hide
Free keywords:
-
Abstract:
Dynamic Global Vegetation Models (DGVMs) with embedded tree demography and flexible plant traits simulate dynamics across tens of thousands of land grid cells, each containing tens of thousands of individual trees, making global simulations computationally expensive. Aggregating cells into climatically coherent clusters and running a single simulation per cluster reduce this cost by orders of magnitude, but cluster quality critically depends on an ecologically meaningful distance metric between cells. We present a framework in which spatially adaptive distance metrics are derived from Köppen–Geiger climate-zone labels via Large Margin Nearest Neighbour (LMNN) metric learning and used for hierarchical clustering. A multilayer perceptron predicts cell-specific positive semi-definite metric matrices, allowing feature weighting to adapt continuously across ecological regimes. We compare a learned metric that captures cross-feature interactions against a simpler variant without interaction terms and the uniform Euclidean metric. Experiments on 67,420 global land cells with 51 climate and soil features, testing cluster counts across five orders of magnitude, show that both learned metrics substantially improve agreement with Köppen–Geiger climate-zone boundaries compared to the Euclidean uniform metric. This improved ecological alignment translates into more accurate biomass reconstruction well beyond the training signal: the learned metrics outperform the Euclidean baseline across cluster counts spanning more than an order of magnitude above the 30-zone signal before all three metrics converge. For cluster counts 2132, the Euclidean metric performs better, showing that at this scale, different features down-weighted by the metric learning—such as soil properties—become more important. The variant without interaction terms achieves biomass reconstruction accuracy equal to the interaction-based metric with directly interpretable feature weights. Evaluated with the individual-based DGVM LPJmL-FIT, more than 80% of cell–year pairs stay within the model-inherent stochastic noise floor already at an energy consumption of only 0.1% of the full-grid cost.
