuracies.3.Propose the cross-correlation-adjusted product method which exploits singledate non-collocated sample units and is proved to be of better performance.Suggestions about model selection in practical scenarios are offered based on analyses about the links between model performances and sample configurations.