Research on Using Ensemble Models to Assess the Impacts of Climate Change on Agriculture Production: A Review

Abstract

The use of artificial intelligence tools in agriculture is growing. In particular, the use of ensemble models. However, there are still few reviews on the use of these models in the study of the impacts of climate change on agriculture. Therefore, the aim of this article is to review the use of such models and perform three key tasks: (1) identify topics in which ensemble models are used, (2) determine the most widely applied model and the predominant crops and regions, and (3) explore future applications and challenges. As a result, it was noted that the first studies, dating back to 2011, applied ensemble models to model invasive species. Since then, research has focused on changes in temperature and precipitation, with at least one study published every year. The most cited studies have dealt with land use classification, emphasizing its relevance to climate change studies. Notably, studies on carbon storage in soil and its capacity to remove CO₂ from the atmosphere have become increasingly relevant. This analysis highlights the growing importance of ensemble models in climate-related agricultural research, outlining trends and key areas for future exploration.

1. Introduction

The intensification of the effects of climate change has the potential to result in an increase in various impacts, including more intense rainfall [1] and more pronounced droughts [2]. These changes in climate result in substantial changes in agricultural production, as evidenced by changes in the forecast of expected crop yields [3]. Accurate crop forecasting is a factor that guarantees food security in different countries [4]. A variety of data sources, including government statistics and satellite imagery, in conjunction with diverse statistical models, can be employed to formulate this forecast.

Machine learning techniques have been used in different agricultural applications such as predicting nutrients in leaves [5] and predicting the yield of crops such as wheat [6]. Ensemble models, including Random Forest (RF), have seen a marked increase in utilization for predicting variables such as leaf area index [7]. Predicting this water content allows for efficient crop management, avoiding water stress, which is extremely important for corn as it is a drought-sensitive crop [8].

Ensemble models were originally developed to reduce the variance of automated decision-making systems, thereby increasing their accuracy [9]. This type of learning is a generic term for a set of methods that combine several inducers to make a decision [10]. The main definition of ensemble learning is that “by combining several models, the errors of a single inducer are likely to be compensated by other inducers and, as a result, the overall prediction performance of the ensemble would be better than that of a single inducer” [10].

Old statistical models and methods based on past data are often insufficient for predicting climate change [3]. Traditional statistical methods generally assume a linear correlation between the dependent variable and influencing variables. However, this assumption does not accurately reflect nonlinear relationships [4]. In such cases, it is necessary to use large data sets to identify complex correlations between variables and make accurate predictions. For this reason, these models have received more attention, and solutions that use them to create more accurate climate change models in agriculture have been proposed.

Machine learning methods outperform statistical methods in different applications [6]. In addition, models such as Adaboost show greater spatial adaptability compared to other approaches [6], which, in regard to climate change, increases the accuracy of the predictability of the impact on agricultural production. The possibility of using and contributing different data sources to these models also demonstrates the importance of understanding the use of these models in the subject matter discussed here.

The utilization of these learning models has exhibited a marked increase. A search of the Web of Science database on 18 April 2025 using the term “ensemble model*” returned more than 8000 articles. It is noteworthy that a significant proportion of these results, more than half in fact, were published in the period between 2021 and 2025. However, there is a paucity of studies that synthesize the content of these studies on the application of ensemble models to the issue of climate change in agriculture. A few reviews have looked at the use of machine learning or ensemble models in areas related to agriculture. However, no review specifically mentions climate change as a key word. Examples of reviews in this area include work on determining organic carbon in soil [11] and its use by scientists studying water issues [12].

These studies raise significant questions regarding the models employed, among other issues. Which subjects related to climate change in agriculture have been utilized? Is the primary focus of these models to predict crop yields, or to anticipate and mitigate natural disasters such as droughts and floods?

The aim of this article is, therefore, to review the use of ensemble models in the field of climate change in agriculture. To this end, using the Web of Science database, a search was conducted for articles on the subject and a review of the results was carried out to assess how these models have been applied to the aforementioned subject.

The review was guided by three key research questions (KRs), which served as the basis for the analysis of the selected studies. The first KR deals with the topics in which ensemble models have been used for agricultural studies in the context of climate change.

Of the models under consideration, which has the broadest application? Which countries and crops are most frequently addressed in the existing literature (KR2)?

How have the applications of these models changed in recent years? Which sustainable development goals have research been concerned with? In addition, we seek to determine the current constraints in pursuing a comprehensive understanding of the subject (KR3).

