Multi-Task Deep Learning for Simultaneous Prediction of Agricultural Biological Traits
Qubahan Educational Journal · 07.11.2026 · 3 (4) · 78-99
DOI: 10.48161/qej.v3n4a26 ↗Kurz gefasst
Fachartikel in der Zeitschrift Qubahan Educational Journal, online erschienen am 07.11.2026. Verfasst von Fatima E. Supan. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).
Abstract (Originalsprache)
Agricultural traits are interdependent, yet prediction pipelines often fit each phenotype independently and discard information carried by correlated traits. This study evaluated whether multi-task deep learning improved simultaneous prediction of maize grain yield, plant height, anthesis date, silking date, and grain moisture by learning a shared genomic-environmental representation while retaining trait-specific output functions. We conducted an empirical analysis using Genomes to Fields records, integrating genotype principal components with target-independent early-season weather, soil, and management covariates. The analytical cohort comprised 16,728 plot records from 1,102 hybrids across 41 location-year environments; 2014-2016 formed the development period, and 2017 remained isolated for temporal confirmation. We compared a task-balanced hard-sharing network with elastic net, random forest, XGBoost, multi-trait GBLUP, matched single-task neural networks, and a cross-stitch multi-task comparator. Validation included nested environment-grouped testing, genotype-ID transfer, kinship-cluster holdout, joint unseen-genotype × unseen-environment stress testing, temporal confirmation, and an independent protocol-replication benchmark using the European DROPS maize panel. The multi-task model retained the lowest standardized error across the principal G2F regimes, with the strongest gains for anthesis and silking; the advantage narrowed but remained positive under kinship-controlled transfer and independent protocol replication. Environment-level leave-one-out sensitivity and exact sign-based inference supported the temporal ranking without treating bootstrap resamples as new biological replicates. The findings indicate that multi-task learning is most defensible when task covariance is biologically justified, predictor timing is target-independent, genetic relatedness is controlled explicitly, and validation matches deployment conditions. The framework provides a rigorous basis for simultaneous prediction of agricultural traits under multiple forms of distribution shift.
Zitieren
APA
Supan, F. E. (2026). Multi-Task Deep Learning for Simultaneous Prediction of Agricultural Biological Traits. Qubahan Educational Journal, 3(4), 78-99. https://doi.org/10.48161/qej.v3n4a26
Vancouver
Supan FE. Multi-Task Deep Learning for Simultaneous Prediction of Agricultural Biological Traits. Qubahan Educational Journal. 2026;3(4):78-99. doi:10.48161/qej.v3n4a26
BibTeX
@article{Supan2026,
author = {Supan, Fatima E.},
title = {Multi-Task Deep Learning for Simultaneous Prediction of Agricultural Biological Traits},
journal = {Qubahan Educational Journal},
year = {2026},
volume = {3},
number = {4},
pages = {78--99},
doi = {10.48161/qej.v3n4a26}
}
EUVIMED