Source-linked AI summary
Crop Yield Prediction Integrating Genotype and Weather Variables Using Deep Learning
Johnathon Shook, Tryambak Gangopadhyay, Linjiang Wu, Baskar Ganapathysubramanian, Soumik Sarkar, Asheesh K. Singh
TL;DR
Predicting crop yield across unseen environments requires models that integrate genotype information with temporally varying weather. The paper uses historical soybean trial records, pedigree-relatedness representations, LSTM recurrent networks, and temporal attention to predict yield and identify influential growing-season windows. The combined approach outperformed RF, LASSO, and the USDA model, while generating hypotheses about weather effects and genotype-by-environment interactions.
Problem
Predicting crop performance in unseen environments while accounting for climatic variability is difficult but important for breeding and production decisions.
Method
The framework combines 13 years of soybean trial data with pedigree-relatedness representations and 30 weeks of weather inputs in LSTM and temporal-attention models.
Results
The LSTM and temporal-attention models outperformed RF and LASSO regression, with RF test RMSE of 9.889 and LASSO test RMSE of 12.779.
Takeaways & Limitations
Attention-based predictions identify influential weather periods and support hypothesis generation about genotype-by-environment responses and breeding under climate variability.
Takeaways & Limitations
The study lacks direct genetic markers and other potentially informative factors such as irrigation, soil fertility, disease pressure, and management variables.
Abstract
from arXiv · showhide
Accurate prediction of crop yield supported by scientific and domain-relevant insights, can help improve agricultural breeding, provide monitoring across diverse climatic conditions and thereby protect against climatic challenges to crop production including erratic rainfall and temperature variations. We used historical performance records from Uniform Soybean Tests (UST) in North America spanning 13 years of data to build a Long Short Term Memory - Recurrent Neural Network based model to dissect and predict genotype response in multiple-environments by leveraging pedigree relatedness measures along with weekly weather parameters. Additionally, for providing explainability of the important time-windows in the growing season, we developed a model based on temporal attention mechanism. The combination of these two models outperformed random forest (RF), LASSO regression and the data-driven USDA model for yield prediction. We deployed this deep learning framework as a 'hypotheses generation tool' to unravel GxExM relationships. Attention-based time series models provide a significant advancement in interpretability of yield prediction models. The insights provided by explainable models are applicable in understanding how plant breeding programs can adapt their approaches for global climate change, for example identification of superior varieties for commercial release, intelligent sampling of testing environments in variety development, and integrating weather parameters for a targeted breeding approach. Using DL models as hypothesis generation tools will enable development of varieties with plasticity response in variable climatic conditions. We envision broad applicability of this approach (via conducting sensitivity analysis and "what-if" scenarios) for soybean and other crop species under different climatic conditions.
Introduction
The paper addresses prediction of soybean performance in unseen environments by combining genetic and weather information, especially temporal weather variability. It proposes an explainable LSTM-attention framework to improve prediction and support breeding decisions.
- Predicting seed yield in unseen and new environments remains a key challenge because extensive genotype-environment datasets are costly to generate.
- Climatic resiliency requires understanding how temperature, rainfall, and other weather variables vary across both locations and time.
- Existing crop and yield models can be limited by parameter estimation, complex unforeseen conditions, or experiments that vary only one environmental factor.
- The proposed framework uses LSTM and temporal attention with 30 weeks of weather data and reduced pedigree representation to predict soybean yield.
- Temporal attention highlights important growing-season periods while preserving prediction accuracy, enabling interpretation of genotype responses to weather.
Methods
The study assembled a 13-year, multi-environment soybean dataset with pedigree-relatedness and weather features, then modeled yearly seed yield using LSTM architectures with and without temporal attention. Genotype clustering, sequence encoding, attention mechanisms, and held-out RMSE evaluation formed the core workflow.
- Data collection and preprocessing: The final dataset contained 103,365 performance records spanning 13 years and 5,839 unique genotypes, with available management information.
- Data collection and preprocessing: Daily weather observations from nearby stations were summarized as maximum, minimum, and average conditions across growing-season time frames and appended to performance records.The growing season was defined as April 1 through October 31.
