Next Article in Journal
Mechanical Harvesting of Marginal Land and Agroforestry Field: New Insights from Safflower for Bio-Product Production
Previous Article in Journal
Influence of Wheat Cultivars, Infection Level, and Climate after Anthesis on Efficacy of Fungicide for Control of Fusarium Head Blight in the Huang-Huai-Hai Plain of China
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Agronomy
Volume 14
Issue 10
10.3390/agronomy14102267
agronomy-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Opportunities to Improve the Recommendation of Plant Varieties under the Recommended List (RL) System

by
Chin Jian Yang
1,*,
Joanne Russell
2,
Ian Mackay
1 and
Wayne Powell
1,*
1
Scotland’s Rural College (SRUC), Peter Wilson Building, King’s Buildings, West Mains Road, Edinburgh EH9 3JG, UK
2
The James Hutton Institute, Invergowrie, Dundee DD2 5DA, UK
*
Authors to whom correspondence should be addressed.
Agronomy 2024, 14(10), 2267; https://doi.org/10.3390/agronomy14102267
Submission received: 26 August 2024 / Revised: 12 September 2024 / Accepted: 26 September 2024 / Published: 1 October 2024
(This article belongs to the Section Crop Breeding and Genetics)
Browse Figures
Review Reports Versions Notes

Abstract

:
Recommended List (RL) is the UK plant variety recommendation system introduced in 1944 for supporting growers in making decisions on variety choices. The current RL system is heavily focused on single-trial analyses developed in the 1980s without making full use of information across varieties and trial sites. Given the statistical advances that have been developed and adopted elsewhere, it is timely to review and update the methods for data analysis in RL. In addition, threats from climate change challenge the prediction of variety performance in future environments. Better variety recommendations, particularly for matching varieties to specific environments can be achieved through the improved modeling of effects from genetics, environments, and genetic-by-environment interactions. Here, we evaluate grain yield data from 153 spring barley varieties that were trialed for RL from 2002 to 2019. Our results show that the current RL system produces poor and inconsistent predictions on variety performance across environments. Improvement in RL can be achieved by using mixed models that account for genetic relationships among varieties, and additional improvement is possible if genetic-by-environment interaction can be modeled accurately. We highlight the relevance and importance of genomics in both variety registration and recommendation.

1. Introduction

Plant breeding, in combination with improved agronomy practices, contributes greatly toward advancing modern crop agriculture and safeguarding food security [1]. Breeding programs are designed to produce new plant varieties with higher yield, qualities, disease resistance, and abiotic stress tolerance [2]. Much of the genetic gain in these traits can be attributed to either transgressive segregation in line breeding or heterosis with dispersed dominance in hybrid breeding [3]. Because many target traits are polygenic and have complex genetic architecture, new varieties are produced through the shuffling of favorable alleles or haplotypes within breeding pools [4]. For example, the Illinois Long-Term Selection Experiment for oil and protein content in maize grain has shown continuous genetic gains over more than a hundred years of selection [5]. Recent developments in precision technologies such as gene editing, remote sensing, and artificial intelligence [6,7,8] could potentially revolutionize future breeding systems to achieve resilient, sustainable, and regenerative agriculture.
As new varieties are released annually by various public and private breeding programs, plant variety recommendation systems help growers choose ideal varieties that fit their market goals and farm environments [9]. The variety recommendation systems typically depend on analyses of trials that span multiple sites and years to characterize agronomic trait stability and rank across environments. In addition to new varieties, several important and high-performing older varieties are often included as checks or controls. Variety recommendation is usually limited to major crops with high economic importance due to the cost of large-scale trials. These systems vary from country to country, and it is known as Recommended List (RL) in the United Kingdom, National Variety Trials (NVTs) in Australia, and Official Variety Testing (OVT) in the United States. RL is conducted by the Agriculture and Horticulture Development Board (AHDB) with the support from the Biomathematics and Statistics Scotland (BioSS), NVTs are conducted by the Grain Research and Development Corporation (GRDC), and OVT is often conducted at the state level by land-grant public universities. Each year, these testing authorities are responsible for analyzing trial data and publishing variety recommendations in various forms such as printed or online tables, variety selection tools, or mobile applications.
Analyzing variety performance in multienvironment trials can be challenging due to the presence of genetic-by-environment interaction (GEI) effects observed in most traits [10]. There are usually multiple replicates in each trial site and at least 10 to 20 trial sites per year depending on the geographical areas covered by the variety testing authorities. The appropriate modeling of genetic, environment, and GEI effects is crucial to ensuring good predictions on variety performance. Variety trial design and analysis are century-old topics and trace back to early work such as Fisher [11], Yates [12,13,14], and Yates and Cochran [15]. The first edition of RL was published by the National Institute of Agricultural Botany (NIAB) in 1944 with only 15 winter wheat varieties [16]. Interestingly, variety recommendation systems in many countries build on the approach developed by Patterson and Silvey [10]. The UK was among the early adopters of this approach and has not changed its RL system since then. On the other hand, the NVTs system in Australia has been modified over the past few decades [17,18,19]. In the US, the OVT systems vary from state to state and are rarely discussed in academic literature. In general, variety performance is usually obtained from a two-stage model where variety means are first computed for within environment (site–year) and followed by across environment. In some cases, a one-stage model is used where raw data from all trialed environments and replicates are fitted together. If appropriate variance components from the first stage are accounted in the second stage, the two-stage model can be equivalent to the one-stage model [20], although this is rarely true in practice due to variability in estimates.
Here, we evaluate the current RL system in the UK and identify possible areas for improvement. Performances of control and new varieties are often presented in a table of trait means within environment, where an environment is defined as site-year combination. This table is presumably derived from the approach of Patterson and Silvey [10], but the information is not publicly available. In the AHDB variety selection tool (https://ahdb.org.uk/variety-selection-spring-barley, accessed on 10 May 2024), variety means can be aggregated from these tables for individual, regional, or all sites from one to five years of trials using simple averages regardless of missing environment. In the context of two-stage multienvironment analysis, the outputs from first and second stages are represented by the tables and aggregated variety means, respectively. The simple means approach in the second stage can be described as a fixed effect model, which can be improved by treating variety means as random instead of fixed effect [21]. Our work here focuses on investigating possible improvements to prediction of variety performance in the RL system based on this approach. For our investigation, we choose spring barley as an example due to its importance in UK agriculture and similarity to the RL system in other major crops such as winter barley, spring/winter wheat, spring/winter oats, and winter oilseed rape. Spring barley is the second most economically important crop in the UK with a production area and value of 1104 thousand hectares and GBP 1815 million, respectively [22]. We focus on grain yield because it has the most complete data in the public domain. We set the trial data in the immediate subsequent year as the prediction target. Because the raw data within each trial site are not publicly available, we cannot evaluate one-stage models nor properly fit genetic-by-environment interactions in our models. First, we evaluate the performance of current system across trial sites and years. Second, we test if the inclusion of genomic information and use of mixed models in the second stage can help improve the system. Third, we estimate the annual yield loss based on the difference in performance between best predicted variety and best observed variety. Finally, we discuss opportunities for further improvement to take advantage of the latest developments in genetics, statistics and artificial intelligence.
Briefly, our results highlight key findings in several areas. First, the current RL system provides poor and inconsistent variety recommendations across environments. Second, while simple means from more trial sites lead to higher prediction accuracies on variety performance, one-year means outperform five-year means due to inaccurate variety ranks involving new and old varieties. Third, alternative methods using Genomic Best Linear Unbiased Prediction (GBLUP) address the shortcomings in the existing variety selection methods by producing better and more consistent variety performance predictions. Fourth, among the eight different methods evaluated here, GBLUP with five years of variety trial data outperforms all considered metrics. Fifth, the use of genomic marker data for variety recommendation fits perfectly with a previously proposed system for variety registration, known as genomic Distinctness, Uniformity, and Stability (DUS) [23].

