Potential risks of future herbicide-resistant weeds in New Zealand revealed through machine learning

ABSTRACT

New Zealand has fewer numbers of herbicide-resistant crop weeds than many other highly developed economies, yet these numbers are likely to increase in the future. A clear indication of the scale of this risk can be derived from the predictable structure in the global occurrence of herbicide-resistant weeds that results from similarities in agronomic and environmental conditions. Hierarchical cluster analysis was used to identify groups of countries with similar herbicide-resistant weed assemblages to New Zealand. A distinct cluster of 11 European countries with strong climatic and agronomic affinities to New Zealand was identified. The combined assemblage of herbicide-resistant weeds within this cluster consisted of 27 species and the potential risk of a species evolving herbicide resistance was calculated as its frequency among these European countries. Species with potential to become herbicide resistant in New Zealand included established crop weeds (e.g. Senecio vulgaris, Tripleurospermum inodorum) as well as species only encountered as contaminants of seed imports (e.g. Alopecurus myosuroides, Apera spica-venti). All eight species already known to be herbicide-resistant in New Zealand were found in the high-risk assemblage and this indicates that the analysis provided a realistic measure of future risk.

KEYWORDS:

• Arable crops
• machine learning
• risk assessment
• sustainable agriculture
• weed management
• weeds

Introduction

The growing number of crop weeds exhibiting resistance to synthetic herbicides poses an increasing problem to sustainable agricultural production (MacLaren et al. 2020; Gaines et al. 2021). Given its level of agricultural intensification and herbicide inputs, New Zealand has experienced fewer problems with herbicide-resistant crop weeds than many other similar countries (Hulme 2022a; Harrington and Ghanizadeh 2023; Hulme 2023). Nevertheless, both the number of weed species becoming resistant to herbicides and the number of cases of resistance to specific herbicide modes of action have been steadily increasing in New Zealand since the 1970s (Ghanizadeh and Harrington 2021). Currently in New Zealand, there are 15 weed species recorded as being resistant to one or more of eight different herbicide modes of action, with the majority of resistance cases found in arable crops (Heap 2023). There is considerable scope for these numbers to increase in the future since many crop weeds in New Zealand are known to have evolved herbicide resistance under similar agronomic regimes overseas implying a significant future risk (Ngow et al. 2020). For this reason, the risk posed by new herbicide-resistant weeds has been highlighted as one of the primary research challenges facing the New Zealand agricultural sector (Goldson et al. 2015; Buddenhagen et al. 2022).

Although the use of species traits has proven useful to distinguish the characteristics that might predict whether a weed will become resistant to herbicides or not (Holt et al. 2013; Hartway et al. 2022; Hulme and Liu 2022), the relatively few herbicide-resistant weeds in New Zealand considerably limits the statistical power of such analyses. Thus alternative attempts to predict the magnitude of this risk in New Zealand have largely relied upon scoring a species in relation to the number of cases worldwide where it has become herbicide resistant (Ngow et al. 2020, 2021). While such methods are a valuable first step, they are limited in that they generate long species lists with no objective cut-off as to how many cases of resistance are needed for a species to be viewed as a risk to New Zealand. Furthermore, using the number of cases found globally results in the grouping together of many different agronomic and environmental conditions that will not reflect the actual situation in New Zealand. The increasing availability of data on herbicide resistance worldwide (Heap 2014) combined with the analytical power of machine learning tools provides an alternative approach to assessing which weed species pose the greatest risk of becoming herbicide resistant in New Zealand. Such an analysis seeks to understand whether there is a predictable structure to the global occurrence of herbicide-resistant weeds that reflects shared ‘hidden attributes’ that are shaped by regional variations in climate, cropping regimes and weed composition and whether these attributes allow prediction of species likely to become herbicide-resistant in new regions. Machine learning tools are particularly well suited to identify the existence of such predictable structure in global herbicide-resistant weed assemblages.

