Source-linked AI summary
Winter wheat yield prediction using convolutional neural networks from environmental and phenological data
Amit Kumar Srivastava, Nima Safaei, Saeed Khaki, Gina Lopez, Wenzhi Zeng, Frank Ewert, Thomas Gaiser, Jaber Rahimi
TL;DR
Winter wheat yield prediction must account for interacting environmental and crop factors across space and time. The study evaluates machine-learning baselines and a one-dimensional CNN using German county data, finding that the CNN outperformed the tested baselines while SHAP and force plots identified influential variables and weeks.
Problem
Crop yield forecasting requires methods that account for interacting genotype, weather, soil, and management factors for timely regional prediction.
Method
The study trains a 1-dimensional CNN and eight supervised machine-learning baselines on weekly weather, soil, phenology, and yield data from 271 German counties between 1999 and 2019.
Results
The CNN outperformed the tested baselines, while nonlinear CNN, DNN, and XGBoost models outperformed linear models across the reported validation metrics.
Takeaways & Limitations
SHAP and force plots identified influential variables and time periods, including DUL, wind speed, and radiation, for interpreting winter wheat yield predictions.
Abstract
from arXiv · showhide
Crop yield forecasting depends on many interactive factors, including crop genotype, weather, soil, and management practices. This study analyzes the performance of machine learning and deep learning methods for winter wheat yield prediction using an extensive dataset of weather, soil, and crop phenology variables in 271 counties across Germany from 1999 to 2019. We proposed a Convolutional Neural Network (CNN) model, which uses a 1-dimensional convolution operation to capture the time dependencies of environmental variables. We used eight supervised machine learning models as baselines and evaluated their predictive performance using RMSE, MAE, and correlation coefficient metrics to benchmark the yield prediction results. Our findings suggested that nonlinear models such as the proposed CNN, Deep Neural Network (DNN), and XGBoost were more effective in understanding the relationship between the crop yield and input data compared to the linear models. Our proposed CNN model outperformed all other baseline models used for winter wheat yield prediction (7 to 14% lower RMSE, 3 to 15% lower MAE, and 4 to 50% higher correlation coefficient than the best performing baseline across test data). We aggregated soil moisture and meteorological features at the weekly resolution to address the seasonality of the data. We also moved beyond prediction and interpreted the outputs of our proposed CNN model using SHAP and force plots which provided key insights in explaining the yield prediction results (importance of variables by time). We found DUL, wind speed at week ten, and radiation amount at week seven as the most critical features in winter wheat yield prediction.
Data acquisition
The study combines county-level winter wheat yield, weather, phenology, and soil data across Germany, preprocessing environmental measurements into weekly features and standardized inputs.
- Data acquisition: Daily meteorological measurements were interpolated to a 1 km grid and aggregated into weekly, area-weighted NUTS3 features.Variables included minimum and maximum temperature, radiation, precipitation, relative humidity, and wind speed.
- Data acquisition: Soil inputs included volumetric crop-available water at wilting point, field capacity, and saturation, plus bulk density to 1.3 m depth.The field-capacity variable is DUL.
- Data acquisition: Sowing, flowering, and harvest dates were obtained from the DWD phenology database, with representative weather and soil values calculated from sampled county grids.Different county sowing and harvest dates produce different growing-season lengths.
- Data acquisition: All features were standardized with z-score normalization to reduce scale dominance and improve numerical stability and training speed.The transformation uses each feature's mean and standard deviation.
Methods
The proposed model combines one-dimensional convolutions for temporal weather patterns with fully connected networks for soil and phenology inputs before predicting yield.
- Methods: The CNN processes weather variables through one-dimensional convolutions to capture temporal and nonlinear effects across the growing season.All six weather variables are processed in the CNN component.
- Methods: The architecture uses convolutional, average-pooling, and dense layers with valid padding throughout.Figure 3 presents convolution parameters as convolution type, filter count, kernel size, and stride size.
- Methods: All layers except the final layers use the ReLU activation function, while concatenation and transpose operations organize intermediate representations.The architecture caption identifies Conv, AvgPool, and Dense as standard convolutional, average pooling, and fully connected layers.
- Methods: Soil and phenology data enter a separate two-layer fully connected network whose output is combined with CNN-derived weather features.The combined representation is passed to additional fully connected layers before final yield prediction.
Random Forest (RF).
Random Forest ensembles decision trees by training trees on bootstrapped samples and combining them into a single model.
- Random Forest (RF).: Random Forest uses multiple decision trees and bagging to form an ensemble model.Bootstrapping randomly selects K samples from the original dataset to train the component trees.
- Random Forest (RF).: K-Nearest Neighbor predicts an output by identifying training instances closest to the test instance using Euclidean distance.The method is non-parametric and supports classification and regression problems.
- Random Forest (RF).: The Euclidean-distance equation computes distances between data points across their dimensions using p, q, n, and the index i.Here, d denotes distance, p and q are data points, n is the number of data points, and i is an index.
LASSO and Ridge Regression.
Lasso and Ridge reduce variance in linear regression by shrinking coefficients, while the other described baselines use tree ensembles, support-vector regression, or boosting.
- LASSO and Ridge Regression.: Lasso and Ridge Regression reduce linear-model variance by shrinking regression coefficients.Lasso uses L1 regularization, whereas Ridge uses L2 regularization.
- LASSO and Ridge Regression.: Lasso can set some coefficients to zero, functioning as a feature-selection method, while Ridge shrinks coefficients without eliminating them.This difference distinguishes the sparsity behavior of the two regularizers.
