PaperlayerView exact revision on arXiv

Source-linked AI summary

Evaluating and improving crop-yield forecasting methods during extreme drought

Shrey Gupta, Yi Ming, George Mohler

arXiv:2608.17971v1cs.LG

TL;DR

The paper addresses crop-yield forecasting during extreme drought when train and test meteorological distributions differ and spatial and temporal data irregularities constrain learning. It compares ML and deep learning models using meteorological predictors and applies sample weighting and feature selection as model modifications. VITA consistently outperforms the ML models with or without modifications, while the adaptations improve ML models but produce little to no improvement for VITA.

  • Problem

    Extreme drought forecasting is limited by feature distribution mismatch and spatial and temporal data irregularities.

  • Method

    The study compares ML and deep learning forecasting models using meteorological data and applies sample weighting and feature selection modifications.

  • Results

    VITA consistently outperforms non-deep-learning ML models with or without model adaptations.

  • Takeaways & Limitations

    SHAP-based model adaptations are effective for non-deep-learning models, while VITA shows little to no improvement from these modifications.

  • Takeaways & Limitations

    Scaling-based solutions improve forecasting performance but remain unable to explain the underlying agroclimatic processes.

Abstract

from arXiv · show

The impact of climate variability on food production has led to the creation of various forecasting models that uses machine learning (ML), numerical weather predictors (NWP) or a hybrid of ML-NWP models to identify structural and physical relationships between meteorological drivers and crop growth, in order to predict crop yield. Droughts, for example the 2012 Midwestern US (Corn Belt) drought, are extreme events that affect crop production and test the limits of these forecasting models. Using 16 meteorological drivers as predictors, we compare ML (non-deep learning) and deep learning forecasting models to predict the county-level corn yield for the extreme drought year, 2012. This forecasting problem is characterized by a dissimilarity between the feature distributions of the training and test data, where the meteorological conditions of the extreme drought year fall outside the range of historically observed values. Additionally, the dataset consists of spatial and temporal irregularities where counties with missing yields introduce spatial sparsity and the use of only a subset of daily values per year introduce temporal sparsity. To overcome this, we use sample weighting and feature selection as modifications to improve our forecasting models. These modifications lead to an improvement for ML models; however, the deep learning model VITA shows little to no improvement. While VITA outperforms the ML models with or without modifications, our current study sheds light on the effect of dissimilarity between train and test feature distributions on forecasting models, compares deep learning versus non-deep learning models, and introduces modifications that are effective for non-deep learning models.

1 Introduction

The study examines corn-yield forecasting during extreme drought as a feature distribution mismatch and data irregularity problem. It compares ML and DL models and evaluates sample weighting and SHAP-derived feature selection as adaptations.

  • Study scope: The authors compare non-neural ML and DL models under feature distribution mismatch and data irregularity constraints.The study asks whether model complexity affects generalization and whether meteorological drivers capture yield variation.
  • Motivation: Extreme droughts challenge forecasting models because representative samples are scarce and test meteorological conditions can fall outside historical training ranges.The study terms this train–test dissimilarity feature distribution mismatch.
  • Study scope: The study evaluates annual county-level corn-yield forecasts using meteorological drivers across the Midwestern United States.The data combine annual county yields with daily meteorological features.
  • Motivation: Data irregularities arise from missing county-level yields and daily meteorological measurements that cover only crop phases rather than the complete year.These create spatial and temporal coverage constraints for county-level forecasting.
  • Approach: Sample weighting and SHAP-derived feature selection are introduced as model modifications to improve forecasting performance.The modifications target adaptation to the study’s data conditions.

2 Background and Related Works

Crop-yield forecasting uses meteorological, land, soil, vegetation, and management information, while this study focuses on meteorological drivers under irregular and extreme-event conditions. Prior work motivates comparing models that forecast yields or drought-related indicators and examining the constraints of the available datasets.

  • Forecasting context: Meteorological drivers alone can explain greater than one-third of yield variability, although soil, physiology, and management also affect yield.Meteorological effects on yield are indirect relative to these dependent factors.
  • Datasets: The gridMET dataset covers 1980–2018 across 1,413 US Corn Belt counties and contains missing-yield and incomplete-season measurement irregularities.Daily measurements cover sowing, growing, and harvesting phases rather than the complete year.
  • Datasets: The study uses NASA Power data for VITA pretraining and gridMET data for fine-tuning and forecasting.NASA Power provides 39 years of data at 0.5° spatial resolution, while gridMET supplies the forecasting dataset.
  • Related work: The open question is which models trained on meteorological features remain useful forecasters during extreme droughts.Prior models include Tradaboost.R2, Extreme Lasso, and VITA, which show promise in drought scenarios.
  • Related work: Prior forecasting studies often combine meteorological data with soil, vegetation, land-use, or management drivers, whereas this setting is intentionally limited to meteorological drivers.The cited studies include heterogeneous datasets and derived drought indicators such as ESI and NDVI.
  • Related work: Related work also forecasts drought indicators rather than yield, including ESI with LSTM models and 2012 drought patterns with SubX models.SubX showed low performance for spring and early-summer precipitation deficits and moderate performance for temperature anomalies.

