EUVIMEDEUVIMED entdecken
Crossref

Uncertainty Aware Crop Yield Forecasting Under Extreme Weather Conditions Using Hybrid Temporal Models

Zhang Wang, Anandakumar Haldorai

Journal of Smart and Sustainable Farming · 22.06.2026 · 122

DOI: 10.64026/jssf/2026011 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Smart and Sustainable Farming, online erschienen am 22.06.2026. Verfasst von Zhang Wang und 1 weiteren Autor:innen. Die Arbeit nennt 28 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The challenge of forecasting yields is increased and complicated by extreme weather events like heatwaves, droughts, floods and cold snaps, as a forecast that is true on average becomes highly misleading in the seasons when it is needed most by those who need it most: farmers, insurers and planners. A framework for uncertainty-aware forecasting that combines a one-dimensional convolutional front-end, a heteroscedastic Gaussian output, a bidirectional gated recurrent unit and temporal attention, as well as normalizing with split-conformal prediction is presented, followed by presenting five members in a deep ensemble, and calibrating the prediction intervals with normalized split-conformal prediction. The evaluation of the framework is done on a dataset of simulated extreme events over 30 years and 100 regions with a strict temporal split (3000 region-seasons, 24 weekly steps, the number of events increasing with the simulation period). The ensemble outperforms the six point-forecast baselines with a test RMSE of 0.557 t/ha (R² = 0.890) and error in extreme seasons of 0.587 t/ha (gradient boosting = 0.671 t/ha). The 90% intervals of its 90% intervals cover 92.7% of the seasons, compared to 83–84% for constant-width conformal gradient boosting for seasons corresponding to heatwave, drought and compound seasons. The calibrated ensemble maintains 90.2% coverage under the same type of intensified extremes it didn't see in training, compared to 87.5% for the uncalibrated ensemble and 79.7% for gradient boosting. The gains are found primarily in the heteroscedastic output and the ensemble but not the convolutional, bidirectional or attention components and conclusions are constrained by simulated data.

Zitieren

APA

Wang, Z., & Haldorai, A. (2026). Uncertainty Aware Crop Yield Forecasting Under Extreme Weather Conditions Using Hybrid Temporal Models. Journal of Smart and Sustainable Farming, 122. https://doi.org/10.64026/jssf/2026011

Vancouver

Wang Z, Haldorai A. Uncertainty Aware Crop Yield Forecasting Under Extreme Weather Conditions Using Hybrid Temporal Models. Journal of Smart and Sustainable Farming. 2026:122. doi:10.64026/jssf/2026011

BibTeX
@article{Wang2026,
  author = {Wang, Zhang and Haldorai, Anandakumar},
  title = {Uncertainty Aware Crop Yield Forecasting Under Extreme Weather Conditions Using Hybrid Temporal Models},
  journal = {Journal of Smart and Sustainable Farming},
  year = {2026},
  pages = {122},
  doi = {10.64026/jssf/2026011}
}
Kostenlos starten →