Machine learning tools, particularly advanced clustering techniques, have previously used the global assemblages of organisms to generate robust and accurate predictions regarding the establishment risks of insect pests (Watts and Worner 2009; Duffy et al. 2021), plant parasitic nematodes (Singh et al. 2014), plant pathogens (Watts 2011; Eschen et al. 2014) and weeds (Morin et al. 2013). Among these tools, hierarchical cluster analysis has proven to be a computationally efficient method that captures the high-level structure found among and within different clusters facilitating clearer explanation of the underlying patterns (Worner et al. 2013, 2015). These powerful techniques, however, have never been applied to prediction of herbicide resistance in weeds. Here, the considerable collation of data on the occurrence of herbicide-resistant weeds worldwide (Heap 2023) is used for the first time as a basis for hierarchical clustering analysis to generate predictions of which weed species are most at risk of becoming herbicide-resistant in New Zealand.

Materials and methods

Data assembly

Data on the occurrence of herbicide resistance in weed species worldwide was extracted from the International Herbicide-Resistant Weed Database (www.weedscience.org) accessed on March 1st 2023. These data on herbicide resistance were used to generate a binary matrix (resistant or not) of weed species (columns) in relation the different countries worldwide for which herbicide resistance has been recorded (rows). The matrix was consolidated to only include weed species known to be resistant to herbicides in at least five countries and, following this filter, only those countries that subsequently had at least five different weed species known to be herbicide resistant. The final matrix comprised 29 countries and 36 weed species. This consolidation was important to reduce the sparseness of the dataset to enable countries to be more robustly clustered and avoid the undue weight of any rare species found to be herbicide-resistant in only one country. Nevertheless, this data consolidation meant that seven out of the 15 herbicide-resistant weed species known from New Zealand had to be excluded since they have not yet been recorded as herbicide resistant anywhere else in the world.

Hierarchical cluster analysis

Hierarchical cluster analysis was used to identify groups of similar countries in relation to the composition of their herbicide-resistant weed flora. The analysis uses machine-learning to partition data into a hierarchy of clusters of countries that show high-within cluster homogeneity and high between cluster heterogeneity (Ma et al. 2020). Thus, the resulting clusters should reveal suites of countries that tend to be more strongly linked in terms of herbicide-resistant weeds. Following Hulme (2022b) an agglomerative algorithm was applied using the Jaccard similarity between two clusters by comparing their most dissimilar members (furthest neighbours). The optimum number of clusters was determined using the Elbow Method by plotting the agglomerative coefficient against the number of clusters and identifying an elbow in the curve where the rate of change in the agglomeration coefficient declined markedly (Madhulatha 2012). Hierarchical cluster analysis was undertaken using SPSS v 29 (IBM Corp. 2019).

Herbicide resistance risk ranking

In order to identify weed species with a high potential of herbicide resistance in New Zealand, the combined list of all herbicide-resistant weed species found across all other countries within the same cluster as New Zealand was generated. The frequency with which different species had been recorded as herbicide-resistant across all the other countries in the cluster was used as an index of the putative risk of a species becoming herbicide-resistant in New Zealand. The relative ranking of weed species known to be herbicide-resistant in New Zealand was used to gauge the likelihood that species not known to be resistant to herbicides might become so in due course. Data on the number of herbicide modes of action and the number of global cases of herbicide resistance were also extracted from the International Herbicide-Resistant Weed Database to see if these values correlated with the species relative ranking.

Results

Hierarchical cluster analysis

Although there was no strongly marked elbow in the relationship between cluster number and the agglomeration coefficient, the steepest decline in the agglomeration coefficient pointed to four major clusters of countries that were linked by the similarity in composition of their herbicide-resistant weed flora (Figure 1). The hierarchical cluster dendrogram, indicated that these four clusters were discriminated at the top of the hierarchy and represent fundamental divisions between groups of countries, though some marked subclusters are also evident (Figure 2). The largest cluster comprised New Zealand and 11 central (e.g. Germany, Switzerland, Poland, the Czech Republic etc.) or northern (e.g. Denmark, Norway, Sweden, the United Kingdom) European countries. Within this cluster, New Zealand appeared more closely aligned with the UK, Norway, and the Netherlands. Australia was included in the second largest cluster that mostly comprised countries that bordered the Mediterranean basin (e.g. Greece, Italy, Israel, Spain, Turkey etc.) but also countries where a large part of the national area could be classified as arid or semi-arid (e.g. Argentina, Iran, South Africa etc.). The third largest cluster included four out of the five largest (by area) countries in the world: Canada, China, the United States and Brazil. Geographic or climatic similarities among the two members (Portugal and Chile) of the fourth and smallest cluster were less clear.

