Application of deep neural networks to forecast 60Co soil to plant transfer factor values based in pedological parameters

Abstract

The soil-plant transfer factor (Fv) is used methods in the computational models for radiological risk assessment by ingestion of radiocobalt-contaminated food. Different soil types, plants types and agricultural practices contribute to a wide dispersion of Fv values, indicating the need to study the criteria that influence root uptake in a regional view. In this scenario, Artificial Neural Networks (ANN) have become a possibility to predict Fv values based on critical pedological parameters. This work aims to apply ANN to evaluate the possibility of predicting Fv for 60Co in reference plants as a function of soil properties considered relevant for transfer processes in the soil-plant system. Through the systematic literature review, mineralogy, organic matter, texture, pH, CEC and nutrients were identified as soil properties that affect Fv values for 60Co. However, although these attributes were not always reported, still it was possible to create databases of Fv for 60Co in radish root and leaf, with pH, organic matter, and CTC as potential edaphic indicators. Learning sets were structured and due to the complexity of the search space and the small amount of available data, deep ANN with regularization (dropout) layers were required to achieve good prediction and avoid overfitting. The best model obtained showed good correlation in the validation and training set, considering the chosen parameters.