To obtain the studies used in the review, searches were conducted in the scientific database Web of Science (Clarivate Analytics, London, United Kingdom). Search queries included terms such as (“Ensemble model*” OR “RF*” OR Adaboost) AND (agricultur* OR crop* OR maize OR soybean OR sugarcane OR rice OR coffee OR wheat OR vine* OR bean* OR corn OR pastur*) AND “climate chang*” to titles, abstracts and keywords. The search was conducted on 20 March 2025, and results were filtered to include studies published up to 2024. Conference abstracts and reviews were excluded. For the analysis, 999 articles were obtained.

Using the bibliometrix library in RStudio (RStudio, version 4.2.2, Boston, MA, USA), data on publications were obtained that concerned the year, most cited articles, most cited authors, production by country, journals with the highest production, and most cited keywords, as well as trend topics. A Sankey diagram was also obtained, which illustrated the evolution of interest in the topics within the database using the bibliometrix library. The utilization of the diagram facilitated the determination of the evolution of subjects of interest over time. Consequently, it was possible to ascertain the subjects that have attracted the most attention in recent years. This information can then be used to discuss future trends. To obtain all these analyses, the keywords used by the authors when describing their articles were considered.

Analyzing the data for a review involves more than just collecting and processing it. It also involves using the information to map out an area of study, identify the most relevant works and main authors, and propose new avenues of research based on gaps in the literature. The bibliometrix library helps compile data, visualize information (e.g., variations over time, relationships between themes, and topics of interest), and produce graphs and other analyses relevant to the article.

Figure 1 presents a flowchart of the review screening process adopted in this study, based on standardized selection criteria from the Web of Science database.

2. Relevant Sections

Ensemble Models and Impacts of Climate Change in Agriculture

Decision-making based on data sets is nothing new in human societies, as we use them quite often—for example, the essence of democracy, where decisions are made based on the votes of a group of people, is in fact based on decision-making based on data sets [9]. Although the original purpose of using ensemble systems is like the reasons why we use collective decisions in our daily lives, i.e., to increase our confidence that we are making the right decision, there are many other explanations for the use of machine learning ensemble models. These include “increasing confidence estimation, feature selection, eliminating missing features, incremental learning from sequential data, merging data of heterogeneous types, learning in stationary environments and addressing unbalanced data problems, among others” [9]. Therefore, ensemble models are models in which the answers obtained come from different models, not just one, as in other machine learning techniques, which increases the confidence of the answer.

Preliminary research indicates that the utilization of ensemble models for the purpose of assessing, measuring, or predicting the impacts of climate change on agriculture has been in effect since 2011, the year in which the inaugural article in the database was published. The article, published by Lemke et al. [13], dealt with the modeling of an invasive species in a given region. The authors employed ensemble models to develop a model of the spread of the invasive species as a function of climatic conditions and land use.

As evidenced by the subsequent articles published in 2012, the issues related to climate change began to become clearer. The second [14] and the third article [15] addressed issues related to increases in precipitation and temperature, respectively. The theme of temperature would remain constant over the years. It is also important to note that after the first publication, all subsequent years had at least one article published on the subject (Figure 2).

Analyzing the areas of concentration of the papers highlighted the five areas with the most articles: Environmental Sciences (40.24% of the articles, n = 402), Remote Sensing (17.51%, n= 175), Geosciences (16.81%, n = 168), Imaging Science Photographic Technology (13.0%, n = 130), and Agronomy (11.31%, n = 113). As each article can relate to more than one area, the total sum is greater than 100%. The topics that were cited in 75% of the articles are shown in

Table 1. The topics most related to the articles.

Topics	Record Count	% of 997
Land Use	251	25.12
Yield	137	13.71
Temperature	133	13.31
Water	133	13.31
Carbon	102	10.21

.

When studying climate change, it is important to consider issues such as temperature, rainfall, and growing time. Even more so when studying how crops adapt to these changes [3]. Also, the topic of temperature appears very frequently in later years, a fact demonstrated by the occurrence of keywords (Figure 3). Of the total number of articles, 376 dealt with some aspects related to temperature. An example of an article on this subject is that published by Long et al. [16], which deals with soil moisture prediction using ensemble models. One of the data points used by the authors is the surface temperature, showing the importance of this data for predicting the amount of water available for crop development.

