Knowledge-based classification of remote sensing data for the estimation of below- and above-ground organic carbon stocks in riparian forests

Floodplain forests play a crucial role in the storage of organic carbon (C_org). However, modeling of carbon stocks in these dynamic ecosystems remains inherently difficult. Here, we present the spatial estimation of C_org stocks in riparian woody vegetation and soils (to a depth of 1 m) in a Central European floodplain using very high spatial resolution remote sensing data and auxiliary geodata. The research area is the Danube Floodplain National Park in Austria, one of the last remaining wetlands with near-natural vegetation in Central Europe. Different vegetation types within the floodplain show distinct capacities to store C_org. We used remote sensing to distinguish the following vegetation types: meadow, reed bed and hardwood, softwood, and cottonwood forests. Spectral and knowledge-based classification was performed with object-based image analysis. Additional knowledge rules included distances to the river, object area, and slope information. Five different classification schemes based on spectral values and additional knowledge rules were compared and validated. Validation data for the classification accuracy were derived from forest inventories and topographical maps. Overall accuracy for vegetation types was higher for a combination of spectral- and knowledge-based classification than for spectral values alone. While water, reed beds and meadows were clearly detectable, it remained challenging to distinguish the different forest types. The total carbon storage of soils and vegetation was quantified using a Monte Carlo simulation for all classified vegetation types, and the spatial distribution was mapped. The average storage of the study site is 428.9 Mg C ha⁻¹. Despite certain difficulties in vegetation classification this method allows an indirect estimation of C_org stocks in Central European floodplains.