Spatial weed distribution models under climate change: a short review

General output

Global climate change affects human health and wellbeing due to more extreme weather events and wildfires, decreased air quality, and diseases transmitted by insects, food, and water. Climate disruptions to agriculture have been increasing and are projected to become more severe over this century, a trend that would affect pest distributions and their effect on crops. In this regard, it is important to find out the future spatial distribution of weeds in order to predict and mitigate their impact on crops. Since the last few decades, species distribution models (SDMs) have been appearing to answer questions about how living beings could behave in the upcoming years. A huge number of groups of plants, animals, fungi, pathogens, etc. are being studied from a conservational and an economic point of view (Bellard et al., 2018; Lantschner, de la Vega & Corley, 2019). According to the latter, weeds are significantly important and must be kept under control in croplands. Many similar studies have been published in the past few years all over the world.

The results yielded were analysed one by one. After applying the criteria for the subject of this review, 59 articles were deemed suitable for a deeper assessment. The number of articles published per year was as follows: 12 in 2017, 7 in 2018, 14 in 2019, 14 in 2020, and 12 in 2021. Forty-two out of 59 articles dealt with only one species, 10 included between two and nine species, and six studied between 11 and 20 species. The remaining reference considered 32 species (Shabani et al., 2020).

Thirty-four out of 59 articles added forecasts in one or more scenarios to the study of the present period. Hindcasting was much rarer and only developed in four of them; only one article used remote periods like the last glacial maximum (~21,000 years BP) and mid-holocene (~6,000 BP), whilst the remaining three articles covered more recent periods during the 20th and the early 21st centuries (Hicks et al., 2021; Mang et al., 2018; Ren et al., 2020; Runquist et al., 2019).

In general terms, the authors of the articles reviewed did not mention the types of crops in which the weeds grew. In a few cases, beet, chaste tree, coconut, eucalyptus, maize, mustard, oats, palm oil, sugarcane, potato, rice, rubber, soybean, sunflower, sweet potato, teak, and wheat crops were mentioned.

Our literature review confirms the concern about weeds and the expected global changes (IPCC, 2007). Hindcasting does not seem to be of great interest to most of the researchers, while forecasting in the next few decades is the issue most studied. The authors of the few articles on hindcasting focused their research on recent periods, and their results could be of interest for studying weed distribution in the last few decades or years. Monitoring their tendency from the past to the future throughout the years and having broader predictions would be useful, but, in our opinion, examining hindcasting from other eras (a thousand years ago) does not make much sense.

Weeds studied

A total of 130 species were studied 194 times in the 59 articles. It should be noted that some researchers published more than one article on the same species.

The species most often studied was Ambrosia artemisiifolia L. (eight times), followed by Ageratina adenophora (Spreng.) R.M.King & H.Rob., Mikania micrantha Kunth, Parthenium hysterophorus L. (six times), and Imperata cylindrica (L.) Raeusch. The rest of the species were only studied once and can be consulted in Material S2.

The species studied mostly came either from non-native areas (108 times) or native and non-native areas together (79 times)—the latter, generally on a global scale. Only seven times were the species studied in their natural habitat. This is consistent with the expected predictions on climate change and the consequences it could have on invasive species (Elith, 2017). The most flagrant evidence for studying alien weeds rather than native ones can be attributed, in our view, to the fact that exotic plants could become more aggressive and invasive in the current context of climate change. Globalization could also be involved in this trend, since nowadays tourism and business promote seed exchange worldwide (Song, Li & Cao, 2018).

The families with the species more studied in the reviewed articles were Asteraceae (36 species) and Poaceae (23 species), i.e., both families assumed 45% of the total. Families with more than one studied weed in the reviewed articles are shown in Table 1. The remaining 19 families were represented by only one species (see Material S2).