According to analyses of the impacts caused by climate change, the occurrence of extreme temperatures (together with storms and forest fires) accounts for 23% of agricultural productivity losses in Asia [17]. Even though elevated temperatures can exert disparate effects on diverse crop species [18], extreme temperatures have the potential to influence various phases of plant development. For corn, for example, temperatures above 35 °C can make pollination unfeasible [19]. Other crops, such as rice and sorghum, have reduced fertility when temperatures rise above 33 °C. This temperature causes reduced pollination, as well as lower yields and grain formation, among other impacts [20,21].

Regarding the issue of soil temperature, some authors have proposed models that anticipate an increase of up to 3 °C [22] in the next years. This increase in temperature can be mitigated by the presence of greater moisture levels in the soil. As posited by Zuo et al. [23], the maintenance of contemporary levels of soil moisture has the potential to curtail global warming by 32.9% under low-emission scenarios. Maraun et al. [24] posit that the consequences of extreme heat can be amplified or mitigated depending on the amount of water present in the soil. Therefore, the utilization of ensemble models to predict these quantities can assist in agricultural management and in determining the impacts of warming in future scenarios.

Despite the prominence of temperature-related articles, the most cited topic, as well as the most cited articles on the subject, dealt with land use classification. The most cited article (1689 citations), by Tang and Huang [25], used ensemble models to classify land cover over a 30-year period in China. According to these authors, “Land cover (LC) determines the energy exchange, water and carbon cycle between Earth’s spheres”. Data on land cover change is extremely important for environmental and climate change studies [26]. A study on the classification of tree species and different crops also featured prominently, being the second most cited article [27]. This type of use of ensemble models goes in the same direction as the most cited article, seeking to classify land uses to help with soil management, water use, and obtaining agricultural statistics. As mentioned above, soil and water management are strongly related to climate change and changes in water availability and crop production capacity.

Changes in land use are of fundamental importance to climate change. In 2005, the U.S. National Research Council [28] recommended including land use and land cover processes as important climate forces in climate change models. This recommendation was made because these processes have significant local, regional, and global climate implications. In some cases, the climate’s response to land use and land cover changes may exceed the impact of increased greenhouse gases [29]. For example, a decrease in vegetation cover can lead to an increase in soil temperature, modifying local climate dynamics. This is why studies on changes in land use are so prominent in the results presented here.

Human activities significantly alter the amount of organic carbon in soil [30]. Soils are very important for removing carbon from the atmosphere, as they contain a carbon stock that is about twice as large as that of the atmosphere and about three times larger than that of vegetation. Therefore, changes in this stock can have a major impact on CO₂ concentrations in the atmosphere [31].

As an indicator of ecosystem health, the measurement of soil organic carbon (SOC) was also highlighted in the articles published on the use of ensemble models in the context of climate change. The study by Were et al. [32], with the objective of measuring the amount of soil organic carbon (SOC) in a given landscape, was the third most cited study.

Therefore, ensemble models are widely used in applications related to climate change in agriculture. These applications are not limited to assessing the impact of rising temperatures on agriculture, but also include land use classification, soil moisture assessment, and carbon content. Thus, these models are diverse and can be used for different purposes, despite the predominance of studies related to temperature. Temperature influences various processes of plant development, such as growth. There is an “ideal” temperature range within which development rates increase almost linearly with temperature [33]. However, increases above this range lead to plant stress, which consequently results in yield losses and other harmful processes. It is because of the influence of temperature on plant development that there is great interest in using ensemble models to predict temperature.

It can be seen in Figure 4 that studies mentioning climate change are more closely related to the topic of productivity (green group). On the other hand, studies on land use change (red cluster) are more closely related to carbon stocks. RF models are related to topics such as vegetation classification (brown cluster), while temperature (blue cluster) is related to a variety of topics, all of which focus on crop yield (such as drought, precipitation, prediction, wheat, etc.).

3. Discussion

3.1. Most Used Models

Among the models mentioned, the one with the most references was RF, which appeared in 789 articles (78.97%). Lima et al. [11] also found that RF was the most widely used model in studies on determining soil organic carbon. This model was developed by Breiman [34]. According to this author, it is a model in which there is “a combination of tree predictors such that each tree depends on the values of a random vector sampled independently and with the same distribution for all trees in the forest”. In this model, the error depends on the strength of the individual trees in the forest and the correlation between them. Therefore, the response of each tree is considered together with the rest to determine the response of the model, which ultimately strengthens the response and reduces the error.