2. Materials and Methods

2.1. Data Collection

Treated yield data for spring barley were downloaded from the AHDB website (https://ahdb.org.uk/knowledge-library/recommended-lists-archive, accessed on 10 May 2024). The data included 153 unique varieties, 18 years (2002 to 2019), and 80 unique trial sites (Tables S1 and S2). The trial sites were further grouped based on “counties”, which were roughly defined based on 48 ceremonial counties in England, 32 council areas in Scotland, 22 principal areas in Wales, and 6 counties in Northern Ireland (Figure S1, Table S1). Genomic marker data for all varieties were obtained from the Improving Winter Barley Malting Quality (IMPROMALT) panel [24] which was genotyped using the barley 50 k array [25]. The original genomic marker data had 43,799 Single-Nucleotide Polymorphism (SNP) markers and 809 varieties that span a much longer period of time. After filtering for the 153 varieties with yield data, we removed any monomorphic markers and retained 24,380 markers for analysis. Even though yield data were available for newer varieties from 2020 onward, their genomic marker data were unavailable and thus were not included in our analysis.
Historic data for 24 out of 37 weather stations were downloaded from the UK Met Office (https://www.metoffice.gov.uk/research/climate/maps-and-data/historic-station-data, accessed on 10 May 2024). These 24 weather stations were chosen based on their proximity to the trial sites. The data included the monthly average of maximum temperature (Tmax, °C), minimum temperature (Tmin, °C), number of frost days, and rainfall (mm) from 2002 to 2019. Temperature range (Trange, °C) was estimated as the difference between the maximum and minimum temperatures. The weather data were only used to highlight climate variation within the UK.
The 2011 InFuse UK boundary data for local authorities were downloaded from the census website (https://borders.ukdataservice.ac.uk, accessed on 10 May 2024) for plotting the map. The local authorities in the boundary data were merged using the R/sf package [26] to form the counties. After merging, an approximated adjustment was applied to correct for part of county Armagh that was misclassified as county Down in Northern Ireland.

2.2. Yield Prediction

In the current UK variety selection tool (https://ahdb.org.uk/variety-selection-spring-barley, accessed on 10 May 2024), there are six available methods for calculating variety performance to help growers with deciding on variety choice. These methods for making yield predictions in each target site were evaluated here and are henceforth referred to by their short names. I1: yield from the same trial site in the previous year. I5: yield mean from the same trial site in the previous five years. R1: yield mean from all trial sites in the same region in the previous year. R5: yield mean from all trial sites in the same region in the previous five years. A1: yield mean from all trial sites in the previous year. A5: yield mean from all trial sites in the previous five years. Details of these methods are provided in Figure S2.
Two alternative methods for yield prediction, I5B and A5B, were evaluated by extending I5 and A5 with the GBLUP approach, hence the names I5B and A5B. Instead of assuming that each variety was distinct, these alternative methods accounted for variety similarities by modeling trait variations according to their genomic relationships. I5B also accounted for random genetic-by-year interaction as GEI, but A5B did not. This is because of model convergence issues when attempting to fit the genetic-by-year interaction in A5B. GBLUP is commonly used in breeding programs with genomic selection or prediction, and its predictive ability tends to scale with the number of individuals. In the case of a variety trial, the number of varieties tends to be low and may not draw the full advantage of GBLUP. Due to the lack of replicated data within each trial site, we were unable to evaluate standard Best Linear Unbiased Prediction (BLUP) models, which have been shown to be effective in improving prediction for variety recommendation [21]. Therefore, the focus of our work here is on the second stage of a two-stage analysis rather than a one-stage analysis. The models for I5B and A5B were fitted using the R/sommer package [27] based on the following general model equation.
y = X β + Z 1 g 1 + Z 2 g 2 + ε
The model terms are described below:
  • y is a vector of yield in n varieties.
  • X is a design matrix for the fixed effects such that the first column is a vector of 1’s for the overall mean, the second to fifth columns are vectors of 0’s and 1’s, indicating which year the observation belongs to, and the remaining columns are vectors of 0’s and 1’s, indicating which trial site the observation belongs to.
  • β is a vector of fixed effects including the overall mean effect, year effects, and site effects.
  • Z 1 is an incidence matrix relating n varieties to observations y.
  • g 1 is a vector of random genetic effect with a normal distribution of N ( 0 , σ g 1 2 K ) .
  • Z 2 is an incidence matrix relating n varieties by p years to observations y.
  • g 2 is a vector of random genetic-by-year effect with a normal distribution of N ( 0 , σ g 2 2 P ⊗ K ) .
  • K is an additive genetic relationship matrix with elements K i j = ∑ k m ( x i k − p k ) ( x j k − p k ) ∑ k m 2 p k ( 1 − p k ) ; x i k is the marker score for variety i at marker k, x j k is the marker score for variety j at marker k, p k is the allele frequency at marker k, and m is the total number of markers.
  • P is an identity matrix of p × p .
  • ε is a vector of residual effect with a normal distribution of N ( 0 , σ ε 2 I ) , and I is an identity matrix.
  • Specifically, the trial site effect was excluded in I5B and the genetic-by-year effect was excluded in A5B.

2.3. Comparing Predicted and Observed Yield

In practice, growers can choose a variety based on the performances of available varieties in previous years. Because we did not have the commercial production data for every variety in each site and year, we used the trial data as a proxy for the prediction target. For example, yield prediction in 2007 was achieved using data from 2002 to 2006, and the predicted yield was compared with the observed trial yield in 2007. Additionally, our criteria for analysis required a site to have trial data in the target year, the year immediately before, and a minimum of three out of five previous years (Figure S2).
The predicted yield from all eight methods was compared with the observed yield in each trial site and year using Spearman’s rank correlations. Given the importance of variety ranks in a variety recommendation system, ranked-based correlations were used instead of standard Pearson correlations. To ensure a fair comparison, the results were restricted to trial sites with available predicted yield from all eight methods. Because of our interest in the methods for identifying variety performance, the correlations were grouped according to method, method by year, or method by site. For each grouping approach, differences in means were tested using t-tests.
The consequences of suboptimal variety choices due to imperfect predictions were quantified as Percent Deficit (PD) in yield for each method, site, and year. To calculate PD, we first obtained the ratio of the difference between the highest and chosen variety’s yield over the difference between the highest and lowest yield. The ratio was then converted into a percentage. For example, if the chosen variety’s highest and lowest yields are 8, 10, and 5, respectively, then the PD would be 100 × 10 − 8 10 − 5 = 40 . All of the values used in the calculations were the observed yield in the target site and year.

