In maize, scientists working with field trials of common rust used an unmanned aerial vehicle platform equipped with multispectral and thermal sensors to monitor disease expression across multiple populations and years. Instead of relying on a single vegetation index such as NDVI or a green to red reflectance ratio, they trained statistical and machine learning models on a suite of 15 remote sensing traits to predict expert visual scores of disease severity.
The team first benchmarked performance by correlating each individual trait with human-assigned visual scores across six population by year datasets, finding that the green to red reflectance ratio was the strongest correlate in five cases and NDVI in one. Building on this reference, they compared a basic ordinary least squares model using only five spectral wavelengths with models that incorporated all traits and applied regularization or nonlinear learning.
A simple five-wavelength model matched or exceeded the best individual index in almost half of 30 out-of-set prediction scenarios, and it often delivered stronger genome-wide association study signals for the main rust resistance locus than the single best index. When the researchers expanded to all 15 traits without regularization, performance deteriorated due to overfitting, with good fits confined to the training data and weak generalization to new populations or environments.
Regularized models corrected that weakness. Ridge regression and LASSO models trained on all traits exceeded the benchmark vegetation index in most prediction cases and improved GWAS signal strength in many comparisons, while nonlinear artificial neural networks provided only limited additional gains and gradient boosted trees performed poorly on the phenotypic data. These results indicate that using the full phenomic feature set with appropriate regularization is more effective than searching for a single best index.
The most pronounced improvement came when the team replaced raw phenotypes with genomic estimated breeding values, which isolate additive genetic effects and reduce environmental noise in the trait values. Phenomic prediction models trained on these genomic breeding values delivered markedly stronger and more consistent GWAS signals, with ridge regression outperforming the benchmark index in nearly all comparisons while still identifying the same key resistance loci with higher statistical power.
Binomial tests confirmed that genomic breeding value based phenomic models systematically outperformed phenotype based approaches across the evaluated scenarios. The authors conclude that breeding and genetics programs should treat phenomic models as flexible, data driven indices tuned to specific traits and populations rather than focusing efforts on a single vegetation index, particularly in situations where symptoms are subtle or confounded by stress and canopy structure.
In parallel, peanut researchers have developed an end to end, drone based phenotyping workflow that uses Meta's Segment Anything Model to automatically delineate individual plots and canopies in high resolution images. Their system ingests UAV imagery from breeding nurseries, applies SAM to segment peanut plants from soil and background, and then calculates plot level canopy and architecture metrics such as cover, shape, and height proxies.
This segmentation pipeline removes the need for manual plot boundary tracing and many hand measured traits that historically consumed substantial time and introduced observer bias. Once the model is calibrated, the same workflow can be applied across different fields and seasons, generating consistent trait outputs that breeders can feed into selection decisions, genomic analyses, or yield prediction models.
The peanut workflow illustrates how foundation vision models can be adapted to agricultural phenotyping, where plot geometry, overlapping canopies, and heterogeneous backgrounds often complicate classical image processing. By treating segmentation as a general vision task rather than a crop specific rule set, the approach can potentially transfer to other species and breeding programs with limited retraining, especially when combined with crop specific post processing to compute relevant traits.
Together, the maize and peanut studies show that drone based phenomics is moving from index centered metrics toward integrated pipelines that span image capture, model based trait prediction, and statistical genetics. In maize, the focus is on using phenomic prediction to sharpen genetic signal and improve disease resistance mapping, while in peanut, the emphasis is on automating trait extraction to reduce labor and increase throughput in field nurseries.
Both efforts underscore the importance of handling large, complex datasets in a way that preserves biological information rather than compressing it into a few indices. As similar workflows are deployed more broadly, breeders are likely to gain faster, more precise measurements of disease responses and canopy architecture across environments, improving their ability to identify and deploy useful alleles in crops facing climate and production challenges.
Research Report:Comparing statistical 'phenomic prediction' models for remote-sensing-based phenotyping of maize susceptibility to common rust
Related Links
International Maize and Wheat Improvement Center CIMMYT
North Carolina State University
Farming Today - Suppliers and Technology
| Subscribe Free To Our Daily Newsletters |
| Subscribe Free To Our Daily Newsletters |
