Source-linked AI summary
Coupling Machine Learning and Crop Modeling Improves Crop Yield Prediction in the US Corn Belt
Mohsen Shahhosseini, Guiping Hu, Sotirios V. Archontoulis, Isaiah Huber
TL;DR
Corn yield prediction in the US Corn Belt has mainly assessed crop modeling and machine learning separately, leaving the value of integrating them uncertain. This study evaluates hybrid APSIM–ML models and finds that adding APSIM variables improves prediction accuracy, especially through hydrological features.
Problem
Crop modeling and machine learning have largely been assessed separately, leaving limited systematic evidence on whether integrating them improves corn yield prediction.
Method
The study couples calibrated APSIM crop simulations with five machine-learning models and ensembles to predict corn yields across Illinois, Indiana, and Iowa.
Results
Adding APSIM simulation variables to machine-learning models improves corn yield prediction, with soil-moisture-related variables the most influential.
Takeaways & Limitations
Hydrological APSIM inputs, particularly drought stress and water-table depth, are key features for improving machine-learning-based corn yield predictions.
Abstract
from arXiv · showhide
This study investigates whether coupling crop modeling and machine learning (ML) improves corn yield predictions in the US Corn Belt. The main objectives are to explore whether a hybrid approach (crop modeling + ML) would result in better predictions, investigate which combinations of hybrid models provide the most accurate predictions, and determine the features from the crop modeling that are most effective to be integrated with ML for corn yield prediction. Five ML models (linear regression, LASSO, LightGBM, random forest, and XGBoost) and six ensemble models have been designed to address the research question. The results suggest that adding simulation crop model variables (APSIM) as input features to ML models can decrease yield prediction root mean squared error (RMSE) from 7 to 20%. Furthermore, we investigated partial inclusion of APSIM features in the ML prediction models and we found soil moisture related APSIM variables are most influential on the ML predictions followed by crop-related and phenology-related variables. Finally, based on feature importance measure, it has been observed that simulated APSIM average drought stress and average water table depth during the growing season are the most important APSIM inputs to ML. This result indicates that weather information alone is not sufficient and ML models need more hydrological inputs to make improved yield predictions.
1. Introduction
The study motivates integrating simulation crop modeling and machine learning because they have largely been assessed separately despite their distinct predictive capabilities. It investigates whether hybrid approaches improve corn yield prediction, which model combinations are most accurate, and which crop-model features are most relevant for ML.
- Motivation: Integrating simulation crop modeling and ML may further increase prediction accuracy because the technologies provide unique capabilities but have mainly been assessed separately.The paper frames this integration as a response to the limitations of evaluating the two technologies independently.
- Background: Simulation crop models use management, cultivar, and environmental inputs with science-based equations, whereas ML learns input–response relationships without researcher-specified transfer functions.Simulation models represent crop physiology, hydrology, and soil C and N cycling; ML is more easily applicable because it does not require expert calibration knowledge and skills.
- Related work: Prior hybrid research was limited beyond a few studies combining crop models with simple regression, although one hybrid model explained 67%, 72%, and 79% of yield variability for September 1st, January 1st, and March 1st predictions.The cited results illustrate existing but limited evidence for combining crop modeling with statistical or ML approaches.
- Research objectives: The paper investigates whether simulation crop modeling plus ML improves corn yield predictions in Illinois, Indiana, and Iowa, identifies the most accurate hybrid model combinations, and determines the most relevant crop-model features for ML.These are the study’s three stated research objectives.
2. Materials and Methods · 2.1. Agricultural Production Systems sIMulator (APSIM)
The study combines APSIM crop simulations with machine-learning models to predict corn yield, using a calibrated US Corn Belt APSIM maize model and geographically aggregated simulation outputs. APSIM data are preprocessed and feature-selected before evaluation across dry, average-to-wet, and test-year scenarios.
- 2. Materials and Methods: The methods describe a hybrid simulation–machine-learning framework, including model inputs, data processing, predictive algorithms, and evaluation metrics.
- 2.1.1. APSIM run details: APSIM is an open-source cropping-systems simulator whose crop, soil-water, carbon, nitrogen, and residue modules interact daily.
- 2.1.1. APSIM run details: The study uses APSIM maize version 7.9 calibrated for US Corn Belt environments, including shallow-water-table simulation and excess-water inhibition of root growth.
