Deep ensembles under distribution shift for crop yield prediction

Investigate and characterize the impact of distribution shifts on deep ensemble models for crop yield prediction, and ascertain whether deep ensembles provide robustness improvements compared to deterministic baselines when training and testing across differing data distributions such as disparate years or regions.

Background

The paper documents severe performance degradation of crop yield prediction models under real-world distribution shifts across years and regions and notes that only a few studies have used deep ensembles in this domain, typically for uncertainty estimation. The authors point out that understanding the behavior of deep ensembles under such shifts remains insufficiently explored.

YieldSAT spans multiple countries, crops, and years, creating natural covariate shifts. The experiments show notable performance drops with leave-one-year-out and leave-one-region-out splits, motivating a systematic assessment of whether deep ensembles can mitigate such shifts.