📄 agriculture

Projecting maize yield in the North China Plain under climate extremes: A hybrid crop model-machine learning approach

This study develops a hybrid framework integrating the APSIM process-based model with machine learning algorithms and feature selection techniques to significantly improve maize yield prediction accuracy in the North China Plain, revealing that future climate extremes will cause substantial yield losses even when accounting for the positive effects of elevated CO₂.

Original authors: Haonan Tan, Dengpan Xiao, Huizi Bai, Yang Lu, Lihua Lv, Peipei Pan, Dandan Ren, Man Zhang, Xiaomeng Yin, Yan Wang, Ying Li, Chen Wang, Maoyuan Sun

Published 2026-08-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Haonan Tan, Dengpan Xiao, Huizi Bai, Yang Lu, Lihua Lv, Peipei Pan, Dandan Ren, Man Zhang, Xiaomeng Yin, Yan Wang, Ying Li, Chen Wang, Maoyuan Sun

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to predict how a giant, hungry plant will do in a garden that's getting hotter and wilder every year. This is the job of agricultural scientists who study crops like maize (corn). They use two main tools to guess the future harvest. The first tool is a "recipe book" called a crop model. It's like a computer program that knows the rules of biology: how much water a plant needs, how sunlight turns into food, and how roots drink from the soil. It's great at following the rules, but sometimes it gets confused when the weather goes crazy, like during a sudden, scorching heatwave or a massive flood. The second tool is a "super-learner" called machine learning. This is a type of artificial intelligence that looks at mountains of past data to find hidden patterns, like a detective spotting clues that a human might miss. It's amazing at guessing outcomes based on history, but it doesn't really understand why the plant is doing what it's doing. The big question scientists are asking is: Can we combine the strict rules of the recipe book with the sharp intuition of the super-learner to get a perfect prediction? This is especially important because extreme weather is becoming more common, and if we can't predict how much food we'll have, it could hurt our food supply.

Now, let's look at what this specific study did. A team of researchers in China decided to build a "hybrid" system to predict maize yields in the North China Plain, a massive farming region that feeds millions. They started with their recipe book, the APSIM model, which simulated how spring and summer maize grow. But they knew this model alone wasn't perfect at handling extreme weather, so they added a layer of machine learning. They taught two different AI detectives—Random Forest and Light Gradient Boosting Machine—to look at the computer's simulation results and then correct them based on 22 different "extreme weather clues." These clues included things like how many days the temperature stayed above a certain limit, how long a drought lasted, or how many heavy rainstorms hit during specific parts of the plant's life. To make sure the AI didn't get overwhelmed by too much information, they used a smart filter (a genetic algorithm) to pick only the most important clues.

The results were like finding a secret weapon. The hybrid model, which combined the crop rules with the AI detective work, was much better at guessing the actual harvest than the recipe book alone. In fact, for spring maize, the hybrid model explained 94% of the changes in yield, and for summer maize, it explained 93%. The old model alone only explained about 64% to 68%. The researchers found that the AI was particularly good at spotting when extreme heat or drought would really hurt the crop, something the standard model tended to underestimate.

When they used this super-powered hybrid model to look into the future, the news was a bit sobering. They simulated what would happen under different climate scenarios, including a very high-emission future where the world gets much hotter. Without any help from rising carbon dioxide levels (which can sometimes help plants grow), the model predicted that spring maize yields could drop by as much as 43.6% and summer maize by 26.8% by the end of the century. Even when they included the "fertilizer effect" of extra carbon dioxide, which does help plants a bit, the hybrid model still predicted significant losses: 25.0% for spring maize and 24.6% for summer maize. The study suggests that while extra carbon dioxide helps, it isn't enough to cancel out the damage caused by extreme heat and weather. The hybrid model showed that the risk of losing crops is actually higher than the standard models predicted, meaning we might be in for a tougher time than we thought if we don't adapt our farming strategies.

Technical Summary: Projecting Maize Yield in the North China Plain under Climate Extremes

Problem Statement
Process-based crop models, such as the Agricultural Production Systems Simulator (APSIM), are established tools for assessing climate change impacts on agriculture. However, they often struggle to accurately quantify the effects of extreme climate events (e.g., heatwaves, droughts, heavy precipitation) on crop yields due to simplified representations of stress mechanisms. Conversely, statistical and machine learning (ML) models excel at capturing complex, non-linear relationships between climate and yield but often lack the physiological and ecological process understanding inherent in mechanistic models. The North China Plain (NCP), a critical maize-producing region, faces increasing frequency and intensity of extreme weather, necessitating a robust framework that overcomes the limitations of standalone models to accurately forecast future yield risks.