The most common families were Asteraceae, Poaceae, and Fabaceae (Devesa Alcaraz & Carrión García, 2012). Therefore, weeds are also well represented by these families, and they have been extensively investigated in the last few years, according to our results. Ambrosia artemisiifolia is an aggressive weed that has been considered from ecophysiological, allergenic, and chemical points of view (Willemsen & Rice, 1972; Bazzaz, 1974; Wang, Xu & Wu, 1993; Klimek, Becker & Raulf, 2014). In fact, according to our review, the Asteraceae species is the one most investigated. Other families like Chenopodiaceae and Amaranthaceae were also of interest, especially the Amaranthus genus. The repetition of some emblematic taxa was due to their concern and interest. For instance, in China, several species were repeatedly studied by Wan & Wang (2019a, 2019b).

Origin of data: variables and databases

Climatic or bioclimatic variables were always used when setting up models, following the example of Bede-Fazekas & Somodi (2020), who frequently considered them. In some cases, these explanatory variables were used alone, whereas in others, they were combined with topographic, pedological, anthropogenic (human footprint), and, to a lesser extent, other biotic variables.

Topographic variables also supported the climatic or bioclimatic ones, although the authors were generally accurate, and correlation matrices were computed in order to prevent collinearity between variables. Human footprints or anthropogenic variables are of interest in modelling (Seiferling, Proulx & Wirth, 2014; Gallardo, Zieritz & Aldridge, 2015), especially when working with weeds (Korres et al., 2016). Biotic variables have been increasing in the last few years, although they are not used very much yet, and, in this review, very few articles considered them. The target species (weeds) may not show the same suitability as forestry or non-weedy species, where biotic variables can provide a better fit to the niche (Woerner & Hackney, 1997; Meier et al., 2010).

The databases for obtaining the information were selected according to the extent of the study area, and the availability of data per country. In general, big countries and a group of countries or continents were the areas most studied, so global databases were frequently used. Local studies, on the other hand, were drafted compiling data from national databases. This argument is in accordance with the resolution, i.e., lower resolution at larger areas and vice versa. The WorldClim project (https://www.worldclim.org) was the most extensive database considered, especially when modelling all around the world, bioclimatic variables and altitude. Other topographic variables, like aspect, were computed with GIS from the latter. In some cases, the Shuttle Radar Topographic Mission (SRTM) was also resorted to, and as for the pedology, the most popular work used was SoilGrids (https://soilgrids.org). Human footprint variables were a resource that the authors did not discard and impacts and disturbances were procured primarily from the Socioeconomic Data and Applications Centre (SEDAC), the History Database of the Global Environment (HYDE), and the Range and Watershed Management Organization (FRWMO). The EarthEnv dataset and the Global Land Cover Facility (GLCF) were also taken into consideration for biotic variables. The more local the study area, the more local the sources were.

Occurrence data for dependent variables were normally obtained from the Global Biodiversity Information Facility portal (www.gbif.org) and sometimes enriched with unpublished data by the authors of the reviewed articles.

Spanning and location of the studies

Nine articles were focused on the world, 13 on a continental scale (Two of them considering two different continents), 21 on countries, 10 on regions, and six at local divisions. Specifically, three references belonged to Africa, 11 to America, 24 to Asia, 11 to Europe, and three to Oceania (Fig. 1). Per country, China published the highest number of the reviewed articles, followed by the USA and India. Brazil in South America and South Africa in Africa were also remarkable on their respective continents.

To model our planet implies working with a lower resolution due to data weight and computers’ performance. This is consistent with other scientific research (Guisan et al., 2007; Franklin & Miller, 2010), and a balanced occurrence and explanatory data in each part of the globe are crucial. Resolution changes can influence the perception and behaviour of the models (Lantschner, de la Vega & Corley, 2019). The smaller study area, however, produces higher resolutions and there are likely to be more accurate presence data for the species.

Regarding location, there was a clear difference in the number of published articles between those administrative areas with facilities to obtaining funding and those that did not invest in the same way. Per continent, Europe is noteworthy, whereas per country, China, the USA, and India are at the forefront. On the other hand, Australia, Russia, South America, and Africa are behind (although Brazil and South Africa are faintly highlighted, respectively). In most cases, differences in their investment to fight climate change are notorious between developing and developed countries (Chinowsky et al., 2011). On a smaller scale, Bhutan, Nepal, and South Korea should be mentioned.

