Performance of statistical and machine learning-based methods for predicting biogeographical patterns of fungal productivity in forest ecosystems

Background The prediction of biogeographical patterns from a large number of driving factors with complex interactions, correlations and non-linear dependences require advanced analytical methods and modeling tools. This study compares different statistical and machine learning-based models for predicting fungal productivity biogeographical patterns as a case study for the thorough assessment of the performance of alternative modeling approaches to provide accurate and ecologically-consistent predictions. Methods We evaluated and compared the performance of two statistical modeling techniques, namely, generalized linear mixed models and geographically weighted regression, and four techniques based on different machine learning algorithms, namely, random forest, extreme gradient boosting, support vector machine and artificial neural network to predict fungal productivity. Model evaluation was conducted using a systematic methodology combining random, spatial and environmental blocking together with the assessment of the ecological consistency of spatially-explicit model predictions according to scientific knowledge. Results Fungal productivity predictions were sensitive to the modeling approach and the number of predictors used. Moreover, the importance assigned to different predictors varied between machine learning modeling approaches. Decision tree-based models increased prediction accuracy by more than 10% compared to other machine learning approaches, and by more than 20% compared to statistical models, and resulted in higher ecological consistence of the predicted biogeographical patterns of fungal productivity. Conclusions Decision tree-based models were the best approach for prediction both in sampling-like environments as well as in extrapolation beyond the spatial and climatic range of the modeling data. In this study, we show that proper variable selection is crucial to create robust models for extrapolation in biophysically differentiated areas. This allows for reducing the dimensions of the ecosystem space described by the predictors of the models, resulting in higher similarity between the modeling data and the environmental conditions over the whole study area. When dealing with spatial-temporal data in the analysis of biogeographical patterns, environmental blocking is postulated as a highly informative technique to be used in cross-validation to assess the prediction error over larger scales.