Cross-Dataset Generalization of Machine Learning Models in Plant Biological Research
Qubahan Educational Journal · 07.11.2026 · 3 (4) · 60-77
DOI: 10.48161/qej.v3n4a25 ↗Kurz gefasst
Fachartikel in der Zeitschrift Qubahan Educational Journal, online erschienen am 07.11.2026. Verfasst von Helen Grace D. Felix. Die Arbeit nennt 33 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).
Abstract (Originalsprache)
Plant biological prediction models are commonly evaluated within the datasets from which they are trained, although practical deployment often requires transfer across populations, sites, years, phenotyping systems, and crop species. This study evaluated cross-dataset generalization of machine-learning models using harmonized maize, rice, soybean, and sorghum data. We aligned productivity, stature, and phenology using ontology-backed functional domains, and processed genomic, climate, soil, and management predictors with training-only transformations. We compared ridge and elastic-net regression, random forest, XGBoost, a conventional deep neural network, VAE-based representation learning, CORAL-DNN, and domain-adversarial neural learning under grouped within-dataset validation, genotype and temporal holdouts, single-source transfer, leave-one-dataset-out testing, multi-source-to-target transfer, and an external sorghum evaluation. The analysis combined RMSE, MAE, R², Spearman correlation, transfer retention, group-aware conformal coverage, dataset-shift diagnostics, clustered uncertainty estimation, ablation, and cross-dataset feature stability. DANN achieved the lowest within-dataset macro sRMSE (0.67) and the lowest leave-one-dataset-out sRMSE (0.81), while XGBoost reached 0.72 and 0.91, respectively. Performance degradation increased with source-target shift (Spearman ρ=0.76, 95% CI 0.61-0.87), and external 90% interval coverage ranged from 0.68 for ridge to 0.86 for DANN. The results show that domain-aware training improved transportability and uncertainty behavior while preserving biologically coherent predictor rankings across the four studied crops and the explicitly harmonized functional domains; they do not establish universal cross-species or raw-trait equivalence.
Zitieren
APA
Felix, H. G. D. (2026). Cross-Dataset Generalization of Machine Learning Models in Plant Biological Research. Qubahan Educational Journal, 3(4), 60-77. https://doi.org/10.48161/qej.v3n4a25
Vancouver
Felix HGD. Cross-Dataset Generalization of Machine Learning Models in Plant Biological Research. Qubahan Educational Journal. 2026;3(4):60-77. doi:10.48161/qej.v3n4a25
BibTeX
@article{Felix2026,
author = {Felix, Helen Grace D.},
title = {Cross-Dataset Generalization of Machine Learning Models in Plant Biological Research},
journal = {Qubahan Educational Journal},
year = {2026},
volume = {3},
number = {4},
pages = {60--77},
doi = {10.48161/qej.v3n4a25}
}
EUVIMED