Using Decision Tree Based Multiclass Support Vector Machines for Forest Mapping

The goal of this study is to develop an automatic supervised classification strategy and to find the necessary data sources to perform automatic tree species classification on single tree level for a large area. The derived forest map is used in a virtual forest testbed to estimate additional forest parameters like diameter at breast height and volume at single tree and stand level. A virtual forest database is populated with the calculated data and further simulations like estimating the effect of cuttings or future forest development can be implemented.. To achieve the goals, a support vector machine based decision tree that uses a decision tree structure and a support vector machine at each tree node is presented for tree species classification. It is compared to a manually deduced decision tree. Further improvements due to the use of additional input data are presented and the results are compared. These improvements include the use of LIDAR height and intensity data and the comparison of results calculated based on SPOT and RapidEye satellite data, which differ not only in resolution but also in available spectral bands. Seven tree species are discriminated, namely oak, beech, pine, larch, spruce, Douglas fir and a group of other deciduous tree species including birch and alder.