Uncertainty Of Object-Based Image Analysis For High Spatial Resolution Remote Sensing Imagery

Recent advances in airborne and spaceborne remote sensing technology together with increasing remote sensing observation practices have led to a rapid development in the acquisition of images with a high spatial resolution.However,the rapid analysis of such images is challenged due to the vast quantity of available images as well as their respective characteristics.That is,traditional per-pixel image analysis cannot successfully process high spatial resolution images due to some new characteristics,for example,distinct terrain details,complex spectral features,and the increased spectral difference of internal categories.Therefore,Object-Based Image Analysis(OBIA),which was designed specifically for high resolution remote sensing imagery,has been receiving more attention from researchers as a novel paradigm of remote sensing image analysis.However,the use of OBIA generates some new problems compared to per-pixel methods,due to the uncertainty of data quality,technological parameters,and processing methods.The main objectives of this study are to analyze the uncertainty of object-based image analysis using high spatial resolution images,including object-based image classification and object-based change detection.Moreover,it also aims at exploiting a high-efficiency classification scheme to label segmented objects.In a first step,the uncertainties relating to Object-Based Image Analysis were systematically analyzed to determine how the accuracy relates to different factors pertaining to the Object-Based Image Analysis procedure.Aside from aiding the development of object-based image analysis technology,the conclusions can also guide the processing of extremely high spatial resolution images in the actual application.Secondly,an unsupervised classification method was proposed.This is a specific information extraction method,used to enhance the efficiency of processing high spatial resolution images in order to make full use of the contextual information.The contents and conclusions of this dissertation are as follows:(1)The uncertainty of multi-resolution segmentation.The first step to studying the uncertainty of object-based image analysis is analyzing remote sensing image segmentation,which could significantly affect the accuracy at each stage of the object-based image analysis.Multi-resolution segmentation was employed as the basic segmentation algorithm,whereby the method was first assessed in order to ensure the credibility of the following research.Secondly,the trans-scale sensitivity of several spatial autocorrelation measures was investigated,including intrasegment variance of the regions,Moran's I autocorrelation index,and Geary's C autocorrelation index.Finally,a top-down decomposition scheme was presented to optimize the segmented objects derived from multi-resolution segmentation,and its potential was examined using an unmanned aerial vehicle image.The experimental results demonstrate that the proposed method is able to effectively improve the segmentation in urban areas or highly consistent areas.(2)Analysis of the segmentation scale and object features in object-based image classifications.Segmented object features and segmentation scale are both important factors impacting object-based image classifications.This study therefore reported the effect of training set size and sampling schemes versus segmentation scale and feature space on object-based image classifications.The results proved that the Overall Accuracy(OA)was significantly affected by the training set size,and that it is also affected by the Segmentation Scale Parameter(SSP).Generally,the SSP should not be too large for a classification with a small training set ratio.By contrast,a large training set ratio is required if the classification is implemented using a high SSP.In addition,we suggest that the optimal SSP for each class has a high positive correlation with the mean area obtained by manual interpretation,which can be summarized by a linear correlation equation.(3)The uncertainty of object-based feature selection techniques.The availability of several feature selection and supervised classification techniques leads to more uncertainty in the process of supervised classification.In this study,such methods were systematically evaluated.Eight supervised feature selection methods were assessed,including five feature-importance-evaluation methods and three feature-subset-evaluation methods.It was demonstrated that it is valuable to perform feature selection prior to object-based classification for both considered classifers,even though a negative impact from both wrapper methods is sometimes observed.Overall,Support Vector Machines Recursive Feature Elimination(SVM-RFE)is a preferable method for both classifiers compared to the other three feature-importance-evaluation methods(i.e.,Gain Ratio,Chi-square and Relief-F),and Correlation-Based Feature Selection(CFS)is the best feature-subset-evaluation method compared to the both wrapper methods.We expect that wrapper method with a polygon-based cross validation is evaluated to improve the performance of wrapper method on object-based classification.(4)The uncertainty of object-based supervised classification techniques.Previous studies have not systematically investigated all relevant factors affecting the classification(segmentation scale,training set size,feature selection and mixed objects),and there is no general agreement regarding the comparison of different classification techniques within OBIA.Therefore,the classification performance was systematically analyzed using a variety of statistical and machine-learning classification methods,and the effect of mixed objects was quantitatively evaluated.Advanced statistical methods and visual inspection were used to compare these factors in two case studies in China,including the Mutiple Comparisons and Analysis of Covariance(ANCOVA).The results indicate that Random Forest(RF)and SVM are highly suitable for OBIA classifications,and confirm the expected general tendency of the overall accuracies to decline with increasing scale.RF and Decision Tree(DT)classifiers are the most stable with or without feature selection,while all other classifiers yielded significant differences between using selected feature sets and using all features.There is an equal response to the mixed objects phenomenon between all classification techniques used.Overall,when considering all factors,no single classifier was consistently superior in all cases,but it was suggested that RF should be considered in most cases.(5)The uncertainty of object-based change detection.Similar to the object-based supervised classification,the diversity of feature space,segmentation strategies,and change detection algorithms lead to more uncertainty in object-based change detection.Four common unsupervised prediction methods with different segmentation strategies and a series of scale parameters using two worldview-2 images were tested.Furthermore,the effect of the extra texture metrics and Normalized Difference Vegetation Index(NDVI)in addition to "spectral-only" was evaluated.This study could guide the selection of the segmentation scale,feature space and change detection methods.The results indicated that a superior performance among prediction methods was achieved at a medium scale,rather than at the fine scale close to the pixel level.The segmentation units generated from images acquired on two dates more significantly contributed to the overall accuracy compared to the single date image segmentation.Multivariate Alteration Detection(MAD)always led to a better accuracy performance than the other methods considered at the same confidence level.Adding texture and NDVI metrics had negative magnitude effects on the detection accuracy,which were not consistent among scales,prediction methods or units.We conclude that an object-based strategy utilizing the segmentation of images acquired on two dates is useful for high resolution image change detection;however,optimizing the threshold for different prediction methods is critical and prediction methods should be further improved in order to deal with additional texture metrics or others.(6)Study of object-based unsupervised classification.Previous studies showed that there is great uncertainty in object-based image classification and therefore the improvement of classification efficiency is difficult.Full use was made of the contextual information provided by objects on the basis of uncertainty studies,and a novel cultivated land information extraction method based on triangulation was presented.On the one hand,it was found that the main factor interfering with cultivated land information extraction is information from densely populated residential areas.On the other hand,the clustered distribution of cultivated land means that the segmented cultivated land plots are larger,due to the characteristics of very high resolution imagery,and therefore the segmented residential areas appear fragmented as a result of abundant internal surface features.We therefore consider the cluster approach as beneficial for eliminating such residential land,in order to achieve the ultimate goal of cultivated land information extraction.The results demonstrate that the proposed method can significantly improve the efficiency of cultivated land information extraction.