Beyond vegetation indices: A phenomic prediction strategy sharpens genetic signals from drone-based crop imaging
image:
Blue boxes represent the final genome-wide association study and are identical between the four approaches. The differences lie in the input of the response variable: A) the input is the visual score (VS) assigned by trained staff and adjusted for experimental design; B) replaces the VS by an individual specific vegetation index (VI) adjusted for experimental design; C) uses a phenomic prediction approach to predict the VS from phenomic data with a trained prediction model; D) instead of phenotypes adjusted to the experimental design only, genomic estimated breeding values (GEBVs) are used in the phenomic prediction of VS.
view moreCredit: The authors
By training statistical and machine-learning models to predict expert visual scores, the study demonstrates that phenomics can match or outperform traditional indices, particularly when combined with genomic information to strengthen genetic analyses.
Remote sensing has transformed crop phenotyping, but disease assessment from drone imagery still often relies on a single vegetation index, which can miss important biological signals. UAV-based high-throughput phenotyping now plays a central role in crop breeding, as multispectral and thermal sensors enable rapid, non-destructive monitoring of plant health. These complex datasets are typically reduced to indices such as NDVI or simple reflectance ratios, each capturing only part of plant physiology. However, index performance varies across environments and disease conditions, especially when symptoms are confounded by stress or canopy structure. Phenomic prediction provides an alternative by integrating multiple traits to predict target phenotypes, raising the question of whether such predictions can improve genetic analyses such as genome-wide association studies.
A study (DOI: 10.1016/j.plaphe.2025.100134) published in Plant Phenomics on 5 November 2025 by Johannes W.R. Martini’s team, International Maize and Wheat Improvement Center-CIMMYT, shows phenomic prediction combining remote sensing and genomic breeding values outperforms single vegetation indices in disease phenotyping and genetic analysis.
The study systematically evaluated whether phenomic prediction can outperform individual vegetation indices in disease resistance phenotyping by integrating UAV-derived multispectral and thermal data with statistical and machine-learning models. As a benchmark, absolute correlations between 15 remote-sensing traits and human-assigned visual scores (VS) were calculated across six population-by-year datasets, identifying the green-to-red reflectance ratio (G) as the strongest correlate in five cases, and NDVI in one. Building on this reference, researchers trained phenomic models to predict VS using two predictor sets—five basic spectral wavelengths and an expanded set including ten vegetation indices—while comparing linear ordinary least squares (OLS), regularized regressions (ridge regression and LASSO), and non-linear approaches (artificial neural networks, ANN; gradient boosted regression trees, GBRT). Using design-adjusted phenotypes, a simple OLS model based on five wavelengths (BT-OLS) matched or exceeded the benchmark correlation in 13 of 30 out-of-set predictions and outperformed vegetation index G in GWAS signal strength in 14 cases, typically identifying the same major resistance locus. In contrast, extending OLS to all 15 traits led to pronounced overfitting, with good performance confined to training data and weak generalization. Regularization mitigated this problem: ridge regression and LASSO applied to all traits (AT-RR, AT-LASSO) exceeded the benchmark in 18–19 of 30 cases and outperformed G in up to 16 GWAS comparisons, while non-linear ANN models provided only marginal additional gains and GBRT performed comparatively poorly on phenotypic data. The most substantial improvement emerged when phenotypes were replaced with genomic estimated breeding values (GEBVs), which isolate additive genetic effects and reduce environmental noise. When phenomic models were trained on GEBVs, GWAS signals became markedly stronger and more consistent: AT-RR based on GEBVs outperformed vegetation index G in 26 of 30 cases while pinpointing nearly identical loci with higher statistical power, and AT-ANN-GEBV showed similarly strong, though slightly weaker, gains. Binomial tests confirmed that GEBV-based phenomic models clearly outperformed phenotype-based approaches, highlighting genomic adjustment as the primary driver of enhanced genetic signal detection.
The findings suggest a shift in how remote sensing data should be used in breeding and genetic research. Rather than searching for the “best” vegetation index, breeders and geneticists can deploy phenomic models as flexible, data-driven indices tailored to specific traits and crops. This approach is especially valuable when disease symptoms are subtle or variable across environments.
###
References
DOI
Original Source URl
https://doi.org/10.1016/j.plaphe.2025.100134
About Plant Phenomics
Plant Phenomics is dedicated to publishing novel research that will advance all aspects of plant phenotyping from the cell to the plant population levels using innovative combinations of sensor systems and data analytics. Plant Phenomics aims also to connect phenomics to other science domains, such as genomics, genetics, physiology, molecular biology, bioinformatics, statistics, mathematics, and computer sciences. Plant Phenomics should thus contribute to advance plant sciences and agriculture/forestry/horticulture by addressing key scientific challenges in the area of plant phenomics.
Journal
Plant Phenomics
Method of Research
Experimental study
Subject of Research
Not applicable
Article Title
Comparing statistical ‘phenomic prediction’ models for remote-sensing-based phenotyping of maize susceptibility to common rust
No comments:
Post a Comment