Mitigating nitrogen (N) loss from rice paddies is essential for enhancing rice productivity while reducing environmental impacts. This study evaluated two machine learning (ML) models (i.e., random forest and extreme gradient boosting; XGBoost) and one deep learning (DL) model (i.e., long short-term memory; LSTM) for predicting ammonia (NH 3 ) and nitrous oxide (N 2 O) emissions under varying fertilizer deep-placement (FDP) conditions. Models were developed using data from a controlled soil incubation experiment ( n = 285) designed to simulate flooded paddy conditions. The models were subsequently validated using an independent field data ( n = 42) collected during a 100-day rice cultivation period. Input features included fertilization depth, days after fertilization, soil properties, inorganic N content, and atmospheric temperature. Model performance was assessed using four metrics, including coefficient of determination (R 2 ) and root mean squared error (RMSE). Under controlled conditions, the ensemble ML models showed stable and reliable predictive performance, with low variability across cross-validation. In contrast, under field conditions characterized by greater environmental heterogeneity and temporal variability, the LSTM model demonstrated superior predictive performance, explaining up to 81% and 97% of the variability in N 2 O and NH 3 emissions, respectively, under field conditions. The LSTM model also provided reasonable estimates of cumulative NH 3 losses, achieving 94–97% agreement with observations. These results suggest that incorporating temporal dependencies is beneficial for capturing N loss dynamics under open-field conditions. Overall, the findings indicate that LSTM-based approaches can serve as a complementary tool for evaluating FDP effectiveness and supporting improved fertilization management in rice paddies.
Kang et al. (2026) studied this question.