3. Results

3.1. Trial Environment

Between 2002 and 2019, 39 out of 108 counties in the UK had at least one spring barley variety trial for RL (Figure 1A and Figure S1, Table S1). Counties in the UK differ among the devolved nations and the definitions for counties used here are described in the Materials and Methods. The number of sites per county ranges from 1 to 5, with Perth and Kinross having the most sites (Aberuthven, Coupar Angus, Invergowrie, Kinross, Perth). Across the 39 represented counties over 18 years, the number of years with trials ranges from 1 (City of Edinburgh, Lancashire, Staffordshire) to 18 (Aberdeenshire, Fife, Norfolk). Similar to the wheat variety trial [28], trial sites for spring barley varieties may not always overlap from one year to another. Overall, spring barley variety trials can be found more often on the east coast than in the center or west coast of the UK (Figure 1A).
The variety trials are spread across the key areas where spring barley is commonly produced, and these trial sites cover a broad range of climate variations. Historic data (2002 to 2019) from 24 relevant UK Met Office weather stations capture information on temperature, frost days, and rainfall (Figure 1B, Table 1). Because the spring barley growing season typically ranges from April (or March if sown early) to August, we restrict the weather data from March to August. As expected, the maximum and minimum temperatures increase as latitude decreases. The switch from colder to warmer climates can be roughly placed between station 10 and 11, which is approximately close to a region known as the Midlands. A smaller temperature range is observed in coastal stations (example with lowest range) over inland stations (example with highest range). The risk of frost during the spring barley season is fairly low, but there are several north and inland stations with higher numbers of frost days such as 2 and 5. Rainfall is evenly distributed across the stations, except for three stations in northwestern UK (4, 5, 6) with higher rainfall.

3.2. Variety Selection Tool

In the variety selection tool provided by the AHDB (https://ahdb.org.uk/variety-selection-spring-barley, accessed on 10 May 2024), users can select one or multiple traits and calculate the performance of all available varieties using six different methods. These methods generate variety performance based on simple means of the chosen trait(s), and the details are provided in the Section 2. Briefly, the trait means can be calculated for each trial site (I), region (R), or all sites (A) from 1 or 5 years of data. For a fair comparison of variety selection methods, only trial sites with data in the target year, the year before, and a minimum of three out of five previous years are considered (Figure S2). These criteria resulted in 29 out of 80 trial sites analyzed (Table 2). AHDB classifies the regions for cereal into North, West, and East, but these regions do not actually reflect the observed climate variations (Figure 1B, Table S1). For any chosen method, users are given ranks of the spring barley varieties from the best to worst. For simplicity, we are focusing on grain yield in our analysis because that is the most important agronomic trait and has the most publicly available data.
Predictions of variety performance are highly variable and vary in success across the years (Figure 1C). Using Coaltown of Balgonie, one of the most trialed sites, as an example, the predicted variety ranks are well correlated with the observed ranks in 2007 but not 2016. In 2007, the best prediction was achieved in method I5, and the best-observed variety matches with the best-predicted variety in methods I5, R1, and A1. When considering all varieties, five-year means generally perform better than one-year means. However, in 2016, none of the methods had a clear advantage over the other, as they all performed badly. The best-predicted varieties ended up with poor yields across all methods.

3.3. Comparison across Methods in Predicting Variety Yield

The best prediction of variety performance across all trial sites and years can be achieved using A5B with a median correlation between predicted and observed variety ranks of 0.571 (Figure 2). The statistical model for A5B treats variety effect as random, accounts for the genetic relationships among varieties, and calculates the variety effect as GBLUP. The improvement in prediction with this approach is due to shrinkage toward means and maximization of correlations between true and predicted genotypic values [21]. In addition, the A5B method takes advantage of the tendency of similar varieties to produce similar phenotypes. The next best prediction comes from both I5B and A1 with a median correlation of 0.518. While I5B uses a similar model to A5B, it uses only data from single trial sites. This discrepancy highlights the benefits of variety performance prediction from multiple over single trial sites. The usefulness of A5B in improving the RL system is exemplified here as its median correlation is 0.053 higher when compared with the best existing method which is A1. Curiously, I5 has the worst prediction with a median correlation of 0.324. Despite different methods showing different predictive abilities, the use of GBLUP offers a clear advantage over other methods.
Methods that use data from more trial sites show higher accuracies in the prediction of variety performance. This pattern follows the order of yield means from individual sites (I), regional sites (R), and all sites (A) (Figure 2). The number of trial sites ranges from 2 to 11 per year for R and 13 to 25 per year for A (Table S1). The median correlations between predicted and observed variety ranks for I1/I5, R1/R5, and A1/A5 are 0.377, 0.474, and 0.488, respectively. The correlations for R-based methods are closer to the A-based methods instead of I-based methods, which could be due to diminishing benefits from additional trial sites and limitations of methods that derive variety performance through simple means.
In contrast to the GBLUP methods, the other methods work better with one year of yield data instead of five years. This observation is consistent across all three comparisons for the pairs of I1-I5, R1-R5, and A1-A5 (Figure 2). The general expectation is that better prediction of variety performance can be attained using more years, as shown when using more trial sites. There are two possible explanations for this contradiction. First, year-to-year yield variation can be large and data from years further apart are less informative toward predicting the yield in a target year. Second, given the nature of the variety recommendation system where new and old varieties constantly enter and exit each year, some varieties have fewer data than others in the predictions with five years of data and lead to an unbalanced estimate of variety means. These issues are partially overcome through the use of GBLUP methods and likely explain the poor prediction of variety performance through simple means. It is evident that the information from five years of data is better used in GBLUP methods than existing non-GBLUP methods.

3.4. Site-Specific Comparison of Methods

Results for the six most used trial sites suggest a slight advantage for GBLUP methods (Figure 3). Four of these sites (Coaltown of Balgonie, Laurencekirk, St Boswells, and Tain) are located in Scotland, and two (Stockbridge, Wymondham) are located in England. Full results for all 29 analyzed trial sites are provided in Figure S3. The only comparison that is significant after applying a Bonferroni correction for multiple testing is between method I1 and A5B (p = 0.0011), where A5B produces significantly higher correlations between predicted and observed variety ranks (Figure 3B). Despite the lack of significance in other pairwise comparisons, the GBLUP methods tended to perform better than their counterparts (e.g., I5B vs. I5 and A5B vs. A5) in all six trial sites.
A previous observation that non-GBLUP methods work better with one year of yield data instead of five years (Figure 2) can again be seen in Laurencekirk, Stockbridge, and Tain (Figure 3). As an extreme example, methods I5, R5, and A5 produce negative correlations when used to predict the 2013 yield at Laurencekirk and Tain with data from 2008 to 2012 (Figure 3A). Similar examples can also be seen at Coaltown of Balgonie in 2013, Stockbridge in 2018 and 2019, and Wymondham in 2019. In many cases, GBLUP methods are able to recover the correlations to similar levels achieved when using one year of data. In contrast to several occasions where non-GBLUP methods produce negative correlations, GBLUP methods, especially A5B, offer better consistencies and lower risk of negative correlations.
Methods aside, varying success in yield prediction can be found across trial sites. Among the six trial sites, Laurencekirk appears to be the most consistent and best predicted over the years (Figure 3A), with a mean correlation of 0.708 across all methods and years. Tain places second with a mean correlation of 0.623. The remaining trial sites have mean correlations of 0.541 at St Boswells, 0.526 at Stockbridge, 0.496 at Wymondham, and 0.434 at Coaltown of Balgonie. Correlations at Coaltown of Balgonie and Stockbridge drop in recent years, while correlations at St Boswells and Wymondham are consistently low in most years.

