Source-linked AI summary
Crop Yield Prediction Using Deep Neural Networks
Saeed Khaki, Lizhi Wang
TL;DR
Predicting crop yield is difficult because genotype, environmental conditions, and their interactions are complex. The paper develops a deep neural network using genotype and environment data, outperforming benchmark models with validation RMSE of approximately 11% of average yield for yield and check yield.
Problem
Crop yield prediction is important but difficult because high-dimensional genetic markers interact with environmental conditions and field-management practices.
Method
The study trains deep neural networks on genotype, yield, weather, and soil data to predict yield outcomes, with feature selection based on the trained model.
Results
DNN outperformed the three benchmark models on almost all measures, with validation RMSE approximately 11% of average yield for yield and check yield.
Takeaways & Limitations
Feature selection reduced input dimensionality without significant accuracy loss and indicated that environmental factors affected crop yield more than genotype.
Takeaways & Limitations
The model’s black-box structure makes it difficult to produce testable hypotheses that provide biological insight.
Abstract
from arXiv · showhide
Crop yield is a highly complex trait determined by multiple factors such as genotype, environment, and their interactions. Accurate yield prediction requires fundamental understanding of the functional relationship between yield and these interactive factors, and to reveal such relationship requires both comprehensive datasets and powerful algorithms. In the 2018 Syngenta Crop Challenge, Syngenta released several large datasets that recorded the genotype and yield performances of 2,267 maize hybrids planted in 2,247 locations between 2008 and 2016 and asked participants to predict the yield performance in 2017. As one of the winning teams, we designed a deep neural network (DNN) approach that took advantage of state-of-the-art modeling and solution techniques. Our model was found to have a superior prediction accuracy, with a root-mean-square-error (RMSE) being 12% of the average yield and 50% of the standard deviation for the validation dataset using predicted weather data. With perfect weather data, the RMSE would be reduced to 11% of the average yield and 46% of the standard deviation. We also performed feature selection based on the trained DNN model, which successfully decreased the dimension of the input space without significant drop in the prediction accuracy. Our computational results suggested that this model significantly outperformed other popular methods such as Lasso, shallow neural networks (SNN), and regression tree (RT). The results also revealed that environmental factors had a greater effect on the crop yield than genotype.
1 INTRODUCTION
Accurate crop yield prediction supports food security, breeding, and farm decisions but requires modeling complex genotype–environment interactions. This paper applies deep neural networks to predict corn-hybrid yield-related outcomes from genotype and environment data.
- Accurate crop yield prediction informs food-security policy, hybrid breeding, and growers’ management and financial decisions.
- Yield prediction must account for genetic-marker effects interacting with multiple environmental conditions and field-management practices.
- Prior approaches modeled phenotype using genotype, environment, and G×E effects, often treating interactions as noise or aggregating environments into mega-environments.
- The paper uses deep neural networks to predict corn-hybrid yield, check yield, and yield difference from genotype and environment data.Deep neural networks learn data representations through stacked nonlinear layers rather than handcrafted features.
- Deep neural networks can reveal more fundamental nonlinear relationships than shallow single-hidden-layer networks, but require advanced hardware and optimization techniques.Their numerous nonlinear activations produce highly non-convex loss functions without guaranteed gradient-based convergence.
2 DATA
The study used a large, multi-environment dataset of maize hybrids combining genotype, yield, weather, and soil information. Training and validation data were constructed from observations collected between 2001 and 2016.
- Dataset scope: 2,267 experimental hybrids were planted in 2,247 locations across the United States and Canada between 2008 and 2016.The dataset supported prediction of corn-hybrid performance in different locations.
- Dataset composition: 19,465 genetic markers characterized each hybrid, while yield data covered 148,452 hybrid-year-location samples.The yield dataset included observed yield, check yield, and yield difference.
- Environmental variables: Environmental inputs included day length, precipitation, solar radiation, vapor pressure, maximum temperature, and minimum temperature.
- Data split: 142,952 samples formed the training dataset and 5,510 samples formed the validation dataset using data from 2001 to 2016.Validation samples were unique hybrid-location combinations, and the 2017 ground-truth responses were unavailable.
3 METHODOLOGY
The methodology preprocesses incomplete genotype data, predicts weather from historical lagged observations, and uses paired deep neural networks to predict yield and check yield. The yield models combine genotype, weather, and soil inputs with deep architectures designed to balance accuracy and overfitting.
- Genotype preprocessing: 37% of genotype data had missing values; markers below a 97% call rate or with minor-allele frequency below 1% were discarded.Genotypes were coded as {−1, 0, 1} for aa, aA, and AA alleles.
