EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-ND 4.0

Uncertainty-Calibrated Machine Learning for Complex Agricultural Biological Data

Sudha Bishnoi

Qubahan Educational Journal · 07.11.2026 · 3 (4) · 38-59

DOI: 10.48161/qej.v3n4a24 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Qubahan Educational Journal, online erschienen am 07.11.2026. Verfasst von Sudha Bishnoi. Die Arbeit nennt 34 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Agricultural prediction increasingly integrates genomic, phenotypic, environmental, soil, and management information, yet high point accuracy alone is insufficient when models are transferred across genotypes and production environments. This study evaluated a leakage-controlled, uncertainty-calibrated framework for maize grain-yield prediction using the public Genomes to Fields (G2F) 2014–2017 resource specification. The analytical frame comprised 31,842 plot records, 1,987 hybrids, and 71 site-year environments. Ridge regression, random forest, XGBoost, a heteroscedastic deep ensemble, genomic best linear unbiased prediction (GBLUP), and a reaction-norm G×E genomic benchmark were evaluated under nested environment-grouped cross-validation, a 2017 temporal holdout, leave-genotype-out validation (CV1), genotype-by-environment cold-start validation (CV00), leave-site-out validation, a genetic-cluster holdout, and an environmental-shift stress test. A prediction-horizon sensitivity analysis distinguished pre-planting, flowering-stage, and grain-fill information availability. We separated predictive uncertainty into aleatoric and epistemic components and calibrated locally scaled intervals using a finite-sample, group-aware hierarchical conformal procedure with group-matched bootstrap confidence intervals. XGBoost retained the lowest mean RMSE in the principal validation settings (1.17 Mg ha−1 in environment-grouped CV; 1.31 in the 2017 temporal holdout; 1.42 in CV1; and 1.59 in CV00), while the reaction-norm genomic benchmark narrowed the advantage of nonlinear machine learning and the deep ensemble remained statistically competitive. Error increased as genomic relatedness to the training set decreased and as both genotype and environment became novel. Environment-blocked calibration achieved 0.90 overall and 0.89 macro-environment coverage for nominal 90% XGBoost intervals, with mean width 3.49 Mg ha−1. Within the G2F trial resource, these results support biologically structured validation, breeding-relevant benchmarks, explicit information cutoffs, and group-aware uncertainty calibration; no claim of cross-program external validation is made.

Zitieren

APA

Bishnoi, S. (2026). Uncertainty-Calibrated Machine Learning for Complex Agricultural Biological Data. Qubahan Educational Journal, 3(4), 38-59. https://doi.org/10.48161/qej.v3n4a24

Vancouver

Bishnoi S. Uncertainty-Calibrated Machine Learning for Complex Agricultural Biological Data. Qubahan Educational Journal. 2026;3(4):38-59. doi:10.48161/qej.v3n4a24

BibTeX
@article{Bishnoi2026,
  author = {Bishnoi, Sudha},
  title = {Uncertainty-Calibrated Machine Learning for Complex Agricultural Biological Data},
  journal = {Qubahan Educational Journal},
  year = {2026},
  volume = {3},
  number = {4},
  pages = {38--59},
  doi = {10.48161/qej.v3n4a24}
}
Kostenlos starten →