3.5. Year-Specific Comparison of Methods

Good yield prediction can be achieved using GBLUP methods in all analyzed years from 2007 to 2019 (Figure 4). A5B slightly outperforms I5B, but they are highly comparable. Each year, A5B and I5B either perform better than non-GBLUP methods or perform similarly to the best non-GBLUP method. Correlations for yield prediction with I5B are significantly higher than I5 (p = 0.0004) and A5 (p = 0.0002) in 2013, as well as I5 (p = 0.0002) and A5 (p = 0.0001) in 2019. Correlations for yield prediction with A5B are significantly higher than A5 (p = 0.0007) in 2013, A5 (p = 0.0007) in 2018, as well as I5 (p = 0.0003), R5 (p = 0.0008) and A5 (p < 0.0001) in 2019. These comparisons are significant after applying Bonferroni correction for multiple testing (Figure 4A). As found in the previous analysis on the six most trialed sites, GBLUP methods are less likely to produce negative correlations on yield prediction.
In addition, GBLUP methods offer higher consistency than other methods over the years (Figure 4B). Correlations from I5B and A5B constantly hover close to 0.5 with smaller spikes and drop-offs. Non-GBLUP methods, especially I5, R5, and A5, suffer from severe correlation drops in a few selected years. This observation again highlights the challenges of using five years of variety trial data due to bias against new varieties with less trial data and the failure of existing methods in the variety selection tool to account for this bias. Among the three non-GBLUP methods that use five years of trial data, I5 clearly underperforms in most years. This result highlights the strong environment and GEI effects on yield in different years.
Methods aside, varying success in yield prediction can again be found across the years. The best year is 2015, with a mean correlation of 0.661 across all methods and trial sites. This is also the year where all methods produce similar correlations (Figure 4). Other good years with mean correlations above 0.500 are 2007, 2009, 2010, and 2011. In contrast, 2013 is the worst year for yield prediction, with a mean correlation of 0.239. Poor yield prediction with mean correlations below 0.400 can be seen in 2008, 2016, 2017, 2018, and 2019. In most of the years where the yield prediction is poor, much higher correlations can be attained using GBLUP methods.

3.6. Quantifying Deficits in Yield

Because of the imperfect yield prediction, all methods result in some amount of deficit in yield across trial sites and years (Figure 5 and Figure 6). Percent Deficits (PD) in yield are calculated as the percentage of the difference between the best-observed yield and the chosen variety’s yield over the difference between the best- and worst-observed yields. The chosen variety is taken from the best-predicted variety using the eight methods. PD at trial sites can be hard to visualize on a map and are therefore merged into counties by averaging PD from all sites within the same county each year (Figure 5). Most counties exhibit moderate PD with small variation among the methods. Extreme PD in a few counties such as Northumberland (Bowsden site), South Ayrshire (Ayr site), and Cambridgeshire (Fulbourn site) needs to be treated with caution as they only have one to two years of predicted data for analysis. Overall, we observe a mean PD of 29 across all methods, trial sites, and years.
While it can be hard to discern the PD from different methods at the geographical level, we find consistent differences among the methods across the years. A1 and A5B produce the lowest Cumulative PD (CPD), followed by comparable CPD in A5 and R1 (Figure 6). On the other hand, high CPD in I1, I5, and R5 suggests that these methods are not suitable for predicting variety performance. Despite I5B and A5B having comparable correlations, I5B is mediocre in terms of CPD, as it is largely hovering between the good and poor methods. Interestingly, there is a discrepancy of about 100 CPD between the best and worst methods over 13 years. As a caveat, we have to assume that all sites contribute equally to the UK annual spring barley production in our estimation of CPD. While this assumption is unlikely to be an exact match to the reality for any given year, it should still produce a reasonable estimate over a period of time as the trial sites are designed to cover spring barley production areas in the UK.

4. Discussion

The UK variety recommendation system builds on the work of Patterson and Silvey [10] and has remained largely unchanged since then. While the trial design and analysis for RL are highly comprehensive at the level of individual trial sites (https://ahdb.org.uk/ahdb-recommended-lists-for-cereals-and-oilseeds-2021-2026, accessed on 10 May 2024), the variety performance has not been well evaluated from a multienvironment perspective. Again, in the context of the two-stage multienvironment analysis, the model performs well in the first stage but not the second stage. The current second-stage model largely focuses on estimating means at each trial site as fixed effects without accounting for phenotypic variations due to genetic relatedness among varieties, environmental effects, and genetic-by-environment interaction (GEI) across trial sites and years. Many examples highlight the importance of GEI in phenotypic variations [29,30,31]. Consequently, we have shown that none of the existing methods for variety selection can reliably and consistently predict the agronomic performance in target environments (Figure 3 and Figure 4).
Our results suggest that the variety recommendation system can be easily improved using the Genomic Best Linear Unbiased Prediction (GBLUP) approach, which can be implemented with existing trial designs. Aside from not needing any modification to the current trial designs, this approach only requires all varieties to be genotyped using a standard platform and a switch from estimating simple means to fitting mixed models in the second-stage analysis. GBLUP is widely used in the selection of lines and crosses in many major plant and animal breeding programs [32,33]. Here, we have shown that methods I5B and A5B are superior to their counterparts, and A5B provides the best prediction of all methods considered (Figure 2). Aside from the overall benefits, we have also demonstrated that the GBLUP approach produces slight improvement across trial sites (Figure 3), reliable and consistent results across years (Figure 4), and minimal yield deficits (Figure 6). For now, we recommend the A5B method for variety performance prediction, as it can be easily implemented in the RL system without requiring significant changes to the existing pipeline for first-stage analysis. In the longer term, further improvement is likely possible by evaluating a single-stage model with complete trial data down to individual replicates in each environment.
Improvement to the variety recommendation system can be further increased by including unreleased or candidate varieties from the National List (NL) and breeders’ lines. Additional data from more individuals and environments allow the GBLUP model to better partition genetic, environmental, and GEI effects [34,35]. Before going into the Recommended List (RL) trials, a candidate variety has to be granted plant variety rights through evaluations for Distinctness, Uniformity, and Stability (DUS) and Value for Cultivation or Use (VCU) in NL trials. Therefore, genotyping candidate varieties may serve dual purposes in applying genomic DUS in variety registration [23,36] and GBLUP in variety recommendation. Because data confidentiality can be an issue with unreleased varieties or breeders’ lines, a recently developed homomorphic encryption method provides a straightforward way to maintain privacy in data sharing [37]. Furthermore, this encryption method does not require further validation, as it has been shown to work well with GBLUP [38].
Over the years, multiple studies on the RL system have been undertaken to determine the best variety recommendation for each crop. Patterson and Silvey [10] described the first stage model for capturing the variations within each trial site which has been used in the current RL system to produce variety performance at the level of individual trial sites. Silvey [39] investigated the entry and exit of wheat and barley varieties in the RL and showed that any recommended variety can last for an average of four years unless for exceptional uses. McVittie et al. [40] studied the method of evaluating milling quality, an important wheat trait, in the RL. Macaulay et al. [41] showed the genetic gain in key barley traits such as yield and hot water extract and traced the pedigree footprint of barley varieties in the RL. Mackay et al. [42] used the historical variety data from RL to estimate genetic gain in wheat, barley, oilseed rape, sugar beet, and forage maize. Marshall [43] investigated the impacts of sowing time and crop rotation on managing weed competition in wheat RL trials. Oberfoster and Marshall [44] compared phenotyping measurements and scales used by different European countries in their RL trials. Dias et al. [45] proposed the use of Bayesian probability concepts for stability in variety recommendations across a wide array of environments. Raymond et al. [28] identified bias in model prediction of variety performance in recommended UK wheat and barley due to the breakdown of the disease resistance allele over time and highlighted the importance of including checks in the trials and models to control for the bias. The outcomes of these studies and ours are complementary and can be used to guide and improve the RL system.
In addition to our GBLUP approach, there are also other methods that may improve the variety recommendation system. For example, treating variety as a random instead of fixed effect with appropriately fitted GEI terms has been shown to improve variety performance prediction [21,46,47]. This method can also be described as switching of the statistical model from Best Linear Unbiased Estimation (BLUE) to Best Linear Unbiased Prediction (BLUP) for variety effect. Genomic marker data are not always necessary for this method, but it does require either fitting data for all replicates within each trial site in the one-stage analysis or carrying over residual variance–covariance matrices in the two-stage analysis [21]. Unfortunately, neither of this information is publicly available so this approach could not be evaluated here. BLUP aside, there is a method known as Additive Main Effects and Multiplicative Interaction (AMMI) which accounts for GEI by fitting only fixed effects in its model [48,49]. Similar to our observations on the challenges in using fixed over random effects, AMMI has been shown to be less accurate than BLUP [50]. The Factor Analytic Linear Mixed Model (FA-LMM) is an improved method based on AMMI, which attempts to identify similar environments and reduce model complexity by fitting environmental principal components as random effects [18,51,52]. Lastly, given recent developments in Machine Learning (ML) and artificial intelligence (AI), there have been many ML/AI-based methods described for applications in variety recommendation systems [53,54,55,56,57].