Figure 1. Elbow plot illustrating the rate of change in the agglomeration coefficient as the number of clusters identified increases. The sharpest decline was found between three and four clusters, identifying four clusters (vertical dotted line) as the most parsimonious number to characterise variation in herbicide-resistant weed assemblages worldwide.

Figure 2. Hierarchical cluster analysis dendrogram identifying four main country clusters in herbicide-resistant weed assemblages. Clusters are labelled 1–4 in order of the number of countries included in each cluster.

Herbicide resistance risk ranking

Across the 11 countries included in the New Zealand cluster, a total of 27 weed species had been recorded as resistant to one or more herbicides (Figure 3). Species within the Poaceae (10 species) and Asteraceae (7 species) were particularly well represented and three genera (Amaranthus, Erigeron and Lolium) accounted for one third of the taxa. Considerable variation occurred among species in relation to the frequency with which they were found to be resistant to herbicides with Alopecurus myosuroides and Chenopodium album being found resistant in ten out of the 11 countries (90.9%), while Erigeron bonariensis, E. sumatrensis and Lolium rigidum were all found in only one country (9.1%). All eight weed species known to be herbicide resistant in New Zealand were found at least once in the 11 European countries comprising this cluster. If this had not been the case, and herbicide-resistant weeds in New Zealand had only been found in countries in one of the other three clusters, then this would have increased the uncertainty regarding the value of the hierarchical cluster analysis to predict likelihoods of weeds developing herbicide resistance.

Figure 3. Herbicide-resistant weed species ranked in terms of their frequency of occurrence among 11 European countries identified as strongly associated with New Zealand using hierarchical cluster analysis. Species associated with blue bars are those already known to be herbicide-resistant weeds in New Zealand.

Weeds resistant to herbicides in New Zealand were found throughout the ranked list of species and suggests, possibly except for the three species found resistant in only one country, that most of the species resistant to herbicides in these 11 European countries have the potential to become resistant in New Zealand. The five highest ranked species would appear to be A. myosuroides, Apera spica-venti, Senecio vulgaris, Tripleurospermum inodorum and Papaver rhoeas. The rank order of species was not correlated with either the number of herbicide modes of action to which a species had become resistant (r = 0.210, df 17, P = 0.388) nor the total number of cases of resistance found for that species worldwide (r = 0.210, df 17, P = 0.824). More detailed examination within the specific subcluster that included New Zealand, the Netherlands, Norway and the United Kingdom, highlighted S. vulgaris to be the only herbicide resistant weed species found in all three European countries that was not resistant to herbicides in New Zealand. However, the sparsity of species data in this subcluster limited its discriminatory power and the generality of any trends in species ranking.

Discussion

Representativeness in cluster composition for New Zealand

Machine learning clustering tools can provide powerful insights into the risks posed by agricultural pests (Worner et al. 2013) but this study is the first to apply such techniques to the assessment of herbicide-resistant weed risks. The hierarchical cluster analysis revealed biogeographically sensible clusters that could be mapped onto the climate similarities among countries. Thus, New Zealand was found in a cluster comprising countries with broadly similar temperate climates while in contrast Australia was more closely linked to countries with Mediterranean, semi-arid, or arid climates. This marked distinction between the clusters that included New Zealand and Australia reflects the marked differences in their agronomic and climatic environments (Harrington and Ghanizadeh 2023). A third cluster comprised large countries that are likely to experience a diversity of climates.