- 2.1.1. APSIM run details: APSIM simulations for Illinois, Indiana, and Iowa were generated with pSIMS on a 5-arcminute grid using cropland-only soil profiles.
- 2.1.2. APSIM output variables used as inputs to ML models: The APSIM outputs contain 22 variables at approximately 40 fields within a county, which were aggregated to county-level values for integration with USDA data.
- 2.1.2. APSIM output variables used as inputs to ML models: APSIM data processing included imputing zero values with feature averages, removing missing-value rows, scaling values between 0 and 1, and cross-referencing the developed data set.
- 2.1.2. APSIM output variables used as inputs to ML models: Feature-selection procedures were applied to retain APSIM variables carrying the most relevant information for prediction.
- 2.1.2. APSIM output variables used as inputs to ML models: Data from 1984 to 2018 were used, with 2012, 2017, and 2018 held out in turn as test years representing extremely dry and average-to-wet conditions.
2.2. Machine Learning (ML)
The ML models use 1984–2018 county-level corn yield observations from Illinois, Indiana, and Iowa, with environmental and management variables as inputs. Inputs include weather, soil, plant population, and planting progress, with min-max scaling applied before model fitting.
- Data and inputs: ML models predict corn yield using soil, weather, and management variables as inputs and observed corn yields as the target.The study covers Illinois, Indiana, and Iowa in the US Corn Belt.
- Data and inputs: The dataset spans 1984–2018 and contains 10,016 yearly average corn yield observations from 293 counties.County-level yields were obtained from USDA-NASS.
- Feature construction: Plant population and planting progress provide additional management-related inputs, comprising one annual plant-population feature and 52 weekly planting-date features.Plant population is measured in plants per square meter, while planting progress records the weekly cumulative percentage planted within each state.
- Feature construction: Weather inputs comprise 260 weekly features covering temperature, precipitation, growing degree days, and incident shortwave radiation.The weather variables are accumulated weekly and obtained from the Iowa Environmental Mesonet.
- Feature construction: Soil inputs comprise 180 features describing soil properties across different soil layers.The soil features include organic matter, sand and clay content, pH, bulk density, wilting point, field capacity, and saturation point.
- Preprocessing: Before fitting ML models, input data were scaled between 0 and 1 using min-max scaling.This preprocessing was intended to give inputs similar ranges, especially for weighted ensemble models.
• Add yearly trends feature
Corn yields showed an increasing trend over time that existing input features could not explain. To account for it, the study added a location-specific linear-regression feature based on year and yield.
- Observed trend: Corn yields increased over time, but no existing input feature explained this trend.The trend was commonly attributed to technological gains, including improvements in genetics, management, and equipment.
- Trend feature: A new feature, yield_trend, was constructed to explain the observed corn-yield trend.The feature was designed to account only for the observed trend in corn yields.
- Trend feature: For each location, a linear regression used YEAR as the independent variable and yield as the dependent variable.Trends were modeled separately because trends differed across sites.
- Trend feature: The predicted yield value for each data point was added to the dataset as a new input variable representing the increasing annual trend.The trend value was calculated for each location and test year using location-specific regression coefficients.
• Aggregated and cumulative weather feature construction
The study expanded weather inputs with aggregated and cumulative features to provide machine-learning models with more climate information.
- Additional weather features were constructed by cumulating existing weather variables for machine-learning inputs.These features were designed to provide more climate information.
- Aggregated precipitation, growing degree days, and shortwave radiation were computed by summing the corresponding weather features.
- Aggregated minimum and maximum temperature were computed by averaging existing temperature values.
• Feature selection
Because the dataset contained many input variables and was prone to overfitting, the study used a two-stage feature-selection procedure to identify essential features and improve model generalizability.
- Feature selection: A two-stage procedure combined expert-knowledge-based feature selection with permutation-based selection to reduce overfitting in highly dimensional training data.The procedure was designed to select the most essential features and build more generalizable machine-learning models.
i. Feature selection based on expert knowledge
Feature selection proceeded in two stages: expert-guided filtering reduced the weather and planting-progress inputs, followed by permutation importance with a tuned random forest to select the top features. Model hyperparameters were tuned using cross-validation and Bayesian search.
- Permutation importance: Permutation importance evaluates a feature by measuring how shuffling its validation or test-set values changes prediction error.A feature is considered important when permutation increases model error.
- Permutation importance: Permutation importance was calculated with a 100-tree random forest, 10 repetitions, and random 10-fold cross-validation before selecting the top 80 input features.The random forest tree-count hyperparameter was tuned using 10-fold cross-validation.