5. Conclusions

Our analysis of the RL reveals that the existing methods produce a poor and inconsistent prediction of variety performance across environments and fail to make full use of multiyear trial data. Alternative methods using GBLUP can overcome these issues, and the best prediction of variety performance is achieved using GBLUP with five years of variety trial data. By having genomic marker information of candidate varieties, we can both improve variety registration and recommendation through DUS and RL, respectively. RL has come a long way since its inception in 1944 and is now celebrating its 80th anniversary in 2024. It is timely to revamp RL by revising the statistical methods for variety recommendation, which will benefit breeders and growers, contribute to sustainable agriculture, and tackle various threats from climate change.

Supplementary Materials

The following supporting information can be downloaded at https://www.mdpi.com/article/10.3390/agronomy14102267/s1, Table S1: Spring barley trial sites from 2002 to 2019; Table S2: Spring barley varieties from 2002 to 2019; Figure S1: Distribution of counties with spring barley trials; Figure S2: Methods and criteria for predicting variety performance; Figure S3: Correlations between predicted and observed yield ranks in all 29 analyzed trial sites.

Author Contributions

Conceptualization, I.M. and W.P.; methodology, C.J.Y. and I.M.; formal analysis, C.J.Y.; resources, J.R.; data curation, C.J.Y. and J.R., writing–original draft preparation, C.J.Y. and W.P., writing–review and editing, J.R. and W.P.; funding acquisition, I.M. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the Scottish Society for Crop Research (SSCR) to I.M.

Data Availability Statement

R scripts, variety data (variety name, site, year, yield), and historical weather station data are available at https://github.com/cjyang-work/RL, accessed on 26 August 2024. Genomic marker data are available upon request from the authors in Looseley et al. [24]. All online links have been archived at http://web.archive.org/ on 10 May 2024.

