Synergizing process-based modeling and data-driven learning for precision nitrogen optimization in winter wheat

CONTEXT Wheat requires effective nitrogen (N) management to ensure food security, improve resource use efficiency, and mitigate environmental pollution. However, coordinating high yields, resource-use efficiency, and low environmental footprints at large scales remains a significant challenge due to high computational demands and the complexity of climate-crop-soil management interactions. OBJECTIVE This study develops a simulation-based decision-support framework for regional precision N management in winter wheat. By integrating crop modeling, machine learning (ML), and multi-objective optimization, the framework explores candidate basal and topdressing N application strategies under multiple agronomic, economic, and environmental objectives. METHODS Focusing on winter wheat in Jiangsu Province, this study calibrated and evaluated the APSIM model using multi-year, multi-site field experimental data. To enable efficient spatiotemporal simulations, an APSIM surrogate model was developed by evaluating four classical ML algorithms (Lasso, RF, SVR, and XGBoost). The surrogate model was then coupled with the Non-dominated Sorting Genetic Algorithm II (NSGA-II) and the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) to identify a baseline compromise N strategy based on a comprehensive assessment of yield, economic, and environmental objectives. RESULTS AND CONCLUSIONS Model evaluation showed that APSIM reproduced wheat phenology reasonably well and achieved moderate performance for yield prediction during independent evaluation (R 2 = 0.55; RMSE = 0.82 t ha −1 ). Among the tested ML algorithms, XGBoost most accurately emulated APSIM-simulated outputs, achieving high performance for yield (R 2 = 0.86, RMSE = 0.53 t ha −1 ) and the APSIM-simulated N 2 O emission indicator (R 2 = 0.81, RMSE = 0.34 kg ha −1 ). In a descriptive comparison with literature-derived benchmarks, the model-derived median total N input under MONM was 191.09 kg N ha −1 , which was 21.79% and 17.96% lower than the FP (244.32 kg N ha −1 ) and EONM (232.91 kg N ha −1 ) benchmark medians, respectively. The median simulated yield under MONM was 6.57 t ha −1 , 8.77% higher than the FP benchmark and 6.01% lower than the EONM median (6.99 t ha −1 ). The median RERI was 17.53, corresponding to values 93.92% and 68.07% higher than the FP (9.04) and EONM (10.43) benchmark medians, respectively. The median APSIM-simulated N 2 O indicator was 0.75 kg ha −1 , corresponding to values 56.82% and 46.99% lower than the FP and EONM benchmarks, respectively. SIGNIFICANCE This integrated framework provides a computationally efficient approach for exploring trade-offs among productivity and environmental sustainability. It supports the rapid screening of candidate N management strategies at the regional scale and provides a simulation-based basis for decision-support development.