Probabilistic modelling of stability and resolution of thematic classes from remotely sensed and digital terrain data

The conventional methods of information extraction from remotely sensed data based on the use of statistical techniques such as Maximum-Likelihood Classification (MLC) and deterministic decisions can lead to inefficient use of these data. These limitations are more apparent in high resolution imagery where pixel-based classifications do not produce satisfactory results. Furthermore the accuracy data provided in the form of confusion matrices provide minimal spatial information on the reliability of the classified maps. In this research, practical techniques for evaluation of the stability of the classification and for more efficient use of the probabilistic measures obtained from the MLC were developed and have led to production of useful spatial information on the reliability of the classification. Practical tests of this research have been performed on Landsat-TM images of two different geographic environments including a semi-natural environment (Tidbinbilla Nature Reserve) and an urban area (Canberra). Based on the stability measure developed here, a paired classification approach for identification of pixels with their two most likely classes is proposed. A practical test of this approach shows that a considerable amount of useful information was obtained about mixed pixels. To overcome the limitations inherent in the single-scale classification of high resolution imagery , a probabilistic multiresolution classification approach based on classification of data of varying scale and integration of the resulting classifications was developed. Accuracy of the resulting classification is shown to be superior to that of the single-scale classification. A spatial distance-based approach for filtering and generalization of the classified map has been developed and has been shown to have advantages over modal filtering, the most commonly used approach for introducing spatial generalization. A logistic modelling approach for efficient evaluation of the classification performance is proposed. In this approach, the probabilistic measures and contextual data obtained from a MLC are used as explanatory variables to model the probability of accuracy. In addition to the MLC, applicability of the proposed approaches for mapping of the accuracy probability has been shown to be successful on the minimum distance to means classification approach