- LASSO and Ridge Regression.: Regression trees recursively partition data through binary splits selected to minimize within-partition squared deviations from the mean.Splitting stops when nodes reach a minimum size or their squared deviation becomes zero.
- LASSO and Ridge Regression.: Support Vector Regression fits a continuous-response function within an epsilon-insensitive tube and identifies support vectors from distances to that tube.The optimization uses a user-defined trade-off parameter C and margin of error ǫi.
- LASSO and Ridge Regression.: XGBoost sequentially builds decision trees so each tree improves the preceding ensemble, while parallelized training increases speed.It is described as an improved gradient-boosted decision-tree method.
Deep Neural Networks (DNN) .
Deep neural networks learn nonlinear representations from data and can recognize learned patterns to predict outputs. The paper also situates traditional single-hidden-layer neural networks as pattern-learning models.
- DNNs learn underlying data representations without handcrafted features through multiple abstraction levels.The described DNN is a feedforward network with multiple hidden layers.
- Nonlinear functions applied across hidden layers make DNN models highly nonlinear.
- DNNs are trained with gradient-based optimization to minimize the task-specific loss function.
- Traditional single-hidden-layer neural networks learn data patterns from training data through weighted connections and use them to predict outputs.
- The study used Python, scikit-learn, TensorFlow, Matplotlib, Seaborn, and QGIS for modeling, visualization, and spatial plotting.
Model interpretability.
The study frames interpretability as a way to explain black-box yield predictions while evaluating models on environmental, phenological, and soil data. It uses standard error and correlation metrics to assess predictive performance.
- Interpretation tools are applied to tree ensembles and deep-learning models to explain yield prediction results.
- The models were trained and tested on weekly meteorological, crop-phenological, and soil variables from 271 German counties between 1999 and 2019.The dataset contained 5,692 instances and 277 column features.
- Hyperparameter values were tested on a validation set to select settings producing the lowest test-data error.Validation used randomly selected cross-validation data.
- Performance was evaluated using mean absolute error, root mean square error, and correlation coefficient metrics.
XGBoost
The section identifies Table 1 as reporting training and validation results for baseline machine-learning models and the proposed CNN, using RMSE, MAE, and correlation coefficient r.
- Table 1 reports RMSE, MAE, and correlation coefficient r for baseline models and the proposed CNN predicting winter wheat yield.
- The correlation coefficient uses predicted, average, and ground-truth target values across n data points.
Results
The CNN achieved the strongest winter wheat yield prediction performance among the tested models, while residual and spatial analyses characterized prediction reliability. SHAP and force plots further identified how environmental and phenological features contributed to predictions.
- Model performance: CNN outperformed the tested methods, attributed to its nonlinear structure and ability to capture temporal dependencies in weather data.XGBoost and DNN showed comparable performance, while linear models performed more weakly than nonlinear approaches.
- Model performance: CNN showed the strongest predicted-versus-ground-truth association, with r = 0.78 and R2 = 0.61, compared with XGBoost at r = 0.71 and DNN at r = 0.70.The CNN predictions explained more than 60 percent of the variability in ground-truth values.
- Residual analysis: Prediction residuals for XGBoost, DNN, and CNN were approximately normally distributed, although the normality-test p-value was borderline at 0.05.The authors emphasize predictive modeling rather than inferential assumptions about the data mechanism.
- Spatial error analysis: CNN prediction errors varied from 0.1 to 77.6% across counties in 2019, with higher errors primarily in eastern and northeastern Germany.Central counties represented lower error values, although the CNN error range exceeded those of DNN and XGBoost.
- Feature interpretation: Feature-interpretation analyses used SHAP values and force plots to explain global feature importance and individual predictions.Force plots represented positive and negative feature contributions toward predicted yield values, while the selected figures covered top features and minimum, maximum, and median yield instances.
- Feature interpretation: Wind speed, minimum temperature, and radiation had the highest normalized weather-feature importance, with values of 13.20, 12.03, and 9.56, respectively.The top-feature analysis also identified minimum and maximum temperature among the leading contributors to yield predictions.
Conclusions
The proposed CNN outperformed the baseline models for winter wheat yield prediction, while SHAP and force plots supported interpretation of feature effects and prediction instances.
- The proposed CNN outperformed all baselines across the test years 2017, 2018, and 2019.
- Nonlinear models, including DNN, CNN, and XGBoost, outperformed linear models to varying extents.The paper attributes this pattern to the complex relationship between yield and interacting genotype and environmental factors.
- SHAP and force plots characterized important features and their values for instances with highest, lowest, and median predicted yields.
Funding
The article reports open-access funding enabled and organized by Projekt DEAL.
- Open-access funding enabled publication of the article.
- Projekt DEAL organized the open-access funding.
- The funding information identifies Projekt DEAL as the organizing body.
Additional information
The article provides supplementary-material access, correspondence information, reprint and permissions guidance, publisher neutrality language, and Creative Commons licensing terms.
- Supplementary material is available online through the article’s DOI.
- Requests for correspondence and materials should be addressed to A.K.S., N.S. or S.K.
- Reprints and permissions information is available through Nature’s reprints website.
- Springer Nature remains neutral regarding jurisdictional claims in published maps and institutional affiliations.
- The article permits reuse, sharing, adaptation, and reproduction under a Creative Commons Attribution 4.0 International License with appropriate credit and change disclosure.