3 Methodology

The methodology introduces comparative forecasting models together with feature engineering and model modifications. These components are used in combination with existing models for the study’s evaluation.

  • Method design: The study introduces forecasting models for a comparative analysis.The models are presented alongside the feature engineering used in the experiments.
  • Method design: Feature engineering and model modifications are applied in combination with the existing forecasting models.The modifications are part of the methodology used to evaluate model performance.

3.1 Prediction Models

The paper compares several machine-learning regressors with VITA, a variational transformer for asymmetric data, to forecast crop yield. The models span linear, tree-based, kernel, extreme-value, transfer-learning, and deep-learning approaches.

  • Machine-learning models: Linear models learn relationships between standardized meteorological features and yield, while tree-based models recursively partition feature space.The linear-model family includes Ridge and Lasso regression; Random Forest is described as an ensemble of trees.
  • Specialized regressors: SVR uses a kernel-based formulation, whereas Extreme Lasso and Huber regression target extreme-value behavior.SVR operates in a reproducing kernel Hilbert space; Extreme Lasso weights extreme residuals, and Huber regression uses a thresholded loss.
  • Transfer learning: TrAdaBoost.R2 transfers information by combining source and target domains and reweighting source samples dissimilar to target samples.Its updates down-weight source samples with larger residuals while up-weighting target samples over boosting iterations.
  • Model roster: The study evaluates machine-learning models alongside VITA, a decoder-free variational transformer model.VITA uses variational inference and attention to capture meteorological sequences and latent weather representations.
  • Deep learning model: VITA embeds weekly meteorological sequences, processes them with a four-layer Transformer encoder, and aggregates latent states with learned temporal attention.The resulting representation is concatenated with historical yields and passed through a yield-prediction MLP.
  • Deep learning model: VITA represents latent weather states probabilistically and is pretrained by reconstructing randomly masked meteorological drivers.Fine-tuning disables the reconstruction term while retaining KL regularization and uses mean-squared error for yield prediction.

3.2 Data and model adaptation

The paper adapts its forecasting pipeline to feature-distribution mismatch and extreme-event conditions through preprocessing, SHAP-based feature selection, sample weighting, and model-specific attention changes. VPD and ETR are identified as the most influential drivers in the 2007–2011 forecasting window.

  • Preprocessing: The preprocessing pipeline aggregates gridMET observations into rolling three-month seasonal windows and derives distributional and historical-yield features.Window statistics include means, dispersion, extrema, quartiles, skewness, and kurtosis; historical yields provide five-year summaries and trends.
  • Explainability analysis: SHAP identifies vapor pressure deficit and evapotranspiration as the most influential meteorological drivers for forecasting during 2007 to 2011.The analysis aggregates features annually and trains Lasso on pre-2007 samples before testing on 2007–2011.
  • Extreme-event adaptation: Four earlier Corn Belt extreme-event years are weighted in the models because they experienced moisture deficits and heat stress associated with yield reductions.The identified years are 1983, 1988, 1993, and 2002.
  • Feature selection: The selected SHAP features are applied across both machine-learning and deep-learning models.The paper refers to the combined preprocessing, feature-selection, and reweighting changes as data and model adaptation.
  • Transfer adaptation: TrAdaBoost.R2 defines the target domain using extreme-event and recent pre-2012 samples to focus on feature-distribution mismatch.The remaining pre-2012 samples form the source domain.

4 Evaluation

The evaluation compares model performance on the 2012 extreme drought year and the normal 2013 year using all features or SHAP-selected features. Models are standardized and assessed with cross-validation and multiple error-oriented metrics.

  • Metrics: The models are evaluated on four metrics: R2, RMSE, bias, and variance.R2 measures captured variance, RMSE measures prediction-error magnitude, bias indicates under- or overprediction, and variance describes residual variation around its mean.
  • Results analysis: The reported results validate model performance on both extreme-drought and normal years, while examining VITA attention weights by year.The evaluation includes attention analyses for the VITA model alongside predictive comparisons.
  • Test scenarios: The test scenarios are the 2012 extreme drought year and the normal 2013 year.This pairing is used to compare model performance under extreme and normal conditions.
  • Feature settings: Each model is tested with either all available features or SHAP-selected features.For Tradaboost.R2 and VITA, the SHAP-feature setting corresponds to the adapted Tradaboost.R2[+] and VITA[+] models.
  • Validation: All models use standardized features and 20-fold cross-validation for the evaluation setup.The VITA model is fine-tuned from a small pretrained checkpoint, using prior historical data with five-fold cross-validation for training.

4.3 Results

Models were evaluated on the extreme drought year 2012 and normal year 2013 using all features or SHAP-selected VPD and ETR features. VITA outperformed the comparative models, while feature selection improved classical ML performance and attention focused on the corn growing season.

