Tree Species Classification in the Bavarian Forest National Park using Hyperspectral Remote Sensing and Site Specific Information

The Bavarian Forest National Park is a unique area of forest stands that developed with low anthropogenic interference into a landscape with remnants of a primeval forest. There is a high interest of the National Park authority to investigate the composition and development of the forest ecosystem. However conventional forest inventories are time-consuming and focus mainly on the forests economic value. The goal of this study is to develop advanced mapping techniques combining remote sensing and in-situ information that meet the demands of the National Park and emphasize the ecological value of a forest. This study is based on a large scale hyperspectral survey flown in 2013. 16 000 hectares of HySpex VNIR1600 and SWIR 320i data were acquired covering 65% of the whole National Park. A previous hyperspectral species mapping exercise was successfully performed using the HySpex VNIR data only. This new approach was now significantly improved using the additional spectral information of the HySpex SWIR sensor as well as stand density information. In addition to the remote sensing data, forest inventory data from 2002 were provided by the National Park and used for validation and partly as additional training data. Clusters of tree crowns from 11 different tree species were located and identified during a field survey in November 2014 and December 2015. This field data was used as training and validation set. By overlaying the pre-selected and field-demarcated tree canopies with the hyperspectral data the tree species spectral libraries were created and used as input for the analysis. Different input parameters and predictor datasets, consisting of spectral and structural data were investigated to enhance the Random Forest classification model. The impact of spectral coverage, geometric resolution and training data on the quality of classification result was also evaluated by comparing previous and new classification results. This work aims at building a site specific classification model transferable to an area wide mapping approach based on the needs of the Bavarian Forest National Park, which will in future be extended to the Šumava National Park. This study reveals the requirements for tree species mapping and shows which spectral/spatial features and data composition generate the best results