An Open Field Phenomics Resource for Multimodal Maize Yield Prediction Across Divergent Environments
Synopsis
This work releases curated 2020-2021 Genomes to Fields imagery from 356 drone flights across 19 environments covering 1,180 maize hybrids, and uses functional principal components of vegetation index and weather trajectories together with genomic information for yield prediction and QTL mapping, finding that combined genomic and phenomic kernels raised held-out hybrid prediction to r = 0.501 when environments were represented in training and 0.408 when environments were also withheld, that accumulated growing degree days offered no consistent advantage over days after planting, that weather contributed modest task-dependent gains, and that NGRDI functional principal components repeatedly mapped to quantitative trait loci on chromosomes 3 and 7.
Interpretation
The study releases what it describes as the largest public field image reference data set to date: 19 G2F environments, 356 temporal drone flights, and 1,180 maize hybrids, with a reproducible analysis pipeline. Compared with previously released genotype, phenotype, and weather data, this resource integrates and publicly shares temporal canopy imagery collected on non-synchronous flight schedules across environments for the same genetic material. Resource scale and provenance are stated in the abstract, introduction, and methods; imagery was stitched with Agisoft Metashape, plots delineated with UAStools and QGIS, and vegetation indices extracted with FIELDimageR, with processed and unprocessed imagery available via Data to Science.
Combining functional principal components of vegetation index and weather trajectories with genomic kernels improved grain yield prediction, with gains appearing in both familiar and unfamiliar environment tasks. Prior multimodal prediction often relied on scalar or single-time-point features; here irregularly sampled trajectories are represented by functional principal components, and two time scales (DAP and AGDD) plus 14 kernel models are compared systematically. Combined genomic and phenomic kernels improved prediction on average by 37.3% in CV2/CV1 and 39.4% in CV0/CV00; M8.G.P-Int was top-performing for CV2 and CV1, and M10 reached CV0 r = 0.440 for DAP and 0.431 for AGDD, with LOEO r = 0.449 and RMSE 2.55 t/ha for DAP.
Indexing temporal data by accumulated growing degree days (AGDD) instead of days after planting (DAP) did not consistently improve yield prediction, and weather information gave smaller, task-dependent gains. The study explicitly tested the physiologically motivated AGDD time scale expected to better align maize phenology, and reports performance similar to DAP, contrasting with the hypothesis that AGDD would be superior. Under M10, CV0 correlations were 0.440 for DAP and 0.431 for AGDD, and LOEO correlations were 0.449 and 0.439; the PTR weather kernel did not consistently improve prediction, which the authors connect in the discussion to prior reports of limited weather contribution.
QTL mapping of NGRDI functional principal components repeatedly detected loci on chromosomes 3 (113-185 Mb) and 7 (127-138 Mb) across environments and testers, with identifiable MAGIC founder allele effects. The analysis links the same temporal functional summaries used for prediction directly to genetic variation, with chromosome 7 NGRDI FPC1 peaks recurring in nine environment, tester, and FPCA-type combinations. A total of 167 significant peaks were identified (71 for DAP and 96 for AGDD), thresholds were set empirically per tester-environment combination using 1000 permutations, linkage disequilibrium blocks in the chromosome 7 interval were small and localized, the NKH8431 allele had a strong positive effect on FPC1, and the PHJ40 allele was negative in all combinations.
Perspective
This resource and analysis target the maize breeding and phenomics community, applying to settings where temporal drone imagery, weather, and genomic markers are used to predict grain yield and to map loci for temporal canopy traits; the design is specific to the 19 environments of the 2020-2021 G2F experiments and the WI-SS-MAGIC hybrid population, with whole-season vegetation index and weather trajectories as model inputs.
The authors note that weather observations were limited to 19 distinct whole-environment samples and that learning representations generalizing to unseen environments remains difficult for flexible models; phenomic prediction may overestimate accuracy because phenomic data are influenced by non-additive and environmental effects, and approaches to adjust accuracy values from models using both genomic and phenomic components still need development; whether larger and more diverse training sets will realize the expected benefits of the TNP remains to be tested; the present analyses used entire-season trajectories, so their utility for earlier-season decisions remains to be established; and the chromosome 3 and 7 regions still require finer mapping and validation.