Acknowledgments

We are grateful to the IMPROMALT consortium for sharing the spring barley genomic marker data. We appreciate valuable discussions with the past and present members of the Principal’s Research Group, including Rajiv Sharma, Ian Dawson, and David Marshall. We thank all reviewers for their time and comments.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Bradshaw, J.E. Plant breeding: Past, present and future. Euphytica 2017, 213, 60. [Google Scholar] [CrossRef]
  2. Cobb, J.N.; Juma, R.U.; Biswas, P.S.; Arbelaez, J.D.; Rutkoski, J.; Atlin, G.; Hagen, T.; Quinn, M.; Ng, E.H. Enhancing the rate of genetic gain in public-sector plant breeding programs: Lessons from the breeder’s equation. Theor. Appl. Genet. 2019, 132, 627–645. [Google Scholar] [CrossRef]
  3. Mackay, I.J.; Cockram, J.; Howell, P.; Powell, W. Understanding the classics: The unifying concepts of transgressive segregation, inbreeding depression and heterosis and their central relevance for crop breeding. Plant Biotechnol. 2021, 19, 26–34. [Google Scholar] [CrossRef] [PubMed]
  4. Scott, M.F.; Fradgley, N.; Bentley, A.R.; Brabbs, T.; Corke, F.; Gardener, K.A.; Horsnell, R.; Howell, P.; Ladejobi, O.; Mackay, I.J.; et al. Limited haplotype diversity underlies polygenic trait architecture across 70 years of wheat breeding. Genom. Biol. 2021, 22, 137. [Google Scholar] [CrossRef] [PubMed]
  5. Moose, S.P.; Dudley, J.W.; Rocheford, T.R. Maize selection passes the century mark: A unique resource for 21st century genomics. Trends Plant Sci. 2004, 9, 358–364. [Google Scholar] [CrossRef]
  6. Wallace, J.G.; Rodgers-Melnick, E.; Buckler, E.S. On the road to breeding 4.0: Unraveling the good, the bad and the boring of crop quantitative genomics. Annu. Rev. Genet. 2018, 52, 421–444. [Google Scholar] [CrossRef]
  7. Zaidi, S.S.A.; Vanderschuren, H.; Qaim, M.; Mahfouz, M.M.; Kohli, A.; Mansoor, S.; Tester, M. New plant breeding technologies for food security. Science 2019, 363, 1390–1391. [Google Scholar] [CrossRef]
  8. Xu, Y.; Zhang, X.; Li, H.; Zheng, H.; Zhang, J.; Olsen, M.S.; Varshney, R.K.; Prasanna, B.M.; Qian, Q. Smart breeding driven by big data, artificial intelligence, and integrated genomic-enviromic prediction. Mol. Plant 2022, 15, 1664–1695. [Google Scholar] [CrossRef]
  9. Federizzi, L.C.; Carbonell, S.A.M.; Pacheco, M.T.; Nava, I.C. Breeders’ work after cultivar development-the stage of recommendation. Crop Breed Appl. Biotechnol. 2012, S2, 67–74. [Google Scholar] [CrossRef]
  10. Patterson, H.D.; Silvey, V. Statutory and recommended list trials of crop varieties in the United Kingdom. J. Royal Stat. Soc. 1980, 143, 219–252. [Google Scholar] [CrossRef]
  11. Fisher, R.A. The Design of Experiments; Oliver and Boyd: Edinburgh, UK, 1935. [Google Scholar]
  12. Yates, F. The principles of orthogonality and confounding in replicated experiments. J. Agric Sci. 1933, 23, 108–145. [Google Scholar] [CrossRef]
  13. Yates, F. A new method of arranging variety trials involving a large number of varieties. J. Agric Sci. 1936, 26, 424–455. [Google Scholar] [CrossRef]
  14. Yates, F. Sampling Methods for Censuses and Surveys; Griffin: London, UK, 1960. [Google Scholar]
  15. Yates, F.; Cochran, W.G. The analysis of groups of experiments. J. Agric Sci. 1938, 28, 556–580. [Google Scholar] [CrossRef]
  16. Recommended Lists for Cereals and Oilseeds 2024/25. Available online: https://projectblue.blob.core.windows.net/media/Default/Imported%20Publication%20Docs/AHDB%20Cereals%20&%20Oilseeds/Varieties/RL2024-25/Recommended%20Lists%20for%20cereals%20and%20oilseeds%202024-25.pdf (accessed on 10 May 2024).
  17. Smith, A.B.; Cullis, B.R.; Thompson, R. Analyzing variety by environment data using multiplicative mixed models and adjustments for spatial field trend. Biometrics 2001, 57, 1138–1147. [Google Scholar] [CrossRef]
  18. Smith, A.; Ganesalingam, A.; Kuchel, H.; Cullis, B. Factor analytic mixed models for the provision of grower information from national crop variety testing programmes. Theor. Appl. Genet. 2015, 128, 55–72. [Google Scholar] [CrossRef]
  19. Gogel, B.; Smith, A.; Cullis, B. Comparison of a one- and two-stage mixed model analysis of Australia’s National Variety Trial Southern Region wheat data. Euphytica 2018, 214, 44. [Google Scholar] [CrossRef]
  20. Piepho, H.-P.; Mohring, J.; Schulz-Streeck, T.; Ogutu, J.O. A stage-wise approach for the analysis of multi-environment trials. Biom. J. 2012, 54, 844–860. [Google Scholar] [CrossRef]
  21. Piepho, H.-P.; Mohring, J.; Melchinger, A.E.; Buchse, A. BLUP for phenotypic selection in plant breeding and variety testing. Euphytica 2008, 161, 209–228. [Google Scholar] [CrossRef]
  22. Table 7.3a to 7.3c Barley; Production, Value, Supply and Use. Available online: https://www.gov.uk/government/statistics/agriculture-in-the-united-kingdom-2022/chapter-7-crops#barley (accessed on 12 September 2024).
  23. Yang, C.J.; Russell, J.; Ramsay, L.; Thomas, W.; Powell, W.; Mackay, I. Overcoming barriers to the registration of new plant varieties under the DUS system. Commun. Biol. 2021, 4, 302. [Google Scholar] [CrossRef]
  24. Looseley, M.E.; Ramsay, L.; Bull, H.; Swanston, J.S.; Shaw, P.D.; Macaulay, M.; Booth, A.; Russell, J.R.; Waugh, R.; Thomas, W.T.B. Association mapping of malting quality traits in UK spring and winter barley cultivar collections. Theor. Appl. Genet. 2020, 133, 2567–2582. [Google Scholar] [CrossRef]
  25. Bayer, M.M.; Rapazote-Flores, P.; Ganal, M.; Hedley, P.E.; Macaulay, M.; Plieske, J.; Ramsay, L.; Russell, J.; Shaw, P.D.; Thomas, W.; et al. Development and evaluation of a barley 50k iSelect SNP array. Front Plant Sci. 2017, 8, 1792. [Google Scholar] [CrossRef] [PubMed]
  26. Pebesma, E.; Bivand, R. Spatial Data Science: With Applications in R; Chapman and Hall: New York, NY, USA, 2023. [Google Scholar]
  27. Covarrubias-Pazaran, G. Genome-assisted prediction of quantitative traits using the R package sommer. PLoS ONE 2016, 11, e0156744. [Google Scholar] [CrossRef] [PubMed]
  28. Raymond, J.; Mackay, I.; Penfield, S.; Lovett, A.; Philpott, H.; Dorling, S. Continuing genetic improvement and biases in genetic gain estimates revealed in historical UK variety trials data. Field Crop Res. 2023, 302, 109086. [Google Scholar] [CrossRef]
  29. Williams, R.M.; O’Brien, L.; Eagles, H.A.; Solah, V.A.; Jayasena, V. The influences of genotype, environment, and genotype by environment interaction on wheat quality. Aust. J. Agric Res. 2008, 59, 95–111. [Google Scholar] [CrossRef]
  30. Zhao, F.; Xu, S. Genotype by environment interaction of quantitative traits: A case study in barley. G3 2012, 2, 779–788. [Google Scholar] [CrossRef]
  31. Lopez-Cruz, M.; Aguate, F.M.; Washburn, J.D.; de Leon, N.; Kaeppler, S.M.; Lima, D.C.; Tan, R.; Thompson, A.; Bretonne, L.W.D.L.; de los Campos, G. Leveraging data from the Genomes-to-Fields Initiative to investigate genotype-by-environment interactions in maize in North America. Nat. Comm. 2023, 14, 6904. [Google Scholar] [CrossRef]
  32. Clark, S.A.; van der Werf, J. Genomic Best Linear Unbiased Prediction (gBLUP) for the Estimation of Genomic Breeding Values; Humana Press: Totowa, NJ, USA, 2013. [Google Scholar]
  33. Hickey, J.M.; Chiurugwi, T.; Mackay, I.; Powell, W. Genomic prediction unifies animal and plant breeding programs to form platforms for biological discovery. Nat. Genet. 2017, 49, 1297–1303. [Google Scholar] [CrossRef]
  34. Edwards, S.M.; Buntjer, J.B.; Jackson, R.; Bentley, A.R.; Lage, J.; Byrne, E.; Burt, C.; Jack, P.; Berry, S.; Flatman, E.; et al. The effects of training population design on genomic prediction accuracy in wheat. Theor. Appl. Genet. 2019, 132, 1943–1952. [Google Scholar] [CrossRef]
