Is Retrieval of Forest Biochemical Traits Stable over Variety of Environmental Conditions?
|Druh||Další prezentace na konferencích|
|Popis||A network of 15 small forested catchments established in the Czech Republic (the GEOMON network) has been used since 1994 for a long-term monitoring of the input-output element budgets and to study the ecosystem response to anthropogenic pressures (e.g., air pollution and soil acidification). The small catchments (22 – 260 ha) are predominantly covered by Norway spruce (Picea abies) forests, they were exposed to different levels of air pollution in the past and represent variety of environmental conditions (e.g., altitude, bedrock, soil types, forest age). During summer 2017 we collected airborne hyperspectral images with the CASI and SASI spectroradiometers (400 – 2400 nm, 162 bands with band width varying between 10 and 15 nm, pixel size of 1 m) and complementary field data on leaf chlorophyll, carotenoid, and water contents and leaf mass per area for 10 out of 15 watersheds (in total 120 trees were sampled at 40 subplots). Taking the advantage of such a rich dataset, the main objective was to optimize the retrieval of biochemical traits for varying conditions in coniferous stands and to test the retrieval stability in respect of varying environmental, forest and data acquisition conditions. Retrievals were optimized for coniferous stands by employing 3D radiative transfer model DART that allowed simulation of look-up tables for variety of forest and environmental conditions (variable biochemical traits, tree height, tree density, leaf area index, topography, and forest floor), as well as for different image acquisition conditions (i.e., illumination geometry). The look-up tables were used to train a machine learning algorithm, support vector regression (SVR) in combination with feature extraction to retrieve biochemical parameters from the airborne hyperspectral images. The field data on leaf biochemical traits indicated that there were no statistically significant differences between the catchments, nonetheless the variation between the subplots is large and it varies from 35 to 60 ug cm-2 for leaf chlorophyll content, from 4.6 to 7.6 ug cm-2 for leaf carotenoid content, from 0.018 to 0.026 g cm-2 for leaf water content, and from 0.014 to 0.024 g cm-2 for leaf mass per area. The preliminary results of retrieved traits from airborne hyperspectral images show systematic overestimation of leaf chlorophyll content and the most stable estimates across the catchments were obtained for leaf water content. The most influential feature of the retrievals was the terrain topography.|