One caveat of these biogeographic patterns is that the clusters could simply reflect similarities among the floras of these regions, rather than specifically trends in herbicide resistance. Countering this view is that the herbicide-resistant weeds analysed in this study are largely cosmopolitan species with large geographic ranges and thus do not represent localised assemblages specific to certain subsets of countries. Indeed, many of the clusters include countries from quite distinct floristic zones (Takhtajan 1986) which would not be expected if the driving force of cluster membership was the similarity in the native flora. In addition, the agricultural weed flora of New Zealand and Australia comprises almost exclusively of non-native species introduced from multiple different regions of the world (Ikegami et al. 2019). Thus at least for these two countries, the patterns depicted in the cluster analysis may reflect underlying agronomic practices, such as the intensity of herbicide use and crop rotation regimes, rather than simply similarity of the native flora.

New Zealand sits within a cluster comprising 11 European countries that have been assigned to the EPPO Maritime agroecological zone in terms of their similarity in crop growth conditions, where crops grow in moderately cool or cold winters and fairly mild summer temperatures, with relatively wet winters and wet to occasionally dry summers (Bouma 2005). Agriculture land-use in the EPPO Maritime region is largely arable production with the dominant crop being winter wheat and other cereals (Hurford et al. 2020). These features mirror the climatic and agronomic situation in New Zealand (Millner and Roskruge 2013). Herbicide inputs are similar across these regions as well (Hulme 2022a, 2023). Thus, it appears this cluster of countries is appropriate to use the herbicide-resistant weed assemblage as a guide to future herbicide resistance risk in New Zealand.

Robustness of the predictions of herbicide-resistant weed risk

The suitability of the species assemblage within the cluster for predicting the potential for weeds to become herbicide resistant in New Zealand is supported by the finding that all eight species in the dataset known to be herbicide resistant in New Zealand occurred in the species assemblage. Most of the species listed as potential risks are established and widespread in New Zealand, the exceptions being A. myosuroides which is currently represented by a local incursion that is the subject of an eradication programme and A. spica-venti that, while not known to be established in New Zealand, is an unwanted organism that is an occasional contaminant of imported crop seed used for sowing (Rubenstein et al. 2021). A further risk, is that such contaminants might themselves be herbicide resistant resulting in the problem being imported rather than evolving locally as a result of prevailing agronomic practices (Shimono et al. 2020). New Zealand imports arable seed from many of the 11 European countries in its cluster and this could be a route for the introduction of weed seeds that have already developed herbicide resistance (Buddenhagen et al. 2021).

There was no clear threshold that might indicate that some of the species established in New Zealand would be unlikely to evolve herbicide resistance. Ten of the species have previously been assessed in terms of their risk of evolving herbicide resistance in New Zealand wheat and barley crops, with seven being classed as high risk, while the remaining three were deemed a medium risk (Ngow et al. 2020). The similarity between the two different approaches to assessing the risk of future herbicide resistant weeds in New Zealand is reassuring, however the hierarchical cluster analysis is likely better tuned since it weighs the prevalence of herbicide resistance in an agronomically more similar region than simply taking account of the global number of cases of resistance found for each species. Supporting this view is the finding that the prevalence of herbicide resistance in the countries included in the New Zealand cluster was not correlated with the global number of cases of herbicide resistance found for the species. A further advantage of the hierarchical cluster analysis is that it can also bring to light species that might otherwise be missed, such as A. spica-venti, because their main threat is through the accidental importation of herbicide resistant seeds (Rubenstein et al. 2021).

Nevertheless, the hierarchical cluster analysis only presents a first screen of the potential risk of herbicide resistance evolving in a particular weed species. To provide a more targeted assessment would require data on the prevalence of the weed species in crops and their likely exposure to herbicides. Weed prevalence has previously been shown to be an important determinant of the likelihood that a weed species will evolve herbicide resistance (Hulme and Liu 2022). Neither weed prevalence nor herbicide application regimes can be assumed to be the same in New Zealand as in other countries. For example, Papaver rhoeas, one of the highest ranked species, is the most common broadleaf weed in winter cereals in Europe (Busi et al. 2018) but although widespread in New Zealand it is relatively uncommon as a crop weed as a result of good control with phenoxy herbicides (Popay et al. 2010). The relative low prevalence of this species in cereal crops in New Zealand would likely result in low exposure to herbicides and thus weak selection pressure for the evolution of herbicide resistance. However, P. rhoeas has evolved resistance to phenoxy herbicides in Europe and is becoming a troublesome weed, especially in warmer climates, requiring a change to agronomic management of cereal crops (Busi et al. 2018). Given this evidence of herbicide-resistance, the successful control of P. rhoeas in New Zealand should not be taken for granted especially since this species will also likely benefit from less predictable precipitation regimes under future climate change (March-Salas et al. 2019). Other weed species identified in the cluster analysis may not currently be problematic in crops but occur in environments such as highways, fence lines or gardens where they will nevertheless still be exposed to herbicides.