  35. Voss-Fels, K.P.; Cooper, M.; Hayes, B.J. Accelerating crop genetic gains with genomic selection. Theor. Appl. Genet. 2019, 132, 669–686. [Google Scholar] [CrossRef]
  36. MacDonald, H.; Henry, R.J. Balancing incentives for innovation in new plant varieties. Nat. Plants 2022, 8, 1206–1208. [Google Scholar] [CrossRef]
  37. Mott, R.; Fischer, C.; Prins, P.; Davies, R.W. Private genomes and public SNPs: Homomorphic encryption of genotypes and phenotypes for shared quantitative genetics. Genetics 2020, 215, 359–372. [Google Scholar] [CrossRef] [PubMed]
  38. Zhao, T.; Wang, F.; Mott, R.; Dekkers, J.; Cheng, H. Using encrypted genotypes and phenotypes for collaborative genomic analyses to maintain data confidentiality. Genetics 2024, 226, iyad210. [Google Scholar] [CrossRef] [PubMed]
  39. Silvey, V. NIAB recommended lists of cereal varieties: An aid to orderly marketing in the United Kingdom. Plant Var. & Seeds 1999, 12, 23–42. [Google Scholar]
  40. McVittie, J.A.; Jellis, G.J.; Flux, A. Evaluation of wheat quality for the UK in the HGCA recommended list system. In Proceedings of the 12th International ICC Cereal and Bread Congress, Harrogate, UK, 24–26 May 2004; pp. 13–17. [Google Scholar]
  41. Macaulay, M.; Russell, J.; Ramsay, L.; Marshall, D.F.; Powell, W.; Waugh, R.; Thomas, W.T.B. Lessons from Barley Recommended Lists; Scottish Crop Research Institute: Invergowrie, UK, 2005. [Google Scholar]
  42. Mackay, I.; Horwell, A.; Garner, J.; White, J.; McKee, J.; Philpott, H. Reanalyses of the historical series of UK variety trials to quantify the contributions of genetics and environmental factors to trends and variability in yield over time. Theor. Appl. Genet. 2011, 122, 225–238. [Google Scholar] [CrossRef] [PubMed]
  43. Marshall, E. Impact of delaying autumn sowing and rotations on wheat yield and quality using AHDB recommended list trial data: Implications for control of Alopecurus myosuroides. Asp. Appl. Biol. 2017, 134, 127–136. [Google Scholar]
  44. Oberfoster, M.; Marshall, E. Descriptive and Recommended Variety Lists in European Countries and Their Scales; BOKU-University of Natural Resources and Life Sciences: Vienna, Austria, 2017. [Google Scholar]
  45. Dias, K.O.G.; dos Santos, J.P.R.; Krause, M.D.; Piepho, H.-P.; Guimaraes, L.J.M.; Pastina, M.M.; Garcia, A.A.F. Leveraging probability concepts for cultivar recommendation in multi-environment trials. Theor. Appl. Genet. 2022, 135, 1385–1399. [Google Scholar] [CrossRef]
  46. Smith, A.B.; Cullis, B.R.; Thompson, R. The analysis of crop cultivar breeding and evaluation trials: An overview of current mixed model approaches. J. Agric Sci. 2005, 143, 449–462. [Google Scholar] [CrossRef]
  47. Molenaar, H.; Boehm, R.; Piepho, H.-P. Phenotypic selection in ornamental breeding: It’s better to have the BLUPs than to have the BLUEs. Front Plant Sci. 2018, 9, 1511. [Google Scholar] [CrossRef]
  48. Gauch, H.G. Model selection and validation for yield trials with interaction. Biometrics 1988, 44, 705–715. [Google Scholar] [CrossRef]
  49. Annicchiarico, P. Additive main effects and multiplicative interaction (AMMI) analysis of genotype-location interaction in variety trials repeated over years. Theor. Appl. Genet. 1997, 94, 1072–1077. [Google Scholar] [CrossRef]
  50. Piepho, H.-P. Best linear unbiased prediction (BLUP) for regional yield trials: A comparison to additive main effects multiplicative interaction (AMMI) analysis. Theor. Appl. Genet. 1994, 89, 647–654. [Google Scholar] [CrossRef] [PubMed]
  51. Piepho, H.-P. Analyzing genotype-environment data by mixed models with multiplicative effects. Biometrics 1997, 53, 761–766. [Google Scholar] [CrossRef]
  52. Piepho, H.-P.; Williams, E. Factor-analytic variance-covariance structures for prediction into target populations of environments. Biom. J. 2024, 66, e202400008. [Google Scholar] [CrossRef] [PubMed]
  53. Newman, S.J.; Furbank, R.T. Explainable machine learning models of major crop traits from satellite-monitored continent-wide field trial data. Nat. Plants 2021, 7, 1354–1363. [Google Scholar] [CrossRef]
  54. Balakrishnan, D.; Kumar, A.P.; Reddy, K.S.K.; Kumar, R.R.; Aadith, K.; Madhan, S. Agricultural crop recommendation system. In Proceedings of the 2023 3rd International Conference on Intelligent Technologies (CONIT), Hubli, India, 23–25 June 2023. [Google Scholar]
  55. Hasan, M.; Marjan, M.A.; Uddin, M.P.; Afjal, M.I.; Kardy, S.; Ma, S.; Nam, Y. Ensemble machine learning-based recommendation system for effective prediction of suitable agricultural crop cultivation. Front Plant Sci. 2023, 14, 1234555. [Google Scholar] [CrossRef]
  56. Han, Y.; Wang, K.; Zhang, Q.; Yang, F.; Pan, S.; Liu, Z.; Zhang, Q. Developing a comprehensive evaluation model of variety adaptability based on machine learning method. Field Crops Res. 2024, 306, 109203. [Google Scholar] [CrossRef]
  57. Shams, M.Y.; Gamel, S.A.; Talaat, F.M. Enhancing crop recommendation systems with explainable artificial intelligence: A study on agricultural decision-making. Neural. Comput. Applic. 2024, 36, 5695–5714. [Google Scholar] [CrossRef]
Agronomy 14 02267 g001
Figure 1. Trial distribution, climate variation, and yield prediction: (A) Geographical distribution of UK counties with spring barley trials. (B) Climate variation from north (1) to south (24) and separation of trial sites into three regions (North, West, East). (C) Predicted versus observed yield ranks from six methods in 2007 and 2016 at Coaltown of Balgonie, which is located in Fife, Scotland, and close to weather station 3.
Figure 1. Trial distribution, climate variation, and yield prediction: (A) Geographical distribution of UK counties with spring barley trials. (B) Climate variation from north (1) to south (24) and separation of trial sites into three regions (North, West, East). (C) Predicted versus observed yield ranks from six methods in 2007 and 2016 at Coaltown of Balgonie, which is located in Fife, Scotland, and close to weather station 3.
Agronomy 14 02267 g001
Agronomy 14 02267 g002
Figure 2. Comparison of eight different methods for predicting variety performance. Each boxplot contains correlations between predicted and observed yield ranks from 29 unique trial sites and 13 years (2007 to 2019). The lower, middle, and upper hinges of the boxplots represent the first, second (median), and third quartiles of the data points. The whiskers extend from the boxes to the furthest data points within 1.5 times the interquartile range. The vertical red line serves as a reference based on the largest median correlation from method A5B.
Figure 2. Comparison of eight different methods for predicting variety performance. Each boxplot contains correlations between predicted and observed yield ranks from 29 unique trial sites and 13 years (2007 to 2019). The lower, middle, and upper hinges of the boxplots represent the first, second (median), and third quartiles of the data points. The whiskers extend from the boxes to the furthest data points within 1.5 times the interquartile range. The vertical red line serves as a reference based on the largest median correlation from method A5B.
Agronomy 14 02267 g002
Agronomy 14 02267 g003
Figure 3. Correlations between predicted and observed yield ranks in six trial sites: (A) Heatmap of correlations for eight methods and 13 years. These six trial sites are highlighted here because they have the greatest number of trials across the years. (B) Scatter plot of correlations for eight methods and 13 years. Significant difference in mean correlations between methods after applying Bonferroni correction is indicated by a black vertical line.
