- 1Department of Natural Hazards, Austrian Research Centre for Forests, Innsbruck, Austria
- 2Department of Forest and Soil Sciences, University of Natural Resources and Life Sciences, Vienna, Austria
- 3Department of Forest Planning, Office of the Tyrolean Government, Innsbruck, Austria
Rising temperatures and drier conditions due to climate change will force Alpine forests towards their ecological limits. For an informed decision on climate-smart tree species composition, we need area-wide information about the current and future moisture regime in forest areas. For this task, soil-vegetation-atmosphere transport (SVAT) models in combination with digital mapping techniques have proven useful tools. However, depending on the selected mapping algorithm (e.g. random forest - RF, generalized additive models - GAM, neural networks - NN) and the selected train-/test-split of the input data, resulting maps can vary considerably. The method of splitting the dataset into training and test subsets can significantly impact model performance and spatial predictions, particularly when imbalanced data is present. Such imbalances in specific training/test splits can lead to inconsistencies in model development and their resulting spatial predictions. For generating maps that represent the moisture regime under current and future climate conditions, spatial consistency and reproducibility are crucial. Models must produce robust spatial patterns that are not merely artifacts of a single training/test split but reflect reliable and consistent predictions.
First, we use a lumped, physically-based SVAT model (LWF-Brook90) for reproducing the moisture regime at 2009 mapped forest sites in Tyrol and Vorarlberg (Austria). We parameterized the model with the individual soil characteristics at the sites, while considering a generic beech forest stand for sake of comparability between the sites. Based on interpolated meteorological observations and bias-corrected climate projections, we derived components of the water balance under current (1991-2020) and future conditions (2036-2065, 2071-2100) on a daily resolution. As an indicator for the moisture regime, the transpiration deficit (Tdef; i.e. the difference between potential and actual transpiration) was identified.
Using digital soil mapping techniques, we generate maps of the mean annual Tdef sum for the selected periods, incorporating geomorphometric and climate-related covariates. Feature selection is conducted using RF (based on feature importance) across multiple training/test splits, selecting the most commonly chosen features to build RF and NN models. GAM, by contrast, employs a smaller, expert-based set of covariates for improved interpretability. To ensure robustness, multiple runs are performed for each algorithm. Forming ensemble means prevents random biases from imbalanced training data, while deviation maps help identifying uncertainties in the mapping process.
Statistical metrics (e.g., R², RMSE) on an independent validation set reveal greater variation within a single algorithm than between different algorithms, complicating the identification of a "better" approach. To address this, we propose a weighted ensemble approach that accounts for performance on independent validation data, enabling reliable and spatially consistent outcomes. The resulting maps will aid in identifying suitable tree species under future climate conditions at the slope scale.
This work was carried out within the WINALP21 project, funded by the INTERREG VI-A program (grant number BA0100020).
How to cite: Huber, T., Zieher, T., Simon, A., Gadermaier, J., and Klebinder, K.: Derivation of robust moisture regime indicator maps in Alpine forests considering climate change – balancing uncertainty by multi-algorithm ensembles, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19804, https://doi.org/10.5194/egusphere-egu25-19804, 2025.