Limitations to machine learning in herbicide risk assessment

An important limitation in the hierarchical cluster analysis is the need for species to have sufficient records of herbicide resistance outside of New Zealand to be considered. Several weed species are known only to be herbicide resistant in New Zealand and nowhere else in the world: Carduus nutans, C. pycnocephalus, Nassella neesiana, Ranunculus acris and Soliva sessilis. These weeds are primarily problems of perennial grasslands and pasture in New Zealand (Ghanizadeh and Harrington 2019) but since they could not be included in the hierarchical cluster analysis it indicates that the results are likely more relevant to arable crops. The high data needs of machine learning tools also prevented more detailed analysis of herbicide resistance risks arising from specific herbicide modes of action or in certain crops. These limitations may be particularly marked when attempting to identify risk at a global scale due to the sparseness of records in some countries but might be overcome if targeting sub-regions in countries with extensive herbicide-resistant weed records such as individual states of the United States.

Risk assessment of herbicide resistance in crop weeds, while a valuable exercise, will only have an impact if it can influence risk management. Identifying a list of 19 new weed species with a high potential of evolving herbicide resistance will provide a basis for more targeted surveys of herbicide resistance in New Zealand (Buddenhagen et al. 2020), ensuring that resources are used effectively by researchers and extension specialists as well as raising awareness among the farming community in terms of which weeds to place on a watch list. Using this list as a starting point, farmers can assess the prevalence of these weeds in their crops and crop margins, consider their current herbicide application regimes and crop rotation programmes, and refine their assessment of future herbicide resistance risks at a farm scale. In addition, hierarchical clustering approaches can also aid in estimating the risk of cross- and multi-herbicide resistance occurring (Hulme 2022b). However, simply being able to detect herbicide resistance more effectively in these future weeds is not the best method of prevention, and knowledge of the weed flora in crop fields combined with an understanding of the high-risk species could inform herbicide management to reduce the risk of future resistance evolution. There are plenty of recommendations as to how farmers can reduce the risk of herbicide resistance evolution (Ghanizadeh and Harrington 2021) but being able to assess likelihoods of such events will depend on knowing the risks individual species in the crop pose in the first place.

Conclusions

A relatively simple, computationally efficient, machine learning tool was able to provide new insights into the future risk of herbicide-resistant weeds in arable crops in New Zealand. Hierarchical cluster analysis identified a distinct group of European countries with strong climatic and agronomic affinities to New Zealand enabling their assemblage of herbicide-resistant weeds to provide an objective indication of risk. The high potential risk weed assemblage included all species known to be herbicide-resistant in New Zealand and was consistent with prevailing views of these risks in arable crops. An advantage of this approach is that it was able to identify high-risk species that pose a threat should they become established in New Zealand following their unintentional entry as contaminants in imports of seed for sowing. Such information is not only useful for border biosecurity but can also assist farmers in assessing the risk their own agronomic practices might play in facilitating the evolution of herbicide-resistance in potentially high-risk weed species.

Acknowledgements

The author thanks Ian Heap for access to a subset of the International Herbicide-Resistant Weed Database. The research was funded by the New Zealand Ministry of Business, Innovation and Employment under the project ‘Managing Herbicide Resistance’ (grant number C10X1806).

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

All data used in this research are available from the International Herbicide-Resistant Weed Database (www.weedscience.org).

Additional information

Funding

This work was supported by Ministry for Business Innovation and Employment [grant number C10X1806].