For example, Han et al. [35] used this model to predict wheat yield. In the results obtained in this work, the authors observed that RF was one of the models that obtained the best results, along with Support Vector Machine (SVM) and Gaussian Process Regression (GPR). Crop yield is one of the characteristics that is most affected by climate change; therefore, works with ensemble models focus on predicting this yield in the scenario of impacts caused by rising temperatures.

This model is also used to determine soil moisture [36]. Different authors report that this type of ensemble model highlights a strong relationship between soil surface variables and the amount of water [37]. Obtaining soil moisture values can help in adopting better irrigation practices [38], as well as predicting meteorological characteristics [39] and the occurrence of droughts and floods [40].

The learning method employed in RFs boasts several advantages [41], including the following:

-

Firstly, joint learning is a method that is known for its ability to avoid the phenomenon of overfitting.

-

Secondly, the bagging has been demonstrated to facilitate the effective operation of RFs when dealing with limited data sets.

-

Thirdly, the training can be administered concurrently.

-

Finally, the implementation of automatic feature selection is facilitated by the utilization of decision tree learning in RFs.

These characteristics explain the large-scale use of this model, with it being the most reported in the research carried out for this review. The model’s superior performance is attributable to its capacity to circumvent the pitfalls of overfitting, a notable advantage over competing models. Consequently, its utilization has become increasingly prevalent and favored among the authors included in the database under analysis.

The model’s ability to operate with limited data makes it a suitable candidate for the research in question. Research into the impacts of climate change on agriculture often lacks substantial data for analysis. The extended production time of certain crops, their seasonality, and the interference of climatic factors (as well as soil), among other factors, result in a reduction in data availability (or the avoidance of repeated data collection). This model is distinguished by its ability to obtain satisfactory results with a reduced data set, which may indicate its prevalence over other models.

Despite the widespread use of RF algorithms in studies on the impact of climate change on agriculture, other methods are also cited. One such algorithm is the SVM, a technique introduced by Vapnik as a kernel-based machine learning model for classification and regression tasks [42]. A notable benefit of its application is its capacity for generalization and its ability to discriminate. Ayalew et al. [43], for example, used this model to predicting the yield of different crops. According to the authors, as well as being highly accurate, this model provides good results for decision-making under the conditions of climate change.

The Extreme Gradient Boost (XGBoost) algorithm is another tool cited in the database consulted. Chen and Guestrin [44] proposed this model in 2016, which is more recent than the other algorithms mentioned here, such as RF. It can be used in both classification and regression problems, an example of which is the work carried out by Magidi et al. [45]. In this article, the authors classified land cover and observed that XGBoost performed better during the classification.

On the other hand, the boosting algorithm AdaBoost [46], which is also mentioned among the works, is used to adjust (by weighting) a sequence of weak predictors, such as small decision trees, in differentiated versions of the training data. The predictions of each predictor are then combined by a weighted sum (or voting in cases where the objective is classification) to produce the final prediction. Among the different applications of this algorithm, it has been used to model the price and production efficiency of beef cattle [47], as well as in work on predicting the leaf water content of crops such as corn [48].

One of the characteristics of boosting algorithms is that they focus on the error in each iteration and try to reduce this error by adding a new model [49]. Thus, to perform this combination of several models, the algorithm explicitly tries to find models that complement each other. The performance of each model influences the reinforcement of the new models [50]. This type of algorithm has been increasingly cited in the field of climate change in agriculture, showing that its use is likely to become a research trend in the coming years.

Because they were proposed in later years than older models, such as the RF, these models are still rarely used (and their participation in the works collected here is reduced). However, there has been an increase in the use of these models. In 2023, only 8 articles mentioned the use of XGboost in their titles, a figure that rose to 23 articles in 2024.

3.2. Countries and Crops

Continuing to answer KR2, the countries that have published the most research on the subject are China and the United States of America, with the latter accounting for 35.2% of the publications (

Table 2. The countries with the most publications on the use of ensemble models in climate change and agriculture research.

Country	Records	% of Articles
China	350	35.2
USA	112	11.3
India	52	5.2
Germany	44	4.4
France	30	3
Italy	28	2.8
Australia	24	2.4
Canada	23	2.3
Brazil	19	1.9
Spain	18	1.8

). These two countries are the largest emitters of greenhouse gases [51], which partly explains their interest in the research topic. However, other factors also apply, such as competition for the artificial intelligence market, as well as the concentration of universities and research centers in these two countries, among other factors.

