Pellet, D. | Search bioRxiv

bioRxiv2025

Explainable AI shows climate impacts on wheat yields: insights from 30 years of field data.

Wheat (Triticum aestivum L.) is amongst the worlds most important staple crops and primary food source for an estimated 35% of the global population. Climate impacts have caused global yield stagnation and quantifying the climatic variables influencing wheat yield is critical to anticipate yield losses and design climate-resilient agricultural strategies. Here, we use a unique 30-year dataset on winter wheat variety trials in six sites across Switzerland, explainable artificial intelligence (XAI) and interpretable machine learning (IML) methods (i.e., decision trees and gradient boosting models combined with post hoc tests) to elucidate climate drivers on wheat yields. We showed based on 405 varieties and over 10,000 observations, that climatic variables such as cumulative solar radiation, precipitation from sowing to harvest and genotype makeup are significant yield drivers. Partial dependence plots and variable interaction analyses revealed, for example, a yield plateau above cumulative solar radiation levels of [~]3000 MJ m-{superscript 2}, suggesting complex genotype-by-environment interactions. These findings suggest that XAI adds important biological interpretability to predictive performance, and reveals the mechanisms how climate affects wheat yields. Our methodological framework and results can inform breeding activities, agronomic management, and adaptation strategies under climate change across environmental conditions in Switzerland and with global ramifications.

plant biology↗

bioRxiv2025

The plant-time-bender model: predicting yield through wheat's perception of time

To address challenges in food security, a better understanding of crop performance under varying and changing environmental conditions is required. Plant Time Warping (PTW) is a deep learning model that integrates high-throughput field phenotyping data with genomic and environmental information to predict wheat yield. PTW leverages image time series, genetic markers, and environmental covariates to learn genotype-specific physiological responses to temperature and vapor pressure deficit. Compared to mere genomic prediction models, PTW demonstrates superior performance when predicting yield in unseen environments across 48 year-locations in Europe. The PTW model captures non-linear growth responses varying with phenological stages and identifies distinct patterns associated with yield performance and stability. Specifically, varieties with higher yield stability exhibit reduced sensitivity to vapor pressure deficit around 1.5 kPa and distinctive temperature responses during emergence and senescence. The learned response pattern enable retrospective and prospective yield predictions, providing a foundation for location-specific variety recommendations and targeted breeding strategies. The integration of phenomic, genomic, and enviromic data has the potential to substantially advance research in climate adaptation strategies for crop production by addressing generalization challenges of predictions to novel environmental conditions. HighlightWe present a novel deep learning model that seamlessly combines high-throughput image data, genomic data, and weather data, enabling better crop predictions for future climates.

genomics↗