- Genotype representation: Pedigree parentage information produced a 5,839 x 5,839 genotype correlation matrix, which was used to represent relatedness where molecular marker data were unavailable.
- Genotype representation: K-means clustering grouped genotypes into five clusters using the correlation matrix to extract relatedness information.The clustering objective minimized within-cluster sum-of-squares, or inertia.
- Model development: Two LSTM-based many-to-one models predicted yearly seed yield: a stacked LSTM and a temporal attention model, each with variants that included or excluded maturity-group and genotype-cluster information.LSTM was selected to learn long-range correlations and address training difficulties associated with long-term dependencies in recurrent networks.
- Model development: The temporal attention model aggregated annotations from all time steps, while the stacked LSTM used the final encoder hidden state as the sequence representation.This design allowed the attention model to use information across the sequence rather than only the final encoded state.
- Training and evaluation: Inputs were scaled to (-1, 1), data were split randomly into 80% training, 10% validation, and 10% test sets, and models were evaluated on test-set RMSE.Training used 200 epochs, Adam with a 0.001 learning rate, mean squared error loss, Keras, TensorFlow, and NVIDIA GPUs.
Results
The models achieved their strongest yield predictions when weather, maturity-group, and genotype-cluster information were combined, while temporal attention highlighted later-season importance for high-yielding genotypes.
- Weekly weather data produced similar performance to monthly and bi-weekly sampling, whereas daily data differed; weekly data was selected as an intermediate scenario.The approximate test RMSE for the similar-performing scenarios was 7.206.
- Adding maturity group caused the largest test-RMSE reduction, and incorporating its classification after the second LSTM layer further improved performance.
- Average relative humidity had the lowest test RMSE in greedy search, followed by average direct normal irradiance and other weather variables.The remaining ranking included maximum direct normal irradiance, maximum and minimum surface temperature, average surface temperature, and average precipitation.
- 0.894 correlation between predicted and observed yields was attained in testing and validation, largely capturing performance differences between environments and years.
- 0.802 coefficient of determination and test RMSE = 7.130 were achieved when maturity group, genotype cluster, and weather variables were all incorporated.The test RMSE was about 14% of the test-set average yield and 44.5% of its standard deviation.
- Test RMSE was 12.779 for LASSO and 9.889 for Random Forest, while both proposed LSTM models outperformed them with the same input features.
- Deep learning models had lower absolute errors than the USDA model for all 12 years except 2011, including 0.03 versus 1.32 in 2014 and 0.35 versus 1.70 in 2015.The comparison used year-wise averages across states because the USDA model does not predict individual locations.
- Attention weights indicated that early-season variables were less important for the highest-yielding genotypes in MG1 and MG7, with increasing importance during August–September.These time phases coincide with crop reproductive phases, but functional validation was outside the study’s scope.
Discussion
The study presents LSTM and attention models as useful for yield prediction, interpreting weather effects, and generating breeding hypotheses, while noting important data and causal limitations.
- LSTM-based prediction can capture temporal weather effects and the relative importance of weather variables for seed yield.
- The framework supports hypothesis generation for breeding climate-resilient varieties and studying genotype-by-environment responses across crops and climatic settings.
- Minimum surface temperature emerged as the most significant weather variable, suggesting nighttime temperatures may matter more than previously suggested.The retrospective design cannot establish this conclusion definitively and calls for further empirical investigation.
- Predictive performance could be strengthened by adding irrigation, soil fertility, disease pressure, genetic markers, genomic data, and sensor-derived information.
- The attention module helps identify potentially important features, reducing the black-box character of deep-learning yield models.
- The approach could provide finer-resolution predictions than the USDA model, including regional and intra-state variation relevant to marketing decisions.
Conclusion
The paper frames deep-learning prediction as a hypothesis-generation strategy for studying climate change and varietal plasticity, while emphasizing that causality remains unresolved.
- Causal understanding of climate change effects on varietal plasticity remains an important next step beyond predictive modeling.
- The proposed strategy combines large, heterogeneous datasets with domain knowledge to generate hypotheses about genotype, environment, and morphology relationships.
Clustering
Clustering groups unlabeled examples using similarity measures, while algorithm choice depends on dataset scale and computational cost.
- Clustering assigns examples to groups based on a similarity measure computed from their feature data.
