Mapping East African tropical forests and woodlands — A comparison of classifiers

In mapping the forest–woodland–savannah mosaic of Budongo Forest Reserve, Uganda, four classification methods were compared, i.e. Maximum Likelihood classifier (MLC), Spectral Angle Mapper (SAM), Maximum Likelihood combined with an Expert System (MaxExpert) and Spectral Angle Mapper combined with an Expert System (SAMExpert). The combination of conventional classifiers with an Expert System proved to be an effective approach for forest mapping. This was also the first time that the SAMExpert had been used in the mapping of tropical forests. SAMExpert not only maps with high accuracy, but is also fast and easy to use, making it attractive for use in less developed countries. Another advantage is that it can be executed on a standard PC set up for image processing.Combining the conventional classifiers (MLC and SAM) with the Expert System significantly improved the classification accuracy. The highest overall accuracy (94.6%) was obtained with SAMExpert. The MaxExpert approach yielded a map with an accuracy of 85.2%, which was also significantly higher than that obtained using the conventional MLC approach.The SAMExpert classifier accurately mapped individual classes. Of the four classes of woodland mapped, the Open Woodland (with Terminalia) and Wooded Grassland classes were more accurately mapped using SAMExpert. The Open Woodland had been previously identified by ecologists, but had never been mapped.