Field-based high-throughput phenotyping technologies, such as drones, are able to provide phenome-wide measurements of plants in much the same way that high-throughput sequencers have provided genome-wide data. Uniquely, phenotyping technologies can screen high numbers of plots repeatedly through the growing period resulting in not only high spatial resolution but also high temporal resolution, helping dissect how different genotypes respond to their environments to maximize fitness in near real-time (SI appendix, Figs. S7 and S8).
As new temporal phenomic markers are difficult to independently measure and validate, one of the first approaches to evaluate phenomic marker utility is to look at heritability/repeatability values over different replicates and environments. This approach is not needed for genomic markers which do not vary over replicates and environments and theoretically have a repeatability near 1, but are also unable to capture environmental interaction in real time. Temporal repeatability (Eq. 3) of VIs were moderate, above ~0.5 for TPP_RGB (SI appendix, Fig. S2) and between 0.26 and 0.66 for TPP_Multi (SI appendix, Fig. S3). Temporal repeatability relied on variation across plant development, biologically more meaningful than using genotypic variation which is static at every time point. Temporal variation captured by drones assesses temporal genotypic variation jointly over time via nested design (Eq. 2). Previously, repeatability has only been calculated between different vegetation indices/CHM and yield at a single time point (16, 31–37); disregarding the temporal genotypic variation occurring across plant growth. Furthermore, previous studies used either one or a limited number time points and analyzed each time point separately.
High dimensional and temporal resolution phenomic data used in predictive plant breeding integrated with high throughput genotyping data discovered underlying genetic causes for many important temporal VI features. For instance, pleiotropy discovered via GWAS identified specific loci controlling more than one VIs (SI appendix, Figs. S13 and S14, Dataset S1). In addition, genomic prediction of temporal VI phenotypes proved that estimated effects of each marker varied through time, causing different prediction accuracy results for temporal phenotypes of the same VIs (Fig. 2). Therefore, instead of depending on discrete genome wide markers as predictors for yield, temporal phenotype data formed by estimated temporal marker effects could better predict certain scenarios (e.g., untested genotypes in tested environment). Predicting grain yield of untested genotypes in a tested environment is an important scenario for public breeding programs because lines developed in public breeding programs are mostly targeted for specific environments. So that figure 3 proved that TPP predicted the grain yield better than GP in CV2 indicating that TPP can be better solution for the public breeding programs for genetic gain. In addition, the predictive ability of TPP in untested genotype untested environments (CV4) was in the same range as that of GS (Fig. 3). This is also an important proof of concept that TPP can be used as widely as GP. Genomic prediction methods have been developed over more than a decade and phenomic prediction methods can likewise be improved. Further optimization and improvement of this approach will likely benefit from the integration of novel crop growth models as genomic prediction has (38).
Phenomic data can predict yield and flowering times via machine learning regressions
Shrinkage factors previously shown as the best performing prediction models when using different hyper parameters have been adapted for predicting both yield (15, 31, 39) and flowering times (15) when different reflection bands were used as predictors. Machine learning models with different regularization parameter settings to predict yield and flowering times (Fig. 1) were more accurate than linear-based prediction models (15, 34). This suggests that temporal variation in VIs do not have a linear relationship to predicted variables. This is because linear models tend to overfit when there are increasing numbers of predictors and with fluctuating collinearity between predictors, such as in phenomic data. Linear models are not capable to explain non-linear relationships between predictors and predicted variables.
Tuning regularization parameters of the ridge, lasso and elastic net-based prediction models is a good approach to deal with model overfitting when high dimensional phenomics data are used in prediction. Tuned regularization parameters in ridge, lasso and elastic net models can lessen coefficients, and predict test data more reliably than linear models. For example, pedigree within flight combination (\({\varOmega }_{i\left(j\right)}\) component in Eq. 2) were found to be statistically significant for all VI and CHM (SI appendix, Fig. S2) indicating a temporal interaction among the pedigree across flight times because of fluctuating temporal phenotype values of VIs (SI appendix, Figs. S7 and S8). Nevertheless, a general trend demonstrated that high- and low- yielding pedigrees segregate according to temporal phenotypes of VIs. This reverse correlation of temporal breeding values of the pedigree through time supports the existence of nonlinear relationships, problematic for a linear model to capture. Because of multiple decision tree learning, the random forest model accounts for non-linearity, limiting overfitting.
