Use of random forest algorithm for predictive modelling of transfer factor soil-plant for radiocaesium: A feasibility...

A German dataset with soil-plant transfer factors for radiocaesium including many co-variables was analysed and prepared for the application of the Random Forest (RF) algorithm using the R libraries 'party', and 'caret'. A RF predictive model for soil-plant transfer factor was created based on 10 co-variables. These are, for example, taxonomic plant family, plant part, soil type and the exchangeable potassium concentration in the soil. The RF model results were compared with the results of two (semi-)mechanistic models. Of the more than 3000 entries in the original dataset, only about 1200 could be used, as this was the largest complete dataset with the largest number of co-variables available. The obtained RF predictive model can reproduce the experimental observations better than the two (semi)-mechanistic models, which are based on many assumptions and fixed parameter values. Model performance was quantified using the metrics of Root Mean Square Error (rmse) and Mean Absolute Error (mae). The RF model was able to reproduce the variability of the data by up to 6 orders of magnitude. The categorical co-predictors, especially taxonomic plant family and plant part, have a greater influence than the numerical co-predictors, such as pH and exchangeable soil potassium concentration. This feasibility study shows that RF is a promising tool to obtain predictive models for transfer factors. However, to build a widely applicable predictive model, a dataset is needed that contains at least thousands of entries for transfer factors and for the most important co-variables and considers a large parameter space.