Optimizing rice in-season nitrogen topdressing by coupling experimental and modeling data with machine learning algorithms

The modern in-season crop N recommendation approaches should have high reliability in promoting agricultural sustainability. These approaches are relevant to soil properties, meteorological conditions, management prac-tices, and crop in-season growing status. This study aims to use machine learning (ML) algorithms to incorporate the above variables as well as the field reactive nitrogen (N) losses (i.e., N damage cost) simulated by a DeNi-trification-DeComposition (DNDC) model to develop a new strategy for optimizing rice in-season topdressing N (TN) usage. Rice field experiments with multiple N treatments and rice varieties were carried out during 2015-2021 at four study sites in eastern China. Four ML algorithms, namely random forest regression (RFR), support vector regression (SVR), lasso regression (LSR), and partial least square regression (PLSR) were used to develop in-season prediction models of yield and reactive N losses by combining soil, meteorological, and management data with crop remote sensing data. The observed in-season agronomic optimum N rates (AONR) that can maximize rice yield at different sites were in the range of 116.5 to 177.4 kg N ha -1, while the in-season economic optimum N rates (EONR) that can maximize marginal revenue (i.e., yield income minus N fertilizer costs and N damage costs) were in the range of 97.4 to 163.6 kg N ha -1. The developed ML models were further used to simulate yield and marginal revenue responses to a series of assumed TN rates (0-300 kg N ha -1, gradient = 20 kg N ha -1). Comparably, RFR model and SVR model were more suitable for determining optimum TN rates, because their simulated response curves of yield and marginal revenue fit the normal regulation (linear plus plateau or single-peak shapes). Independent validation results showed that the in-season AONR and EONR predicted by RFR and SVR well accorded with the observed values (R2 >= 0.64, RRMSE <= 18.3 %), and the accuracy of ML models containing both historical and in-season meteorological information is superior to ML models that contain in-season meteorological information only. The proposed ML-based strategy is expected to help the regional rice production systems precisely manage N use, improve net profits, and reduce environmental footprints.