Delimiting the spatio-temporal uncertainty of climate-sensitive forest productivity projections using Support Vector Regression
As climate change makes many traditional empirical growth approaches not functional for forest dynamics modelling, new climate-sensitive models are needed. However, using these newly developed models for extrapolation, such as predicting forest productivity for new areas or future scenarios is still...
Guardado en:
| Autores principales: | , |
|---|---|
| Formato: | article |
| Lenguaje: | EN |
| Publicado: |
Elsevier
2021
|
| Materias: | |
| Acceso en línea: | https://doaj.org/article/ff2b1272137b403392d607f03446971b |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| Sumario: | As climate change makes many traditional empirical growth approaches not functional for forest dynamics modelling, new climate-sensitive models are needed. However, using these newly developed models for extrapolation, such as predicting forest productivity for new areas or future scenarios is still a difficult task. In this study, we proposed a method for delimiting the uncertainty of climate-sensitive extrapolations of forest productivity (site index, SI) using the regularisation approach implicit in distance-based Support Vector Regression. As a case study, we predicted forest productivity with a dataset of 165 permanent research plots of radiata pine forests in Galicia (NW of Spain) as a function of bioclimatic variables from the Worldclim 2 raster datasets. The developed model was based on the radial basis kernel and, after calibrating it using cross-validation, produced adequate performance metrics, explaining up to 56% of the site index’ variability. Then, we predicted forest productivity for the Galician territory basing on climate raster maps for current conditions and six future scenarios (using different Global Climate Models) and evaluated the resulting maps by delimiting the surfaces with predictions strongly regressed to the mean. This analysis revealed that the extrapolations for unseen climatic conditions were extremely regularised, even for current climate, being 60–99% of the territory regressed to the observational site index mean. In other words, the validity area delimited for the fitted model was narrow in comparison with the prediction extent. These results imply that the climatic conditions in these areas/scenarios were too different from the training datastet for making reliable predictions, at least under the optimum model setup defined by cross-validation. However, when we reduced the σ parameter, responsible for controlling distance-based regularisation, we observed a noticeable increase in validity area of the model, together with a drop in performance. This fact revealed the existence of a trade–off between highly specific models, with high performance and a small applicability area, and more generalisable models, with a broad validity area but lower performance. We concluded that the tested methodology could be a useful starting point for assessing the spatio-temporal uncertainty of forest productivity predictions in the future. |
|---|