EUVIMEDEUVIMED entdecken
Crossref

Interpretable Transformer-Based Yield Forecasting Under Climate Extremes: A Multi-Layered Explainability Framework and Drought-Year Validation on the Canadian Prairies (2021)

Seema Sharma, Dr. Praveen Gupta

International Journal of Emerging Technologies and Innovative Research · 27.07.2026 · 28

DOI: 10.48175/ijetir-6655 ↗

Kurz gefasst

Fachartikel in der Zeitschrift International Journal of Emerging Technologies and Innovative Research, online erschienen am 27.07.2026. Verfasst von Seema Sharma und 1 weiteren Autor:innen. Die Arbeit nennt 56 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Accurate and interpretable crop yield forecasting under climate extremes is essential for food security, agricultural risk management, and policy decision-making. This study presents a multi-layered explainability framework for transformer-based yield prediction, validated during the severe 2021 drought across the Canadian Prairies — a natural out-of-distribution stress test. We trained a Temporal Fusion Transformer (TFT) on a multi-source geospatial data stack encompassing Sentinel-2 surface reflectance, MODIS vegetation indices, RADARSAT constellation synthetic aperture radar backscatter, SMAP soil moisture, ERA5-Land climate reanalysis, and AAFC crop inventory maps. The model achieved strong overall performance (R² = 0.84, RMSE = 0.42 t ha⁻¹) across 2017–2020 but degraded predictably during the 2021 drought (R² = 0.71, RMSE = 0.68 t ha⁻¹). To understand this degradation, we deployed a comprehensive explainable artificial intelligence (XAI) pipeline: (1) TFT native variable selection networks revealed dynamic reweighting from vegetation indices toward soil moisture and thermal stress indices; (2) attention rollout analysis identified a temporal shift in model focus from mid-season (July) to early-season (June) predictors approximately 4–6 weeks before harvest; (3) SHAP analysis quantified global and local feature attributions, showing that evaporative stress index (ESI) and surface soil moisture anomalies became dominant drivers in 2021; (4) LIME confirmed instance-level fidelity of explanations; and (5) Grad-CAM applied to spatial input rasters highlighted drought-sensitive regions in southwestern Saskatchewan and southeastern Alberta. Critically, early-stress signals emerged in attention and SHAP time series by late June 2021, supporting Hypothesis H4 that interpretable transformer outputs can provide actionable early warnings 4–6 weeks’ pre-harvest. These findings demonstrate that multi-layered explainability not only builds trust in deep learning forecasts but also reveals mechanistic insights into crop response to extreme climate events, with direct implications for crop insurance, drought relief allocation, and on-farm adaptation decisions.

Zitieren

APA

Sharma, S., & Gupta, D. P. (2026). Interpretable Transformer-Based Yield Forecasting Under Climate Extremes: A Multi-Layered Explainability Framework and Drought-Year Validation on the Canadian Prairies (2021). International Journal of Emerging Technologies and Innovative Research, 28. https://doi.org/10.48175/ijetir-6655

Vancouver

Sharma S, Gupta DP. Interpretable Transformer-Based Yield Forecasting Under Climate Extremes: A Multi-Layered Explainability Framework and Drought-Year Validation on the Canadian Prairies (2021). International Journal of Emerging Technologies and Innovative Research. 2026:28. doi:10.48175/ijetir-6655

BibTeX
@article{Sharma2026,
  author = {Sharma, Seema and Gupta, Dr. Praveen},
  title = {Interpretable Transformer-Based Yield Forecasting Under Climate Extremes: A Multi-Layered Explainability Framework and Drought-Year Validation on the Canadian Prairies (2021)},
  journal = {International Journal of Emerging Technologies and Innovative Research},
  year = {2026},
  pages = {28},
  doi = {10.48175/ijetir-6655}
}
Kostenlos starten →