Interpolating paleovegetation data with an artificial neural network approach

To drive an atmospheric general circulation model (AGCM), land surface boundary conditions like albedo and morphological roughness, which depend on the vegetation type present, have to be prescribed. For the late Quaternary there are some data available, but they are still sparse. Here an artificial neural network approach to assimilate these paleovegetation data is investigated. In contrast to a biome model the relation between climatological parameters and vegetation type is not based on biological knowledge but estimated from the available vegetation data and the AGCM climatology at the corresponding locations. For a test application, a data set for the modern vegetation reduced to the amount of data available for the Holocene climate optimum (about 6000 years B.P.) is used. From this, the neural network is able to reconstruct the complete global vegetation with a kappa value of 0.56. The most pronounced errors occur in Australia and South America in areas corresponding to large data gaps.