Upscaling with the dynamic two-layer classification concept (D2C): TreeMig-2L, an efficient implementation of the forest-landscape model TreeMig
- 1Department of Environmental Systems Science, Swiss Federal Institute of Technology ETH, 8092 Zürich, Switzerland
- 2Dynamic Macroecology, Landscape Dynamics, Swiss Federal Institute for Forest, Snow and Landscape Research WSL, Zürcherstrasse 111, 8903 Birmensdorf, Switzerland
- anow at: Max Planck Institute for Meteorology, Bundesstrasse 53, 20146 Hamburg, Germany
Abstract. Models used to investigate impacts of climatic changes on spatio-temporal vegetation dynamics need to balance required accuracy with computational feasibility. To enhance the computational efficiency of these models, upscaling methods are required that maintain key fine-scale processes influencing vegetation dynamics. In this paper, an adjustable method – the dynamic two-layer classification concept (D2C) – for the upscaling of time- and space-discrete models is presented. D2C aims to separate potentially repetitive calculations from those specific to single grid cells. The underlying idea is to extract processes that do not require information about a grid cell's neighbourhood to a reduced-size non-spatial layer, which is dynamically coupled to the original two-dimensional layer. The size of the non-spatial layer is thereby adaptive and depends on dynamic classifications according to pre-specified similarity criteria.
I present how D2C can be used in a model implementation on the example of TreeMig-2L, a new, efficient version of the intermediate-complexity forest-landscape model TreeMig. To discuss the trade-off between computational expenses and accuracy, as well as the applicability of D2C, I compare different model stages of TreeMig-2L via simulations of two different application scenarios. This comparison of different model stages demonstrates that applying D2C can strongly reduce computational expenses of processes calculated on the new non-spatial layer. D2C is thus a valuable upscaling method for models and applications in which processes requiring information about the neighbourhood constitute the minor share of the overall computational expenses.