One of the most common explanations for the low participation of developing countries in research on artificial intelligence or machine learning is their limited access to technology. However, “this is only true because the technology was developed assuming the conditions of developed countries and, therefore, current machine learning (ML) tools encode the infrastructure and cultural conditions of developed regions” [52]. But even with the challenges posed by infrastructure limitations, skill gaps, and other factors, there are a wealth of opportunities in several areas, including agriculture, that can benefit from the profound impact that AI could have on these nations [53].

Although there has been significant growth in recent years of scientific research in some countries, such as China, India, Brazil, South Korea, Turkey, and Iran, which went from a share in total world publications of 7% in 1996 to 27.8% in 2018 [54], there is still a large concentration of articles published in European countries and the USA. According to data from other studies, between 1990 and 2019, around 74% of publications in the main economic development journals were written by researchers not based in southern countries [55]. This fact is repeated in other areas, which explains the data discussed here. Access to funding, difficulties with writing in English, and a lack of structure and international partnerships may all explain these results.

Despite having a smaller share of the total number of articles (39.7%) analyzed here, developing countries have seen an increase in research in this area. In 2015, only 5 articles were published by these countries, while by 2024 the number had risen to 128 articles. However, this is still far below the total number of publications from developed countries. Developing countries have less capacity to adapt to the impacts of climate change than developed countries [56]. In addition, these countries will suffer more from the economic consequences of climate change than richer countries [57]. Therefore, expanding research on the use of ensemble models to predict the impacts of climate change on underdeveloped countries could be essential to help mitigate these impacts.

The preponderance of developed countries as the primary publishers of articles is also reflected in the occurrence of the main crops evaluated in these studies. While studies with wheat, corn, or rice in their title accounted for 167 articles, other crops (mainly from tropical regions) were less represented. For instance, a review of the titles of the papers revealed that only 15 papers contained the terms “sugar cane”, “cassava”, or “beans”. Of these, 12 were exclusively about sugar cane. This shows that only three crops (corn, rice, and especially wheat) account for most of the work analyzed.

Research on wheat has been carried out for a variety of purposes, including growth modeling [58] and predicting yields [59]. Among the most frequently cited articles on this subject is that by Feng et al. [60]. When modelling the impact of climate change on the yield of this crop, the authors observed that drought events during the grain filling and vegetative stages, as well as heat events just before anthesis, were the three most serious impacts that cause yield losses for wheat.

The prediction of crop yields is also highlighted in the articles published on corn [61,62]. However, work on modeling the impacts of climate change is also well cited. This is the case of the article published by Falconnier et al. [63], in which the authors observed that, in cases of low nitrogen fertilization, corn yields are more affected by excessive rainfall.

Due to concerns about the impact of climate change on crop yields, predicting this factor is a constant theme in the works analyzed here. For rice, the most cited articles also deal with this issue [4,64]. On a different subject, Bo et al. [65] carried out a study to assess the benefits of non-continuous flood irrigation in reducing greenhouse gas emissions in rice. Using an RF model, the authors observed that the non-continuous flood can be used in 76% of the mapped areas worldwide. This finding indicates that the implementation of this method would lead to a 47% reduction in the global warming potential (GWP) of methane (CH₄) and nitrous oxide (N₂O) emissions from rice production, or a 7% reduction in the total GWP. Additionally, this approach would result in a 25% reduction in irrigation water usage while maintaining yield levels.

Makowski et al. [66] carried out a quantitative synthesis of the effects of temperature, CO₂, rainfall, and adaptation on global crop productivity. Using ensemble models, these authors observed results showing that, for C3 crops (such as rice, soybeans, beans, and wheat), there will be yield reductions of around 2.4% for +1 °C. For plants such as corn, yield losses will amount to −10% for +4 °C.

Other authors, with fewer publications, have also analyzed production alternatives to mitigate climate change. For example, Han et al. [67] attempted to predict the primary production of a wheat and corn crop rotation system. The authors found that factors such as soil water content (SWC), evapotranspiration (ET), and leaf area index (LAI) largely explained the variation in production. In addition, drought and water deficit are the challenges the authors identified for this type of production.

As described in

Table 3. The performance of RF compared to other models.

Article	R2 RF	R2 Others	RMSE RF	RMSE Others
[30]	0.8	0.8	<750 kg ha−1	<750 kg ha−1
[68]	0.983	0.903	0.297	0.727
[69]	0.817	0.716	129.9	152.7
[70]	0.64	0.56	-	-

, the RF performs similarly to, or in most cases, better than other models in predicting the yield of different crops. These results also help to explain the predominant use of this model in the articles cited, as well as the increased use of ensemble models in applications related to the area studied here.

