Assessing the robustness and generalizability of machine learning models for predicting selenium content in rice: a case study from the Pearl River Delta and Eastern Guangdong, China.

Abstract

Crop selenium uptake, influenced by complex factors, has prompted extensive research to predict the Se content in crop grains, leading to the development of various prediction methods. However, the practical application of these models is limited by geographical constraints and variations in independent variables. This study selected two distinct regions in Guangdong Province, China: the Pearl River Delta (PRD), a Quaternary plain region, and Heyuan, a hilly region characterized by outcrops of clastic rocks. A total of 205 paired rice and rhizosphere soil samples (PRD: 2016) and 60 paired samples (Heyuan: 2023) were collected to assess model robustness and generalizability. The results showed that 82.93% and 30.00% of soil Se ≥ 0.40 mg/kg and 72.68% and 38.33% of rice grain Se content ≥ 0.04 mg/kg were found in the PRD and Heyuan, respectively. However, no significant positive correlation was observed between soil Se and rice grain Se content in either area. Further studies found that the main influencing factors of rice grain Se content were soil SiO₂, Al₂O₃, total organic carbon (TOC), S, and pH. The model was applied to the dataset for both time periods separately, yielded strong results, indicating that the model is robust and does not fluctuate greatly with the time of sample collection. The five feature subsets were used to predict the two regions separately with significant results. This indicates that the subset of predictive model features is highly generalizable, and the differences in the lithology of the soil parent materials and topography do not significantly affect the prediction results.