Phenomic prediction reached up to ~0.80 for grain yield and flowering time prediction (Fig. 1) higher than previously reported prediction accuracies (31–33, 35–37). (31) showed use of raw reflected bands instead of ratios (e.g. vegetation indices) performed better in prediction models. (34) further reported using all bands simultaneously increased prediction accuracy instead of VIs alone. However, reflected bands used in past studies derived from five to nine time points, lower time dimension data than what we generated in our study. This suggests that predictors derived from additional time points could play an important role on increasing the prediction ability of the models; more so than using the predictors as either raw reflectance bands or vegetation indices.
Genomic prediction for temporal traits can vary depending on the time points of growth
TPP_RGB phenomic data tested using genomic prediction to identify temporal marker effects and their prediction accuracies for each VI and Weibull_CHM throughout time (Fig. 2) demonstrated that genomic markers could predict an individual’s VI or Weibull_CHM value through cross validation using other individuals at the same stage. This demonstrated that certain stages and VIs have more genetic determination and are more heritable.
Temporally varying marker effects on the phenotype of VIs resulted in phenotypes at different timepoints of VIs and Weibull_CHM having different correlations with yield (SI appendix, Figs. S5 and S6) as well as different prediction abilities for dependent variables (Fig. 2). A dynamic pattern of marker effects as shown here has so far been overlooked in genomic prediction/selection of yield. (4) underlined that predicting the candidate genotype using the phenotype information collected from across multiple environments may be more accurate than using the genetic markers in the prediction model. Similarly, instead of predicting grain yield fitness by whole genome marker effect approaches such as RR-BLUP and GBLUP, including the temporal phenotypic variation occurring across growth into prediction models can result in more accurate fitness prediction as phenomic data already contain temporal marker effects. This study also showed that specific loci can explain different phenotypic variance across more than one derived VI (SI appendix, Figs. S13 and S14, Dataset S1) signifying pleiotropic effects of certain markers for the VIs. These pleiotropic effects have various associations with developing young tissues, inflorescence, and yield.
Phenomic prediction can perform similarly to or outperform genomic prediction
Phenomic data (TPP_Multi and TPP_RGB) predicted grain yield as well as genomic data using ridge regression (Fig. 3) but different results were observed depending on the cross validation scheme. TPP_RGB contained 35 VIs derived from only RGB bands and Weibull_ CHM belonging to fifteen time points (525 phenomic features) resulting in an accuracy of 0.71; this accuracy was same as the accuracy of 0.71 belonging to GP containing the 153,252 segregating whole genome markers. However, when TPP_Multi, which contains the 89 VIs derived from the multispectral bands and Weibull_CHM belonging to twelve time points (1068 phenomic features), were used in the prediction the yield, prediction accuracy reached up to 0.80; substantially higher than both GP and TPP_RGB supplied for the untested genotype in tested environments schemes (CV2) (Fig. 3). Moreover, in the most challenging cross-validation scheme, untested genotypes in untested environment (CV4), GP, TPP_RGB and TPP_Multi performed approximately equally as their prediction accuracies were around 0.50 ± 0.05 (Fig. 3). These empirical findings suggest, for the first time, that increasing temporal as well as spectral information can be used to predict fitness substantially better than genomic prediction. This also suggests that temporal and continuous phenomic data can be better predictors than discrete genomic data in prediction and selection of high yielding genotypes. In the only two previous phenomic prediction studies reported to date, (2) used 3,076 NIRS bands at a single timepoint, while (40) used 1,050 NIRS bands on grain samples. (40) then showed these NIRS bands outperformed genomic selection which used 84,259 SNP markers in wheat. Overall, phenomic selection is an emerging approach that may remove the cost of genotyping each year that is required by genomic prediction/selection. Adding a temporal component into phenomic prediction has innumerable known and yet to be discovered advantages.
In summary, this study demonstrated the predictive capability of phenomic data for complex traits in maize, yielding as much as genomic markers frequently applied in plant selection over the past 20 years. UAS surveys over the experimental field plots supplied temporal traits as predictors to facilitate the selection of untested genotypes in untested environments. Growing more plants and measuring them accurately are critical steps to drive effectiveness of selection intensity and accuracy resulting in higher genetic gain over time. This study exemplified that screening more plants and measuring them thanks to repetitive UAV flights across plant growth may results in greater genetic gain than genomic selection when phenomic prediction/selection is applied routinely.