- Weather prediction: Weather variables were predicted for 2016 using 2001–2015 historical weather data.Weather prediction was included because weather affects yield but is unknown beforehand.
- Weather prediction: 24,717 samples trained each weather-variable network using four previous years at the same location.Four years produced the best results among the tested periodic lags.
- Yield prediction: Two deep neural networks predicted yield and check yield separately, with their output difference used to predict yield difference.This structure was more effective than one network for yield difference because genotype and environment relate more directly to yield and check yield.
- Yield prediction: Each yield network used 21 hidden layers with 50 neurons per layer, selected as the best balance between prediction accuracy and limited overfitting.Weights used Xavier initialization, and training used mini-batch SGD with Adam at a 0.03% learning rate.
- Yield prediction: The deep network accepted genotype, weather, and soil data, with residual shortcuts, maxout activations, and L2 regularization across hidden layers.L1 regularization was added to the first layer to decrease dimensionality.
4 RESULTS
The results show that DNN outperformed the benchmark models across nearly all measures, while maintaining low regional errors and approximately preserving yield distributions. Prediction accuracy deteriorated with predicted rather than ground-truth weather, demonstrating sensitivity to weather quality.
- Model comparison: DNN outperformed Lasso, SNN, and RT on almost all performance measures.Lasso’s weaker performance reflected its linear structure, which could not capture epistatic or G×E interactions and nonlinear environmental effects.
- Regional errors: 207 of 244 validation locations had prediction-error RMSE below 15.Prediction error was consistently low for most locations in the validation dataset.
- Distributional properties: DNN approximately preserved the ground-truth yield distribution, but its predicted-yield variance was lower.The lower variance indicates that predictions were more centralized around the mean.
- Weather sensitivity: DNN prediction accuracy deteriorated when predicted weather replaced ground-truth weather.The comparison indicated that yield prediction was sensitive to weather prediction quality and that perfect weather predictions could improve results.
5 ANALYSIS
The analysis found that environmental inputs, especially soil and weather, explained more yield variation than genotype. Guided backpropagation identified influential variables and enabled feature reduction without a significant accuracy loss.
- Component importance: DNN(W) and DNN(S) performed approximately equally and significantly better than DNN(G), indicating environmental components explained more yield variation than genotype.The comparison used models trained separately on genotype, soil, or weather data, alongside an average-phenotype baseline.
- Feature importance: Guided backpropagation ranked input variables by backpropagated gradient magnitude, with larger gradients indicating greater importance.The method averaged last-hidden-layer activations across validation samples before propagating gradients to the input space.
- Feature importance: Percentage of clay and soil pH were more important than the other soil conditions among the 8 evaluated soil variables.The soil variables included available water capacity, organic matter, cation exchange capacity, and saturated hydraulic conductivity.
- Feature importance: Solar radiation and temperature had considerable effects on yield variation across environments, with higher yield associated with low temperature and high solar radiation.Lower temperature was linked to longer growth duration and greater radiation interception; precipitation was also identified as important.
- Feature selection: Selecting 50 genetic markers and 20 environmental components did not significantly reduce DNN prediction accuracy.The selected features were ranked by their estimated effects and used to evaluate the feature-selection method against corresponding full-feature results.
6 CONCLUSION
The study presented deep neural networks for crop-yield prediction from genotype and environment data, learning nonlinear gene–environment relationships and achieving superior performance in the 2018 Syngenta Crop Challenge. Although the model is a black box, backpropagation-based feature selection identified important features to improve interpretability.
- 6 CONCLUSION: The approach demonstrated superior performance in the 2018 Syngenta Crop Challenge using large datasets of corn hybrids.
- 6 CONCLUSION: Deep neural networks predicted yield, check yield, and yield difference from genotype and environment data.The networks learned nonlinear and complex relationships among genes, environmental conditions, and their interactions from historical data.
- 6 CONCLUSION: The model’s black-box structure made it difficult to produce testable hypotheses and biological insights despite capturing genotype-by-environment interactions.
- 6 CONCLUSION: Backpropagation-based feature selection successfully found important features and was used to make the model less of a black box.
DATA AVAILABILITY STATEMENT
The study analyzed data provided by Syngenta through the 2018 Syngenta Crop Challenge, which was publicly available during the challenge period. Researchers can request access by contacting Syngenta directly.
- DATA AVAILABILITY STATEMENT: Syngenta provided the study data through the 2018 Syngenta Crop Challenge, with public access from September 2017 to January 2018.Researchers seeking access may contact Syngenta directly.