Model algorithm and validation

The most popular algorithm used was Maximum Entropy (MaxEnt) with thirty-seven out of 59 articles exclusively applying this software. In addition, nine studies considered several algorithms; MaxEnt was also present in five of them, so that the latte was employed in 71.2% of the cases. Random forest (RF) followed MaxEnt, although at a considerably greater distance. This algorithm was used in six articles; only in one case was it used alone. By contrast, CLIMEX was found in four articles by itself, whereas once it was used together with other algorithms. Other algorithms like artificial neural network (ANN), classification tree analysis (CTA), generalized boosting model (GBM), generalized linear models (GLM), multiple adaptive regression splines (MARS), logistic regression (LOGIT), etc. were used less (Table 2).

Among the various model validation methods, Area Under the Curve (AUC) stood out significantly. Forty-five articles took it into consideration, of which 28 only considered AUC. The true skill statistic (TSS) was also largely used in the reviewed works: 15 times, of which only once was considered without any other validation method. Besides AUC and TSS, two articles also used a third method called Kappa. Bayesian information criterion (BIC) and root mean squared error (RMSE), among others, were also methods applied.

As can be seen from the results, most of the authors agreed on using MaxEnt and AUC for their respective works, which is consistent with other SDM studies not on weeds. Currently, both MaxEnt and AUC seem to be robust methodologies in this field (Cobben et al., 2015; Cao et al., 2016). In general, AUC gave accurate values according to Swets (1988), who ranked those values as follows: low accuracy (0.5–0.7), potentially useful (0.7–0.9), and high accuracy (>0.9). Wan, Wang & Yu (2019) established 0.7 as a relevant threshold value. The fact that TSS is normally used together with AUC may be due to the weakness of the former. Thus, other alternatives are being investigated to assess SDMs, such as the odds ratio skill score (ORSS) and the Symmetric Extremal Dependence Index (SEDI) (Wunderlich et al., 2019).

MaxEnt is a software programme frequently used worldwide to estimate the potential distribution of invasive plants (Mainali et al., 2015; Merow, Smith & Silander, 2013). This algorithm can perform accurate predictions with few occurrence data (Pearson et al., 2007), although a good management of the dependent and independent or explanatory variables, together with an appropriate election of the resolution, improves MaxEnt performance (Wan, Wang & Yu, 2019). The striking dominance of MaxEnt in plant models contrasts with other groups reviewed, including plants, animals, and pathogens; CLIMEX is commonly used in pest risk assessment (Lantschner, de la Vega & Corley, 2019). However, it has been found that the best statistical SDMs and environmental data substantially outperformed CLIMEX and ecophysiological data when a pest was modelled (Early et al., 2022). The popularity of MaxEnt led to investigating it more profoundly, allowing improvements to be made when managing data and settings. MaxEnt’s strength is that it is easy to use (Mainali et al., 2015; Merow, Smith & Silander, 2013), producing robust outputs when data are scarce, irregularly sampled, or show minor location errors (Elith et al., 2006; Pearson et al., 2007). Moreover, when the study area so allows, spatial filtering can minimize omission errors (false negatives) and commission errors (false positives) (Kramer-Schadt et al., 2013). The number of occurrences, the number of environmental variables, and the spatial scales can also be managed to enhance the performance of MaxEnt. Choosing an appropriate number of occurrences and variables is crucial, whereas using a spatial resolution of 5.0 arc-min to model invasive plants globally seems suitable (Wan, Wang & Yu, 2019).

The use of ensemble models has been encouraged because niche-based modelling algorithms can then be compared, and they can help to understand how their approaches perform on their own. In the current review, ensemble models were used in some cases (nine out of 59 reviewed articles), although only the final consensus maps were provided. In our opinion, this information should not be wasted as it would motivate the authors to publish and discuss their results separately by algorithm.