- Pairwise similarity methods scale as n^2, whereas k-means scales linearly with n and is therefore suited to larger datasets.
Modeling Approach
The framework uses stacked LSTMs to encode weather and genotype-related sequences, while temporal attention aggregates time-step information into an interpretable context vector for yield prediction.
- LSTM architecture: LSTM gates regulate which information is removed, added, or exposed while preserving relevant long-range sequence information.Input, output, and forget gates control updates to the cell and hidden states.
- Stacked LSTM encoding: Two stacked LSTM layers sequentially encode the input time-steps, with the first layer’s sequence of encodings passed to the second layer.The model updates hidden and cell states at each time-step.
- Temporal attention: Temporal attention computes a weighted sum of hidden-state annotations to form a context vector for prediction.The attention weights represent each time-step’s contribution to the prediction.
- Temporal attention: A jointly trained dense alignment model scores how well each time-step aligns with the prediction and produces the attention weighting.The alignment model is parameterized as a feedforward neural network.
Experiments
Experiments varied the temporal resolution of weather inputs while keeping the model architecture fixed, balancing prediction accuracy against training efficiency.
- Input sequence experiments: The dataset contains 214 daily time-steps with seven weather variables, which were downsampled to weekly, biweekly, and monthly sequences.The resulting sequence lengths included Tx = 30, Tx = 15, and Tx = 7.
- Input sequence experiments: RMSE values were nearly identical across input sequence lengths except for Tx = 7.The study selected Tx = 30 because it enabled faster training while retaining long-range temporal information.
Adding Maturity Group, Genotype Cluster Informations
The experiments evaluated how maturity group, genotype cluster, and weather variables contribute to prediction, including separate analyses for northern and southern locations.
- Maturity group and genotype cluster: The second maturity-group architecture, concatenating MG at every time-step and before prediction, produced the lowest RMSE among the tested approaches.The three architectures differed in where MG information was injected.
- Maturity group and genotype cluster: The same second-approach design gave the lowest RMSE when genotype cluster information was added.This approach concatenated cluster information at every time-step and after the second LSTM layer.
- Weather-variable search: The weather-variable greedy search ranked variables by iteratively selecting the addition that produced the lowest RMSE.MinSur was selected first and ADNI second; RMSE remained between 7.19 and 7.20, with little impact from later additions.
- Regional analyses: Separate greedy searches assessed weather-variable importance for northern locations defined by MG 0 to 4 and southern locations defined by MG 4 to 8.The results were reported independently in Tables 6 and 7.
Performance Comparison of two models with RF, LASSO
The study compared Stacked LSTM and Temporal Attention models with RF and LASSO across input sequence lengths and input-information variants, then examined which information was needed for best performance.
- Model comparison: Both LSTM-based models were compared with RF and LASSO using the same input features while varying sequence length and included information.Each model used variants with weather variables alone or with additional maturity-group and genotype-cluster information.
- Model comparison: The two proposed LSTM models showed comparable accuracy, so subsequent input-information experiments used the Stacked LSTM model.Performance was evaluated using test RMSE and R2.
- Input information: The best Stacked LSTM performance required maturity-group, genotype-cluster, and weather-variable information together.This conclusion was based on the comparative experiments reported for different inputs.
Comparison with USDA Model
The study compares the Stacked LSTM model with the USDA’s state-wise, weather-based soybean yield model using year-wise averages across states. It also examines how training-data availability relates to test-set performance across metagroup and genotype-cluster combinations.
- Model comparison: The Stacked LSTM model is compared with the USDA’s weather-based soybean yield model using year-wise average test-set performance across states.The USDA model uses linear regression and predicts yield state-wise rather than for individual locations; absolute error is computed between predicted and actual yield.
- Data availability and performance: Figure 9 examines test-set performance in relation to training-data availability for metagroup and genotype-cluster combinations.The figure is designed to gain insights into performance based on data availability in the training set.
- Data availability and performance: Test RMSE values are plotted as a heat map across all metagroup and genotype-cluster combinations, alongside a training-sample-to-unique-location ratio.The ratio estimates data availability and distribution for each combination, including the reported highest-RMSE case at metagroup 7 and cluster 1.