Abstract

Creation of accurate vegetation maps from optical satellite data requires use of reference data to aid in interpretation or to verify map results. Reference data may be taken, for example, from field visits, aerial photo-interpretation, or ground-based inventories. National inventories are a potential source of reference data useful in land cover mapping projects. This thesis addresses aspects of mapping forest and alpine vegetation in Sweden through combined use of optical satellite data and inventory data. Issues such as reference and satellite data pre-processing, spatial scale, quantity and quality of reference data, and classification methods have been examined. Optical satellite data with pixel sizes ranging from 10 to 300 m have been used together with reference data from the Swedish National Forest Inventory (NFI), National Inventory of Landscapes in Sweden (NILS), a point sample based on the Terrestrial Habitat Monitoring program (THUF), and a forest stand database. Results include modifications to common remote sensing methods, such as introducing iterative adjustment of prior probabilities in Maximum Likelihood classification, and improved topographic normalization (C-correction) of satellite data. Probability-based samples such as NFI, NILS and THUF provide data necessary for assignment of prior probabilities, estimation of continuous values, and are useful as training and validation data. For managed boreal forest stands, coarser pixel (60 m) AWiFS data were nearly as effective for stem volume estimation as SPOT 5 data (10 m). On the other hand, the most accurate classification of detailed alpine vegetation types (72.9% overall accuracy) was from SPOT 5 data combined with elevation derivatives, while classifications of Landsat TM (25 m), AWiFS, and MERIS (300 m) were less accurate. Non-parametric methods (e.g., random forests, decision/regression trees) produced higher classification accuracies than traditional parametric methods for alpine vegetation. The quantity of reference data affected classification accuracy, as more reference data produced higher map accuracy, although other factors such as distribution and quality of the reference data should be considered. As seen in this thesis, the characteristics of the landscape exert an influence on satellite and training data requirements, classification methods and resulting map accuracy.