Integrating the user's perspective into object-based land cover mapping

Segmentation of remotely sensed data is increasingly used to create spatially connected groups of pixels, commonly called objects, which are then used as the basic spatial unit in land cover mapping via image classification. There are many methods for image segmentation, and numerous outputs are possible, so numerous that is often impractical to classify all of them and then evaluate each. For this reason accuracy assessment of image segmentation is a necessary step to select a suitable result for object-based classification, hopefully the one affording the highest possible classification accuracy. Commonly the assessment of the accuracy of image segmentation is based on only the geometric properties of the objects derived (e.g. shape). A consequence of this approach is that all segmentation errors are regarded implicitly as being equally serious. However, the sensitivity of a specific map user to error may vary as a function of his/her needs and the classes involved. This thesis argues that a more appropriate assessment of a segmentation output is to consider the thematic content of the objects as well as their geometric properties. This allows the assessment to be tailored to the needs of the specific user. A metric that expresses the degree of thematic quality of objects from a user’s perspective, the thematic similarity index (TSI), is proposed. Then, a geometric-thematic method for image segmentation accuracy assessment is described, which combines a traditional method from the literature with the TSI. The perspectives of three users (a wolf researcher, a general user of land cover information, and the climate modelling community) were adopted in several case studies to analyse the TSI and the new method. The results show that the TSI is able to accommodate the user’s needs into image segmentation accuracy assessment, with the geometric-thematic method allowing the selection of a segmentation output more suited to the user than that from the use of the standard geometric-only approach. Furthermore, the use of the geometric-thematic method in operational contexts is illustrated. This includes a proposal for training an image classification in which mixed objects are used for training (which can increase classification accuracy), and using weighted estimators of classification accuracy which are able to assess the quality of a land cover map from the perspective of the user. This thesis thus integrates the user’s needs in all the main stages of an object-based image classification, which proved to be beneficial for land cover mapping production.