Influence of Soil Type on the Reliability of the Prediction Model for Bioavailability of Mn, Zn, Pb, Ni and Cu in the Soils of the Republic of Serbia

The principles of sustainable agriculture in the 21st century are based on the preservation of basic natural resources and environmental protection, which is achieved through a multidisciplinary approach in obtaining solutions and applying information technologies. Prediction models of bioavailability of trace elements (TEs) represent the basis for the development of machine learning and artificial intelligence in digital agriculture. Since the bioavailability of TEs is influenced by the physicochemical properties of the soil, which are characteristic of the soil type, in order to obtain more reliable prediction models in this study, the testing set from the previous study was grouped based on the soil type. The aim of this study was to examine the possibility of improvement in the prediction of bioavailability of TEs by using a different strategy of model development. After the training set was grouped based on the criteria for the new model development, the developed basic models were compared to the basic models from the previous study. The second step was to develop models based on the soil type (for the eight most common soil types in the Republic of Serbia—RS) and to compare their reliability to the basic models. From the total number of developed models by soil type (80), 75% were accepted as statistically reliable for predicting the bioavailability of TEs by soil type and 70% of prediction models had a higher determination coefficient (R2), compared to the basic models. For the Fluvisol soil type, all prediction models were accepted, while the least reliable prediction was for the Planosol type. As in the previous study of bioavailability prediction for TEs, the prediction models for Cu stood out, with more than half of the models with R2 greater than 0.90. Results of this study indicated that the formation of a testing set by soil type derives models whose predictions are more reliable than the basic ones. To improve the performance of prediction models, it is necessary to include additional physicochemical parameters and to conduct an adequate analysis of extensive testing sets with more comprehensive statistical techniques.