- Hyperparameter tuning: Bayesian search with 20 iterations tuned machine-learning hyperparameters under a 10-fold cross-validation procedure.Bayesian search was selected because it can tune hyperparameters faster over a continuous range of values than exhaustive grid search.
3. Results
Adding APSIM outputs improved all tested ML models, reducing prediction error and bias, with the strongest average benefits for stacked ensembles. Soil- and weather-related variables were more influential than phenology variables, especially drought stress and water-table depth.
- Overall hybrid-model performance: Adding APSIM features improved all 11 ML models, while weighted ensembles outperformed other ML models on average.The hybrid approach remained better than the benchmark even for LASSO, which had the smallest RRMSE decrease.
- Overall hybrid-model performance: Stacked ensembles benefited most from APSIM inclusion, and all ML models produced less biased predictions with APSIM inputs.Including APSIM variables also improved residual distributions and reduced overall prediction bias in the 2018 test year.
- Extreme-weather-year evaluation: 5% to 43% decrease in prediction RMSE was achieved by hybrid models over benchmarks in the extreme dry year 2012.Predictions were less accurate in 2012 than in 2017 and 2018 for both ML and hybrid models, but hybrid models still improved over benchmarks.
- Partial APSIM-feature inclusion: For crop-related inputs, stacked regression made the best predictions, while stacked random forest produced the least biased predictions.With soil- and weather-related inputs, stacked regression achieved relatively low prediction error and bias.
- Partial APSIM-feature inclusion: Soil- and weather-related APSIM variables and crop-related variables influenced ML predictions more significantly than phenology-related variables.Some soil- and weather-based ensembles even outperformed models using all APSIM variables, indicating that not every APSIM variable was useful.
- APSIM-feature importance: AvgDroughtStress, AvgWTInseason, and CropYield were the most important APSIM features, with water-related variables prominent among influential inputs.Across the top 14 non-trend features, seven were APSIM variables and seven were weather inputs; yield_trend was the most important overall feature.
4. Discussion
Coupling APSIM variables with machine learning improved county-scale corn yield prediction, with soil-water variables contributing most to the gains. However, the approach struggled in the dry 2012 year and would face additional uncertainty when APSIM operates without observed test-year weather.
- Hybrid prediction: 7% to 20% RMSE reduction resulted, on average, from introducing APSIM variables into machine-learning yield prediction models.The study proposed a hybrid simulation–machine learning approach for county-scale crop yield prediction.
- Hybrid prediction: Soil-water-related APSIM variables contributed most to improving yield prediction, and APSIM inclusion improved machine-learning predictions in 2012, 2017, and 2018.Neither the machine-learning nor hybrid model could sufficiently predict yields in the dry 2012 year.
- Feature construction: A newly constructed trend feature was by far the most important input for predicting corn yields.The feature was designed to help machine-learning models capture yearly increasing corn-yield trends.
- Future directions: Future work could add remote-sensing data to the machine-learning prediction task and assess the importance of each data source.This extension follows the demonstrated benefits of adding simulation crop-model inputs.
- Limitations: Additional uncertainty would arise in real-world applications because APSIM inputs used the full weather of each test year, whereas forecasting requires unknown weather.The study identifies forecasting-mode APSIM simulations as an issue for future investigation.
5. Conclusion
Including APSIM simulation inputs improved yield prediction accuracy across all designed machine-learning models. Soil-water variables—especially growing-season average drought stress and average depth to water table—were the most important APSIM inputs.
- Conclusion: APSIM inputs improved yield prediction accuracy across all designed ML models.The improvement occurred when additional inputs from the simulation cropping systems model were included.
- Conclusion: Soil-water-related APSIM variables were the most important inputs for ML yield prediction.The analysis particularly identified growing-season average drought stress and average depth to water table.
- Conclusion: Growing-season average drought stress and average depth to water table were the leading APSIM variables.Both variables describe soil-water conditions during the growing season.
Author Contribution Statement
MS led the research and manuscript drafting, while GH oversaw the study, SA provided data and guidance, and IH prepared the APSIM data.
- MS conducted the research and wrote the manuscript’s first draft.
- GH secured funding, oversaw the research, and reviewed and edited the manuscript.
- SA provided research data and guidance and reviewed and edited the manuscript; IH prepared the APSIM data.