Methodology
The study developed a hybrid simulation framework integrating the APSIM-maize model with machine learning algorithms (Random Forest [RF] and Light Gradient Boosting Machine [LGBM]) and feature selection techniques.

Study Area and Data: The research focused on 51 agricultural meteorological stations across the NCP, covering both spring and summer maize growing regions. Historical data (1981–2010) included meteorological variables, soil properties, and crop management practices. Future projections utilized 27 Global Climate Models (GCMs) from CMIP6 under two Shared Socioeconomic Pathways (SSP245 and SSP585) for the 2040s and 2080s.
Extreme Climate Indices (ECIs): Twenty-two ECIs were constructed across three key growth stages (sowing-to-jointing, jointing-to-flowering, flowering-to-maturity). These indices categorized extreme high temperatures, drought (using SPI and SPEI), and extreme precipitation events.
Feature Selection: Two methods were employed to identify the most informative predictors from the 22 ECIs:
- Stepwise Regression (SR): A linear method to eliminate redundant predictors.
- Genetic Algorithm (GA): A non-linear optimization method inspired by natural selection to find optimal variable subsets.
Hybrid Model Construction: Six hybrid models were constructed by combining APSIM outputs (phenology, biomass, yield) with selected ECIs as inputs for RF and LGBM algorithms. The models were: APSIM+RF+SR, APSIM+RF+GA, APSIM+RF+All, APSIM+LGBM+SR, APSIM+LGBM+GA, and APSIM+LGBM+All.
Evaluation: Model performance was assessed using the coefficient of determination ( $R^2$ ), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE) against observed yield data. The best-performing model was subsequently used to project future yields under scenarios with and without $CO_2$ fertilization effects.

Key Results

Model Performance: The standalone APSIM model explained 64% of yield variability for spring maize and 68% for summer maize. All hybrid models significantly outperformed the standalone APSIM. The APSIM+RF+GA model achieved the highest accuracy, explaining 94% of spring maize yield variation and 93% of summer maize yield variation. Compared to APSIM alone, this hybrid approach improved simulation accuracy by 30% for spring maize and 31% for summer maize.
Feature Selection Efficiency: The Genetic Algorithm (GA) consistently identified a smaller subset of predictors (12 ECIs) that maintained high predictive accuracy, demonstrating that non-linear feature selection effectively mitigates the "curse of dimensionality" without sacrificing performance.
Future Yield Projections:
- Under the high-emission SSP585 scenario for the 2080s, the hybrid model projected significantly larger yield losses than the standalone APSIM. Without $CO_2$ fertilization, losses were projected at 43.6% for spring maize and 26.8% for summer maize.
- When $CO_2$ fertilization effects were included, the hybrid model still predicted substantial losses of 25.0% (spring) and 24.6% (summer).
- The standalone APSIM consistently underestimated yield losses compared to the hybrid model, particularly under high-emission scenarios, suggesting it fails to fully capture the severity of extreme climate impacts.
Spatial and Temporal Trends: Yield reductions are projected to be more severe in the 2080s than the 2040s and under SSP585 compared to SSP245. Spatial heterogeneity indicates that central and southeastern plains face the most severe risks.

Significance and Claims
The authors claim that the developed hybrid framework addresses the inherent limitations of standalone process-based models in capturing the non-linear and severe impacts of extreme climate events on crop productivity. By integrating mechanistic understanding (APSIM) with data-driven pattern recognition (RF/LGBM) and optimized feature selection (GA), the study provides a more reliable tool for forecasting future crop yields.

Key conclusions include:

Superior Risk Assessment: The hybrid model reveals greater yield reduction risks than traditional models, indicating that relying solely on process-based models may lead to overly optimistic projections of future agricultural resilience.
Limitations of $CO_2$ Fertilization: While elevated $CO_2$ concentrations exert a positive effect on maize yield, the study finds this benefit is insufficient to offset the negative impacts of climate change and extreme weather events.
Strategic Value: The framework offers a scientific basis for formulating regional agricultural adaptation strategies, such as breeding stress-resistant varieties and adjusting sowing dates, specifically tailored to the distinct extreme climate threats facing spring and summer maize in the NCP.

More like this