Knowledge-guided machine learning for improving crop yield projections of waterlogging effects under climate change

Abstract

Extreme precipitation poses a significant threat to crop production, often underestimated by process-based models. State-of-the-art models also struggle with high-resolution spatial applications due to process complexity. Here, we developed a Knowledge-Guided Machine Learning (KGML) framework that integrates machine learning with a waterlogging-enabled APSIM (Agricultural Production Systems sIMulator) to simulate wheat yield change under climate change in the Yangtze River Basin, China. Using transfer learning, this KGML framework transferred waterlogging processes to eight gridded crop models, enabling more accurate yield projections. We found that KGML could accurately replicate the behavior of the improved APSIM model under waterlogging conditions, achieving an R2 of 0.83 and an RMSE of 272.3 kg/ha for yield loss simulations. Soil properties were identified as the primary factors influencing yield losses under waterlogging, highlighting the importance of optimizing soil conditions to mitigate the adverse impacts of excessive water. Across different scenarios, the improved crop model ensembles projected greater crop yield losses compared to the original simulated outputs, with additional losses (compared to the historical period) around 5.9%–7.3% during the two periods. Although global climate models were the primary source of uncertainty in T1 (2029–2059), crop models contributed more to uncertainty in T2 (2069–2099). The improved ensemble reduced uncertainty from crop models compared to the original. This study highlights the potential of KGML to improve crop models, offering valuable insights for climate impact assessments and resource management. We believe our results can help national and local authorities make informed crop yield decisions under climate change.

Keywords

• Waterlogging
• Wheat yields
• Random forest
• Uncertainty analysis
• Crop model