Modeling Reflectance of Urban Chestnut Trees: a Sensitivity Analysis of Model Inversion for Single Trees

Trees in urban environments have important ecological functions, but often suffer from suboptimal environmental conditions. Chestnut (Aesculus hippocastanum) is one of the main tree species in Germany’s capital Berlin, constituting 5% of all urban trees. Recently, chestnut trees are exposed to amplified stress levels due to the horse chestnut leaf miner (Cameraria ohridella), a parasitic insect whose larvae feed exclusively on chestnut leafs. Detecting plant damages at early stages and monitoring chestnut tree stress levels are thus important to manage urban canopies in Berlin and elsewhere. Hyperspectral data offer unique opportunities to derive detailed vegetation parameters for wide areas. Vitality parameters are commonly based on the characterization of plant physiology attributes and several studies point out the usefulness of radiative transfer modeling in this context, particularly for locally heterogeneous objects such as tree canopies. However, the selection of pixels that fully represent a single tree is not easy, because the model is highly sensitive to mixed pixels or varying illumination conditions. This is challenging, because the average chestnut crown diameter is only 15m and HyMap images offer at best a ground sampling distance of 4m, resulting in a low number of pure pixels. Tree geometry introduces further disturbance through diverse illumination constellations. In this research, we use a combination of the radiative transfer models PROSPECT and SAIL to derive parameters of chestnut tree stress levels for the city of Berlin based on HyMap images. Our goal was to quantify the influence of pixel selection methods on the quality of the model output. Reference data was collected for 11 trees and included leaf area index, leaf chlorophyll-, waterand, dry matter content. Different methods and statistical approaches to select input pixels for the modeling were compared based on several statistical parameters. Our results clearly highlight that choosing the optimum selection method is a crucial step in the model parameterization, and different selection methods lead to significantly different model results. In our case, using an averaged signal of the illuminated crown areas results in the best agreement between the model output and reference data.