Mapping of spruce and pine fractional coverage at 1 ha resolution for entire Bavaria

In Central Europe spruce and pine are severely affected by the impacts of climate change. In several regions a significant decline in their distribution is observed. To cope with this threat, the Bavarian State Institute of Forestry produces climate risk maps for the next decades. To be effective however, locational information of the two tree species is required. Such basic information is not yet available at the necessary spatial resolution. The aim of this study is to generate distribution maps for spruce and pine for entire Bavaria (70.500 km2) at a resolution of about one hectare. For each hectare cell, the fraction coverage of the two tree species is to be specified as well as associated uncertainties. In order to meet these user-defined requirements, a two-step methodology combining satellite imagery at metric to deca-metric resolution was developed. In a first step, tree species maps with a high level of detail were generated from 8-band multispectral WorldView-2 data with 0.5 to 2.0 m spatial resolution. As reference, inventory data from the Bavaria State Forest enterprise was used. Where necessary, additional reference samples were derived from stereo interpretation of aerial images. From this data, detailed tree species maps were generated for roughly 40 sites (each about 100 km2 large) well distributed across Bavaria. For the object-based mapping, spectral information and textural indices were used. The textural measures were generated at several scales with a discrete stationary wavelet transformation (using Red, Near Infrared and NDVI as inputs). The classification itself was performed using Random Forest (RF). Features used in the classification were selected by means of RF’s importance measures. The generated tree species maps were used in a second step as reference information (targets) to generate the fractional coverage maps for the entire country using neural nets. For the upscaling, Landsat multi-temporal data complemented by high resolution RapidEye imagery was used as predictor variables. From the sites with detailed tree species maps, data from these two satellite sensors were extracted and used to train a neural network for estimating the fractional coverage of the two tree species. After network training, the models were applied to the entire area