  • Model comparison: Models trained on SHAP-selected VPD and ETR features improved prediction performance compared with models using all features.The study identifies VPD and ETR as influential drivers because they combine temperature, humidity, and radiation effects.
  • Model comparison: VITA without modifications outperformed all competitive models, while VITA and VITA[+] had comparable R2 and RMSE scores.The comparison covered extreme-year and normal-year evaluation settings.
  • Model comparison: VITA performed well under feature distribution mismatch and data irregularities, unlike classical ML models.The authors report that meteorological drivers explained more than half of yield variation.
  • Normal-year validation: On the normal year 2013, VITA[+] had the best forecasting performance, whereas TrAdaBoost.R2[+] showed no improvement over TrAdaBoost.R2 with SHAP features.The authors associate the reduced effectiveness of domain adaptation with the absence of feature distribution mismatch in 2013.
  • Attention analysis: VITA and VITA[+] assigned higher attention to weeks 25–33, corresponding to June–July corn growing-season months, across 2012 and 2013.The authors interpret this consistent emphasis as evidence that self-supervision prioritizes growing-phase months.

4.4 Ablation Study

The ablation study separates the effects of year weighting, SHAP feature selection, and growing-season attention in TrAdaBoost.R2[+] and VITA[+]. Growing-season attention most strongly affected VITA[+], while the two modifications worked best together for TrAdaBoost.R2[+].

  • VITA[+] ablation: Removing year weights or feature attention affected VITA[+] performance, with growing-season attention having the largest effect.The ablations varied year weights, selected features, and attention during the growing season.
  • Overall ablation results: VITA still outperformed all VITA[+] variations, indicating that year weights and growing-season attention were useful when applied together as modifications.The comparison refers to results in Tables 4 and 2.
  • TrAdaBoost.R2[+] ablation: SHAP features VPD and ETR improved TrAdaBoost.R2[+], whereas extreme-year weights reduced performance when paired with all features.The ablation compared an all-feature weighted model with a SHAP-feature model without extreme-year weights.

4.5 Machine learning using ’few’ test data samples

The study tests whether a small number of labeled extreme-year county samples can improve machine-learning forecasts. Randomly sampled and weighted samples performed best at n = 20, while VPD-based stratification underperformed random sampling.

  • Random sampling: For n = 20, the Lasso model achieved its best performance, which improved as the sample weights increased.The experiments used 5 to 30 randomly selected samples with weights ranging from 5 to 200.
  • Sampling strategy: The experiments used 20 counties selected through uniform sampling across sorted counties before applying stratified sampling with equal coverage.The selected counties were visualized in Figure 4.
  • Sampling strategy: VPD-based stratified sampling underperformed random sampling, indicating that yield variability cannot be determined solely using VPD.The counties were sorted by mean April–August growing-season VPD before stratified selection.
  • Implications: The results suggest that a small amount of labeled county data can provide useful information for forecasting yield variability during the extreme drought year.Meteorological data from the growing season can support forecasting, but the optimal number of test samples remains unresolved.

5 Discussion and Future Work

The discussion links the 2012 forecasting challenge to feature distribution mismatch and spatiotemporal data irregularities. It reports especially large VPD and ETR deviations in 2012 and identifies more complete, higher-resolution, and more local data as future directions.

  • Feature distribution mismatch: The 2012 feature distribution differed from training data, with especially large z-score deviations for VPD and ETR.Precipitation and minimum temperature were less skewed than VPD and ETR relative to their observed values.
  • Feature distribution mismatch: VPD and ETR had heavier tails in 2012 than in normal 2009 and extreme 1993, with 2012 deviating from the earlier drought year.The authors report that 2012 contained more extreme values than historically observed.
  • Data limitations: The gridMET dataset uses only 210/365 daily values, which may miss long-term precipitation and temperature trends contributing to yield variation.The authors propose studies using complete data to address this irregularity.
  • Data limitations: VITA relies on the coarse NASA Power dataset at 0.5° resolution, motivating future use of higher-resolution climate data for pretraining.The discussion also identifies higher-resolution grid data such as GFDL AM4 or ERA5 as a possible direction.
  • Future work: County-level forecasting differs from forecasting smaller 5x5 km regions, so more local prediction would help test VITA’s learning and performance.The authors also note that models can underperform when training data is scarce.

6 Conclusion

VITA consistently outperforms non-deep-learning models for crop-yield forecasting under the 2012 drought’s feature distribution mismatch and spatiotemporal irregularities. Model adaptation improves ML performance, while VITA shows little to no improvement from these modifications.

  • VITA shows little to no improvement from the model-adaptation modifications.Its performance advantage remains present with or without the modifications.
  • VITA consistently outperforms machine learning models with or without model-adaptation modifications.The comparison concerns forecasting crop yield during feature distribution mismatch scenarios such as the 2012 US Corn Belt drought.
  • Pretrained deep learning models like VITA can overcome feature distribution mismatch and spatiotemporal irregularities using meteorological drivers alone.The conclusion links this capability to strong prediction performance despite the dataset’s irregularities.
  • SHAP feature selection and extreme-year weighting improve the performance of machine learning models.These modifications are described as model adaptation for the data.
  • ML explainability-based modifications are most useful when large-scale models cannot be deployed.The study contrasts this use case with deep learning models whose performance improves as data quality and scale increase.
Loading 2608.17971v1…