Our goal is to develop advanced techniques for tree species mapping based on hyperspectral remote sensing in combination with other remote sensing and in situ measurements that meet the demands of the National Park. This approach needs to be adapted to the heterogeneous appearance of the forest.

This work aims at building a model transferable to an area-wide mapping of tree species based on the needs of the Bavarian Forest National Park. It reveals the requirements for tree species mapping and shows which spectral/spatial features and data combinations generate the best results within a Random Forest modelling approach.

The study is based on airborne hyperspectral data acquired with the HySpex VNIR-1600 sensor (160 spectral bands, 400-990 nm, 1.6 m spatial resolution). Additional full waveform LiDAR data, including a Digital Surface Model, Digital Terrain Model and a Digital Canopy Height Model, were available for the analysis. Individual tree crowns as well as clusters of tree crowns from 13 different tree species were located and identified during a field survey. The field-demarcated tree canopies were used as reference data for creating the feature database.

Several preprocessing steps including atmospheric correction, spectral and spatial polishing, bidirectional reflectance distribution function (BRDF) effect correction as well as ortho-rectification of the hyperspectral imagery were conducted before the analysis. A band selection procedure based on principal component analysis, band correlation, and band variance was performed to identify the most appropriate spectral bands for species discrimination, resulting in a set of 53 spectral bands. Seven different combinations of hyperspectral, structural and terrain-specific parameters contained in the feature database were investigated in a Random Forest Modelling approach to ascertain which variables enhance the overall classification accuracy. A classification model using all available parameters in the feature database yielded an overall accuracy that is 17 percentage points higher (94%) compared to using only the preselected spectral bands (77%). For most of the 13 tree species, the final classification model achieved individual class accuracies of more than 90%.