Skip to main content
Research article, 2003

Comparative Performance of Classification and Stratification Methods in Remote Sensing and Forest Applications

Norström, Fredrik


New technology, like the Global Positioning System (GPS) and Landsat TM, gives interesting alternatives to improve predictions of the ground truth. In this work simulation models were created to evaluate statistical properties for Landsat TM images in forestry applications. The common factor of the applications are classification of pixels based on remotely sensed data. Classifications were used for three different objectives. The first objective was stratified sampling with classes consisting of pixels classified to the same forest class. The performance of the stratified sampling was evaluated from comparisons with simple random sampling. For the second objective, comparisons of different classification methods and their sensitivity for an imperfection in the coordinates of the field data and the corresponding satellite pixel were performed. Different levels of imperfections were used that corresponded to positional errors that often exist in a real situation. The last objective was to construct probabilities of occurrence, on pixel level, for a forest class. A probabilistic classifier was introduced for that purpose. The probabilities can be used both to measure how reliable a traditional classification method is, and for area estimates. The objectives were evaluated with simulation models. In those models, spatial dependency between pixels were included. Comparisons between stratified sampling and simple random sampling show that stratified sampling, based on remotely sensed data, produce more precise inventory estimates of wood volume tha simple random sampling. Remotely sensed data are likely to be useful for more efficient sampling designs. When an imperfection in the coordinates were added some of the used classification methods showed some sensibility for it. The sensibility differed between the classification methods but no general conclusion could be drawn about what method that was best or worse when an imperfection existed. The performance of the probabilistic classifier indicated that it could be a useful alternative


Monte Carlo methods; spatial statistics; forest inventories; tessellations; stratified sampling; imperfections

Published in

Research report (Centre of Biostochastics)
2003, Volume: 2003, number: 2, pages: 1-82
Publisher: SLU

    SLU Authors

    Permanent link to this page (URI)