|Paper title||Towards globally applicable Spatial Prediction Models|
|Form of presentation||Poster|
Global-scale maps provide a variety of ecologically relevant environmental variables to researchers and decision makers. Usually, these maps are created by training a machine learning algorithm on field-sampled reference samples and the application of the resulting model to associated remote sensing based information from satellite imagery or globally available environmental predictors. This approach is based on the assumption, that the predictors are a representation of the environment and that the machine learning model can learn the statistical relationships between the environment and the target variable from the reference data.
Since field samples are often sparse and clustered in geographic space, machine learning based mapping requires, that models are transferred to regions where no training samples are available. Further, machine learning models are prone to overfit to the specific environments they are trained on, which can further contribute to poor model generalization. Consequently, model validations have to include an analysis of the models transferability in regions where no training samples are available e.g. by computing the Area of Applicability (AOA, Meyer and Pebesma 2021).
Here we present a workflow to optimize the transferability of machine learning based global spatial prediction models. The workflow utilizes spatial variable selection in order to train generalized models which include only predictors that are most suitable for predictions in regions without training samples.
To evaluate the proposed workflow we reproduced three recently published global environmental maps (global soil nematode abundances, potential tree cover and specific leaf area) and compared the outcomes to the original studies in terms of prediction performance. We additionally assessed the transferability of our models based on the AOA and concluded that by reducing the predictors to those relevant for spatial prediction, we could greatly increase the AOA of the models with negligible decrease of the prediction quality.
Meyer, H. & Pebesma, E. Predicting into unknown space? Estimating the area of applicability of spatial prediction models. Methods in Ecology and Evolution 2041–210X.13650 (2021) doi:10.1111/2041-210X.13650.