Presentation Information
[U02-P06]Machine Learning–Based Meta-Modeling of APEX for Efficient Prediction of Agricultural Nutrient Discharge under Future Climate Scenarios
*Seungwon Seok1, Jaeyoung Jung1, Hyunsoo Park1, Huiwon Yang2, Donghyeon Kim2, Taegon Kim1,2,3 (1.Department of Agricultural Engineering, College of Agricultural & Life Sciences, Jeonbuk National University, Jeonju, 54896, Korea, Republic of South Korea, 2.Department of Smart Farm, Jeonbuk National University, Jeonju, 54896, Republic of Korea, 3.Institute of Agricultural Science & Technology, Jeonbuk National University, Jeonju, 54896,)
Keywords:
Meta-modeling,Machine learning,Nutrient Discharge,Agricultural management,APEX,Climate change
Agricultural production relies on diverse crop management practices, nutrient losses from chemical fertilizer use remain a major source of agricultural nonpoint source pollution. The APEX model has been widely used to simulate nutrient transport under different agricultural management and climate conditions, supporting decision-making for nonpoint source pollution mitigation. However, the high computational cost and long simulation time of the APEX model limit its applicability to large-scale scenario analyses. This study aims to develop a machine learning–based surrogate meta-model capable of efficiently reproducing APEX simulation outputs, providing a foundation for applying future climate change scenarios. A synthetic dataset was generated using 40 years of meteorological data from five representative weather stations in South Korea and approximately 2.5 million agricultural management scenarios varying in nitrogen fertilization rates (0–200 kg/ha), slope gradients (0.0-0.40), and crop types (corn, cabbage, and rice), with rice simulated using the APEX-paddy module. Monthly meteorological inputs included precipitation characteristics, solar radiation, temperature, humidity, and wind speed. A RandomForest-based surrogate meta-model was developed to predict total nitrogen (TN) and total phosphorus (TP) runoff, and agricultural land-use layers were applied to assess climate change impacts separately for paddy and upland fields. The developed meta-model demonstrated excellent predictive performance across all crop types, achieving R² values of 0.99 for both total nitrogen (TN) and total phosphorus (TP). Scenario simulations using the surrogate meta-model were conducted approximately 800-1000 times faster than conventional APEX simulations, providing a practical decision-support framework for evaluating climate adaptation scenarios and optimizing nutrient management strategies across paddy and upland agricultural systems.
Acknowledgement: This work was carried out with the support of “New Agricultural Climate Change Response System Establishment (Project No. RS-2024-00396736).” Rural Development Administration, Republic of Korea.
Acknowledgement: This work was carried out with the support of “New Agricultural Climate Change Response System Establishment (Project No. RS-2024-00396736).” Rural Development Administration, Republic of Korea.