Uncertainty Analysis In Object-oriented Remote Sensing Image Classification

As the appearance of high spatial resolution remote sensing imagery, object-oriented remote sensing image classification technique has become a research hot spot recently. It segments the remote sensing image into objects firstly, and then extracts various object features for classification. In object-oriented remote sensing imagery classification, each involved procedure such as data processing, image segmentation, object feature extraction and selection, classification may introduce different uncertainty which will finally result in the uncertainty of the classification result.Analysis of the uncertainty in classification is helpful for users to understand where the uncertainty comes from. This information provides the fundamental for reducing the uncertainty of classification result. Moreover, the accurate assessment of classification accuracy is not only useful for users to understand the quality of thematic data, but also useful for validating the methods used in classification.This dissertation takes the object-oriented remote sensing image classification as the major clue to study the uncertainty in classification qualitatively and quantitatively. The whole classification procedure is simplified as a linear series system consisted of three main steps, including image segmentation, object feature extraction and classification. Firstly, the uncertainty sources in classification are analyzed. Then, uncertainty assessment measures are defined based on objects to assess the image segmentation and classification accuracy.The major works are listed as follows:(1) The sources of uncertainty and the propagation of uncertainty in object-oriented remote sensing imagery classification are analyzed.(2) Based on the principle of supervised image segmentation assessment, the quantitative measures of "over-segmentation", "under-segmentation" and "object boundary" uncertainty are defined to assess the image segmentation accuracy. The segmentation accuracy assessment and scale selection experiment validates the effectiveness of these segmentation uncertainty measures. Supervised image segmentation accuracy assessment method is proved to be an effective way to select suitable segmentation scales for different classes of ground objects.(3) The object location uncertainty and its effect on object feature extraction and classification are analyzed to study the selection problem of image resampling and object boundary transferring schemes. Based on the analysis result, an optimal object location method for object-oriented multi-source remote sensing imagery classification is proposed.(4) Based on the review of principles and methods of remote sensing image classification accuracy assessment, a complete set of object-based remote sensing imagery classification accuracy assessment methods are proposed. Firstly, an effective object-based sampling scheme is proposed, and then the object-based confusion matrix is constructed to assess the classification accuracy. Two methods to construct the confusion matrix based on objects are analyzed. The first one treats each object as one element in classification error matrix construction, and the other weights each object with its area to construct the classification error matrix. The accuracy assessment result of the former method reflects the distinguishing power of the classification, namely the power of classifying the objects correctly. It provides the map producer a way to assess the classification performance and select the proper methods for classification. The accuracy assessment result of the latter method reflects the area proportion of the whole or each category of objects that are correctly classified on the thematic map, namely the thematic map accuracy. It gives the users valuble information of the quality of the classification result, according to which the applicability of the classification result for a certain application task can be determined.(5) The impact of object positional uncertainty on the object-based thematic map accuracy assessment result is analyzed. A thematic map accuracy assessment method considering the object positional uncertainty is proposed. This method takes the components of each accuracy assessment object with different classes into classification information statistic to construct the error matrix and assess the thematic map accuracy. It can avoid the effect of object positional uncertainty and assess the thematic map more accurately.The main innovative points are as follows:(1) An image segmentation accuracy assessment method considering both the difference of the regional inner pixel consturction and object boundary is proposed. It can assess the image segmentation result from three aspects, including "over-segmentation", "under-segmentation" and "object boundary" uncertainty. Based on these assessment measures, an optimal segmentation scale selection method is designed which can meet various remote sensing task requirements.(2) Based on uncertainty analysis of object location on object feature extraction and classification, a proper object location scheme for latter object-based multi-source feature extraction and classification is designed.Uncertainty analysis result of object location indicates that when using different object location scheme the posional accuracy requirements of object boundary is different. In object location, when low spatial resolution images are resampled to high spatial resolution, the object statistical features and classification accuracy are little affected by the object boundary uncertainty; transfer of raster or vector object boundaries are both adoptable. Whereas when images are geo-registered without changing spatial resolution boundary uncertainty has a significant influence on the statistical value of texture feature and tends to induce the instability of classification results, the object location uncertainty cannot be disregarded unless it is controlled in a certain limited range. It is therefore suggested to resample images to high spatial resolution before transferring objects to them. In this way, if only errors of the initial object boundary and image registration are controlled in one pixel range, the object boundary uncertainty is not going to result in a large difference in object feature calculation and classification, and raster and vector object boundaries are both adoptable in object location. Users can focus on the latter object feature extraction and classification procedures without considering the influence of object location uncertainty.(3) A complete set of object-based remote sensing imagery classification accuracy assessment methods are proposed, including the stratified random sampling scheme, the sample number determination method and confusion error matrix accuracy assessment methods that are applicable for assessing the classification performance and the thematic map accuracy. Firstly, representative testing samples are gathered by stratified random sampling method. Then, the confusion matrix constructed by weight each object with equal weight and and object area are used to assess the classification performance and the thematic map accuracy respectively.The object equal-weighted accuracy assessment method can assess the distinguishing power of classification, namely the power of classifying the objects correctly, quite well. It provides the map producer a way to assess the classification performance and select the proper methods for classification. The object area-weighted accuracy assessment method can assess the area proportion of the whole or each category of objects that are correctly classified on the thematic map, namely the thematic map accuracy. It gives the users valuble information of the quality of the classification result, according to which the applicability of the classification result for a certain application task can be determined.With consideration of the negative effect of object positional uncertainty on the thematic map accuracy assessment, the paper proposes an object "components" based classification thematic map accuracy assessment method to more accurately assess the thematic map accuracy. In this method, the inner pixel sets of different classes within the testing object are called object "components", according to the true reference class information of these object "components", the classification information of testing objects are firstly gathered to construct the confusion matrix, and then the classification accuracy is assessed. This method can assess the classification thematic accuracy more precisely.Moreover, a sample reference data collection method for object "component" based assessment method is proposed. It is suggested to collect the reference information of testing objects under the rule that:for the objects that are of pure or less important category, only record its major component; for other objects that are of impure or important categories, record more detailed reference information for assessing the thematic map accuracy as accurate as possible.Further research is needed on the following issues:(1) The mathematic modeling of the uncertainties in classification needs further research. Error propagation functions can be established to provide the quantitative base for error propagation analysis in classification.(2) The discrepancy measure is the key in supervised image segmentation. How to define discrepancy measures that are proper for assessing image segmentation result needs more work. To assess the segmentation accuracy more comprehensively, combination of different discrepancy measures to measure segmentation accuracy needs further research. On the other hand, the unsupervised image segmentation accuracy assessment methods should be studied in the future to meet the assessment requirement of image segmentation without reference data.(3) To meet the expanded classification accuracy assessment requirements, methods for the classification accuracy assessment should be studied further.(4) More complicated sampling methods should be studied in the future to promote the object-based remote sensing image classification accuracy assessment method in practice.