Figure 3. Correlations between predicted and observed yield ranks in six trial sites: (A) Heatmap of correlations for eight methods and 13 years. These six trial sites are highlighted here because they have the greatest number of trials across the years. (B) Scatter plot of correlations for eight methods and 13 years. Significant difference in mean correlations between methods after applying Bonferroni correction is indicated by a black vertical line.
Agronomy 14 02267 g003
Agronomy 14 02267 g004
Figure 4. Correlations between predicted and observed yield ranks over 13 years: (A) Scatter plot of correlations for eight methods and trial sites within each year. Significant differences in mean correlations between methods after applying Bonferroni correction are indicated by black vertical lines. (B) Line plot of mean correlations for eight methods shown from 2007 to 2019.
Figure 4. Correlations between predicted and observed yield ranks over 13 years: (A) Scatter plot of correlations for eight methods and trial sites within each year. Significant differences in mean correlations between methods after applying Bonferroni correction are indicated by black vertical lines. (B) Line plot of mean correlations for eight methods shown from 2007 to 2019.
Agronomy 14 02267 g004
Agronomy 14 02267 g005
Figure 5. Distribution of yield deficits across counties. Percent Deficits (PD) in yield due to incorrect predictions of best local variety are quantified for each trial site and year. The heat map shows the county-level PD as the means across all trial sites and years within each county.
Figure 5. Distribution of yield deficits across counties. Percent Deficits (PD) in yield due to incorrect predictions of best local variety are quantified for each trial site and year. The heat map shows the county-level PD as the means across all trial sites and years within each county.
Agronomy 14 02267 g005
Agronomy 14 02267 g006
Figure 6. Cumulative Percent Deficits (CPD) in yield over the years. PD is averaged across all trial sites within each year to obtain the annual mean PD, which is then summed over the years from 2007 to 2019 to highlight the CPD in yield for each method.
Figure 6. Cumulative Percent Deficits (CPD) in yield over the years. PD is averaged across all trial sites within each year to obtain the annual mean PD, which is then summed over the years from 2007 to 2019 to highlight the CPD in yield for each method.
Agronomy 14 02267 g006
Table 1. Weather station data. Out of 37 weather stations with archived records from the UK Met Office, 24 with close proximity to spring barley trial sites are shown here. For each weather station, temperature (maximum/minimum/range), number of frost days, and rainfall are provided as means from March to August from 2002 to 2019.
Table 1. Weather station data. Out of 37 weather stations with archived records from the UK Met Office, 24 with close proximity to spring barley trial sites are shown here. For each weather station, temperature (maximum/minimum/range), number of frost days, and rainfall are provided as means from March to August from 2002 to 2019.
IDWeather StationLong. (°)Lat. (°)Tmax (°C)Tmin (°C)Trange (°C)Frost (Day)Rain (mm)
1Nairn57.593−3.82115.267.058.201.9356.80
2Braemar57.011−3.39613.954.889.075.4866.70
3Leuchars56.377−2.86115.367.308.062.0261.44
4Paisley55.846−4.43016.138.337.791.0983.03
5Eskdalemuir55.312−3.20514.415.938.483.93129.14
6Ballypatrick Forest55.181−6.15313.847.566.280.8896.95
7Durham54.768−1.58516.197.688.511.6356.65
8Whitby54.481−0.62416.108.327.771.0452.10
9Armagh54.352−6.64916.347.978.371.4367.24
10Bradford53.813−1.77216.198.327.871.2366.22
11Sheffield53.381−1.49016.978.938.040.8165.04
12Waddington53.175−0.52217.128.718.411.0055.88
13Shawbury52.794−2.66316.917.669.252.2054.92
14Lowestoft52.4831.72716.819.487.320.7149.73
15Cambridge NIAB52.2450.10218.338.529.811.4944.68
16Ross-On-Wye51.911−2.58417.978.769.021.2658.65
17Oxford51.761−1.26218.519.059.460.9753.87
18Heathrow51.479−0.44919.1610.049.110.6446.79
19Manston51.3461.33717.359.687.670.4846.52
20Chivenor51.089−4.14716.969.627.350.6665.27
21Yeovilton51.006−2.64117.738.299.442.0352.41
22Hurn50.779−1.83517.988.219.772.2557.35
23Eastbourne50.7590.28517.1710.596.580.2848.14
24Camborne50.218−5.32715.289.835.450.2269.87
Table 2. Summary of spring barley trial sites. Out of 80 unique trial sites, 29 with sufficient data were included in our analyses. For each of these sites, additional geographical information, count of analyzed years, and closest weather stations are provided.
Table 2. Summary of spring barley trial sites. Out of 80 unique trial sites, 29 with sufficient data were included in our analyses. For each of these sites, additional geographical information, count of analyzed years, and closest weather stations are provided.
SiteLat. (°)Long. (°)CountyRegionNationCountWeather Station
Tain57.812−4.055HighlandNorthScotland10Nairn, 28 km
Sandend57.685−2.748AberdeenshireNorthScotland1Nairn, 65 km
Inverurie57.284−2.374AberdeenshireNorthScotland2Braemar, 69 km
Laurencekirk56.832−2.468AberdeenshireNorthScotland11Leuchars, 56 km
Coupar Angus56.547−3.264Perth and KinrossNorthScotland1Leuchars, 31 km
Dundee56.465−2.971City of DundeeNorthScotland1Leuchars, 12 km
Perth56.394−3.432Perth and KinrossNorthScotland5Leuchars, 35 km
Kinross56.208−3.423Perth and KinrossNorthScotland1Leuchars, 40 km
Coaltown of Balgonie56.185−3.126FifeNorthScotland11Leuchars, 27 km
East Saltoun55.901−2.840East LothianNorthScotland8Leuchars, 53 km
Lanark55.674−3.776South LanarkshireWestScotland3Paisley, 45 km
St Boswells55.570−2.646Scottish BordersNorthScotland10Eskdalemuir, 46 km
Ayr55.458−4.629South AyrshireWestScotland1Paisley, 45 km
Strabane54.827−7.463TyroneWestN Ireland4Armagh, 75 km
Crossnacreevy54.556−5.849DownWestN Ireland6Armagh, 57 km
Ballywalter54.543−5.484DownWestN Ireland1Armagh, 79 km
Abergavenny51.825−3.019MonmouthshireWestWales2Ross-On-Wye, 31 km
Bowsden55.669−2.014NorthumberlandNorthEngland2Eskdalemuir, 85 km
Bainton53.957−0.533East YorkshireEastEngland6Whitby, 59 km
High Legh53.354−2.451CheshireWestEngland4Sheffield, 64 km
Horncastle53.208−0.113LincolnshireEastEngland7Waddington, 28 km
Edgmond52.775−2.412ShropshireWestEngland5Shawbury, 17 km
Kings Lynn52.7520.402NorfolkEastEngland8Cambridge NIAB, 60 km
Wymondham52.5691.115NorfolkEastEngland11Lowestoft, 43 km
Fulbourn52.1830.222CambridgeshireEastEngland1Cambridge NIAB, 11 km
Callow52.005−2.739HerefordshireWestEngland4Ross-On-Wye, 15 km
Cirencester51.718−1.969GloucestershireWestEngland4Ross-On-Wye, 48 km
Stockbridge51.117−1.486HampshireWestEngland9Oxford, 73 km
Kingsbridge50.283−3.777DevonWestEngland4Chivenor, 93 km
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Yang, C.J.; Russell, J.; Mackay, I.; Powell, W. Opportunities to Improve the Recommendation of Plant Varieties under the Recommended List (RL) System. Agronomy 2024, 14, 2267. https://doi.org/10.3390/agronomy14102267

AMA Style

Yang CJ, Russell J, Mackay I, Powell W. Opportunities to Improve the Recommendation of Plant Varieties under the Recommended List (RL) System. Agronomy. 2024; 14(10):2267. https://doi.org/10.3390/agronomy14102267

Chicago/Turabian Style

Yang, Chin Jian, Joanne Russell, Ian Mackay, and Wayne Powell. 2024. "Opportunities to Improve the Recommendation of Plant Varieties under the Recommended List (RL) System" Agronomy 14, no. 10: 2267. https://doi.org/10.3390/agronomy14102267

APA Style

Yang, C. J., Russell, J., Mackay, I., & Powell, W. (2024). Opportunities to Improve the Recommendation of Plant Varieties under the Recommended List (RL) System. Agronomy, 14(10), 2267. https://doi.org/10.3390/agronomy14102267

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop