henhao
Sign in
Plant BiologybioRxiv
Preprint

An open field phenomics resource for multimodal maize yield prediction across divergent environments

Aaron J. DeSalvioiD, Peiman MohseniiD, Alper AdakiD, Seth C. MurrayiD, Mustafa ArıkiD, Raymond K. W. WongiD, Jinha JungiD, Dayane Cristina LimaiD, Alejandro Castro Aviles, Edward S. BuckleriD, Nick G. DuffieldiD, Jode W. EdwardsiD, David ErtliD, Sherry A. Flint-GarciaiD, Michael A. GoreiD, Candice N. HirschiD, James Brendan HollandiD, Shawn M. KaeppleriD, Jarrod O. MilleriD, M. Cinta RomayiD, James C. SchnableiD, Maninder P. Singh, Erin E. SparksiD, Addie ThompsoniD, Jacob D. WashburniD, , Natalia de LeóniD

Abstract

Temporal drone phenotyping captures crop development, but irregular flight schedules complicate comparisons across environments. We release curated imagery from 356 flights across 19 Genomes to Fields environments containing 1,180 maize (Zea mays L.) hybrids. To evaluate its utility, we integrated functional principal components of vegetation index and weather trajectories with genomic information. Combinined genomic and phenomic kernels improved yield prediction, reaching correlations up to r = 0.501 for held-out hybrids in environments represented in training and 0.408 when environments were also withheld. Accumulated growing degree days offered no consistent predictive advantage over days after planting, and weather contributed modest, task-dependent gains. A transformer neural process learned directly from irregular observations, serving as a novel application of neural process models in agriculture. Mapping vegetation index functional principal components identified recurrent quantitative trait loci on chromosomes 3 and 7. This resource and its reproducible analyses guide the use of temporal spectral data for crop prediction and genetic discovery.

PDFDOI
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
1
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 7 Nov 2026 · from the full text of version 1 (18 Sept 2026) · CC BY 4.0

Drone imagery from 356 flights in 19 Genomes to Fields maize environments; vegetation-index functional principal components complement genomic kernels, predicting yield of held-out hybrids at r up to 0.501.

Genomic markers do not change over time, and drone flights fall on different dates in each environment, so temporal canopy data are hard to compare or combine across trials. The authors describe the 2020-2021 Genomes to Fields flight set as the largest public field image reference set to date; they use it to compare calendar vs thermal time indexing and kernel vs neural process models, and to link the same temporal summaries to QTL.

Key results

  1. ShownAdding phenomic matrices (functional PC scores of normalized green-red difference index, NGRDI, curves) to genomic matrices predicted grain yield better than either source alone. This held for new hybrids (CV1), new environments (CV0), both (CV00) and leave-one-environment-out (LOEO) validation.

    Genomic+phenomic models (M6 and above) raised correlation on average by 37.3% (CV2/CV1) and 39.4% (CV0/CV00) over either source alone; LOEO average r 0.427 vs 0.321 for single-source models (33.0%). Best r = 0.501 for held-out hybrids in trained environments, 0.408 with environments also withheld; M10 LOEO r = 0.449, root mean squared error (RMSE) 2.55 t/ha (days-after-planting scale). · Abstract; 2.4 Familiar environments–kernel-based…; 2.5 Unfamiliar environments–kernel-based…; 2.7 Leave-one-environment-out–kernel-based…

  2. ShownIndexing curves by accumulated growing degree days (AGDD) instead of days after planting (DAP) aligned development across sites better and spread NGRDI variation differently across components. It did not consistently improve yield prediction.

    NGRDI FPC1 explained 48.3% of functional variation with DAP (FPC2 24.7%) vs 35.5% with AGDD (FPC2 33.4%). M10 CV0 r = 0.440 (DAP) vs 0.431 (AGDD); LOEO r = 0.449 vs 0.439, RMSE 2.55 vs 2.77 t/ha. Pooled correlations with yield best linear unbiased estimates reached r = 0.55 (DAP FPC1) and 0.60 (AGDD FPC3). · 2.2 DAP– and AGDD-indexed FPCA captured…; 2.5 Unfamiliar environments–kernel-based…; 2.7 Leave-one-environment-out–kernel-based…; 3 Discussion

  3. SuggestedWeather variables summarized as functional PCs tracked environment-level yield. As a prediction kernel, however, the photothermal ratio (PTR) weather term added only small and inconsistent gains.

    Across 19 environments, PTR FPC1 correlated with environment-level yield at r = 0.66, photoperiod FPC1 at 0.62 and FPC2 of minimum temperature at 2 m at -0.64. The PTR kernel did not consistently help; M5.G.W-Int had slightly higher correlations than genomic-only M3.G-Int, with lower RMSE under DAP but higher under AGDD. · 2.2 DAP– and AGDD-indexed FPCA captured…; 2.5 Unfamiliar environments–kernel-based…; 2.7 Leave-one-environment-out–kernel-based…; 3 Discussion

  4. ShownA transformer neural process (TNP) read raw, irregularly timed vegetation-index and daily weather records plus genomic relationship rows. It slightly beat the best kernel model for new hybrids in trained environments (CV1) but had lower correlation in the other schemes.

    CV1 r = 0.512 vs 0.501. Best TNP correlations relative to kernel counterparts: -25.6% (CV2), +2.1% (CV1), -18.4% (CV0), -20.0% (CV00), with RMSE +32.4%, -0.4%, +0.2%, +0.4%. Adding genomic inputs gave a substantial gain over the vegetation-index-only TNP; weather gains were smaller and variable. · 2.6 TNP analysis of raw, irregular temporal…; 3 Discussion; 4.9 A neural process for multimodal yield…

  5. ShownQuantitative trait locus (QTL) mapping of NGRDI functional PC scores repeatedly detected regions on chromosomes 3 and 7 across testers and environments. At the chromosome 7 locus, alleles from different multi-parent advanced generation intercross (MAGIC) founders had opposite effects.

    167 significant peaks (71 DAP, 96 AGDD) from 29 tester-environment scans with 1000-permutation thresholds; concordant intervals at 113-185 Mb (chr 3) and 127-138 Mb (chr 7), where linkage disequilibrium blocks were small; chr 7 FPC1 peaks recurred in nine environment-tester-FPCA combinations; the NKH8431 allele raised FPC1, and PHJ40 alleles lowered it in all combinations. · 2.8 QTL mapping of functional PCs uncovered…; 4.10 Quantitative trait locus (QTL) mapping of…

Methods, data and reproducibility

Vegetation indices from RGB drone orthomosaics and daily weather series were summarized by functional principal component analysis (fdapace) on days-after-planting or accumulated growing-degree-day scales. Scores fed 14 Gaussian-process kernel models (BGLR) with genomic relationship matrices, compared with a transformer neural process on raw series, and were QTL-mapped (qtl2).

  • RGB drone imagery (orthomosaics and plot-level vegetation indices) from 2020-2021 Genomes to Fields trials, released via Data to Science (D2S) · maize (Zea mays) · n = 356 flights across 19 environments
  • Hybrid grain yield trials: WI-SS-MAGIC lines crossed with testers PHP02, PHK76 and PHZ51, plus commercial and experimental checks · maize (Zea mays) · n = 1,180 hybrids; 10,109 genotype/environment combinations
  • Hybrid genotypes from the 2024 G2F Genotype by Environment Prediction Competition (CyVerse) · maize (Zea mays) · n = 437,214 markers, 238,782 after filtering; 383 MAGIC inbreds
  • Daily weather from NASA POWER (via EnvRtype) and G2F in-field weather stations · n = 19 environments
Methods & tools
Agisoft Metashape 2.0.2, UAStools, QGIS, FIELDimageR, EnvRtype, fdapace, lme4, TASSEL / rTASSEL, AGHmatrix, BGLR, qtl2, MaizeMine v1.5
Reagents & environment
Agisoft Metashape Version 2.0.2 and MaizeMine v1.5 have version numbers; R package versions are not given in the main text. Metashape parameters and TNP settings are in the Supplementary Methods.
Assay conditions & compute
Portions ran on Texas A&M High Performance Research Computing; no hardware, runtime or memory figures are given.
Reproducibility gaps
The text cites a GitHub repository (code), D2S imagery with DOIs in Table S7, and CyVerse genotypes, but prints no URLs or accessions. TNP architecture and training details are only in the Supplementary Methods.
Limitations (6)
  • Covers 19 environments over two seasons and uses full-season trajectories. The authors call for sites beyond Texas and Wisconsin, tests in future seasons and broader germplasm, and prediction from truncated in-season series. · stated by the authors
  • Phenomic prediction may overstate accuracy because canopy data carry non-additive and environmental effects. No correction yet exists for models that use genomic and phenomic components together. · stated by the authors
  • Weather came from only 19 environment-level samples, which makes it hard to learn representations that transfer to unseen environments, especially for the flexible TNP. · stated by the authors
  • The chromosome 3 and 7 QTL regions still need validation and finer mapping. · stated by the authors
  • Some choices were made by looking at yield before cross-validation: PTR FPC1 was picked by its yield correlation across all 19 environments, and the component counts (5 DAP, 7 AGDD) were fixed after exploratory prediction runs. Both could inflate held-out accuracy. · noted in this summary
  • The TNP and kernel models differ in inputs (raw DAP-indexed series vs NGRDI FPCs plus a PTR kernel) and in seeds (five vs 10), so their performance gaps mix model class with representation. · noted in this summary

Who should read:Read this if you build genomic prediction models for multi-environment trials and want public drone imagery plus a leakage-aware way to turn irregular flight series into phenomic kernels.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISAdak '22Jin '24Sweet '25Liang '17Vižintin '26Creach '26
75% semantic matchPlant Biology
A. Adak, S. C. Murray, S. L. Anderson·G3·Nov 29, 2022