3.3. Changes in Applications and Prospects

As demonstrated in Figure 5, the terminology employed in relation to the subject under discussion has undergone a gradual transformation over time. Sankey diagrams were developed to represent the flow of energy or materials in various networks and procedures and are now commonly utilized for this purpose. Furthermore, they also represent flows and associations of information and their transitions in quantitative detail. At each node, the number of input weights is equivalent to their respective output consequences. Sankey diagrams are a visual representation of information transformations over time, and communications between each node can be explored [71].

Even with the passage of time, the terms that refer to the RF model remain among the most cited. However, some significant thematic changes are beginning to be observed. In the period between 2011 and 2016, the detection of changes in land use was highly relevant. However, in the following period (2017–2021), research on this topic used the same model (RF) and other topics began to appear, such as studies on drought, biodiversity, and others.

In the last period, 2022 to 2025, research using RF (for classification and regression) began to address issues such as soil organic carbon and carbon sequestration. This change occurred because, in recent years, there has been growing interest in forms of cultivation and management that help to sequester carbon from the atmosphere and not just mitigate the effects of climate change. In this regard, there is research such as that conducted by Meng et al. [63]. The present authors posit that, for a certain category of soil, there has been a diminution in the quantity of soil organic carbon (SOC). Consequently, monitoring of this content is imperative to ascertain the adoption of more sustainable practices, as well as to identify areas where actions are necessary to mitigate this increase in carbon loss.

Mahmood et al. [72] also argue that SOC is used to assess soil health, “indicating the agricultural productivity potential of soils and correlating with other functions such as water capacity and soil biodiversity”. In addition, these stocks are increasingly recognized for their importance in climate change mitigation strategies. Using RF, these authors developed a model to predict the amount of SOC. They observed that soil pH and the type of cover are the most important factors for this prediction.

The utilization of ensemble models for the purpose of simulating and observing land use has been a constant theme since the publication of the initial papers about climate change. Considering the repercussions engendered by agricultural practices and the pursuit of strategies to mitigate their environmental impact, the implementation of diversified land use has emerged as a pivotal solution. This approach aims to curtail the deleterious effects stemming from agricultural activities by promoting multifaceted utilization of land resources.

The concept of land use multifunctionality is a critical term in the study of land use functions, with the potential to enhance comprehension and facilitate the realization of sustainable development [73]. The concept of multifunctionality can be identified across various spatial scales, including the region, landscape, and grid level. This underscores the importance of ensemble models in this field. These models facilitate more precise identification and classification of land use at the site under assessment. Therefore, the use of these models is crucial to measuring the multifunctionality of the site being assessed.

According to Bouma [74], multifunctionality of land use is essential for achieving some of the UN’s (United Nations) sustainable development goals (SDGs), such as good groundwater and surface water quality (SDG6), low greenhouse gas emissions and carbon capture (SDG13), and biodiversity and nature conservation (SDG15). In addition, multifunctionality is also related to food security. This theme appears as one of the trend topics related to the research discussed here (Figure 6).

Food security can be defined as the ability to provide food under all contingencies [75]. From a climate change perspective, a 3 °C increase in global temperature could lead to massive losses in the yield of different crops [76], thus increasing the instability of food production and causing food insecurity. Some measures, such as reducing crop losses, changing eating habits, diversifying crops, and stabilizing production in the face of climate change, among others, are seen as necessary to mitigate the effects of climate change and guarantee food security [77].

The use of ensemble models can help to model future scenarios of susceptibility to the effects of climate change. An example of this is the work published by da Silva et al. [78], which modeled land use change because of climate change effects. They found that for one region of Brazil, croplands may lose approximately 8% of their suitable area, while pastures are expected to expand by up to 30%. These results serve as indicators for decision-making and the adoption of public policies to minimize these effects.

Cheng et al. [79] also used ensemble models to determine which food production indicators are most closely related to food security in China. A study of this type was also carried out by Gyamerah et al. [80] who determined, through ensemble modeling, that increases in temperature and rainfall could favor corn production in Ghana and, thus, guarantee greater food security. However, these increases must be contained within a limit so that they do not do more harm than good. The topic of food security appeared more frequently after 2020, but should remain as a topic of interest within this theme for years to come.

Other SDGs also featured prominently in the search analyzed here (

Table 4. The five sustainable development goals most related to the articles.

Sustainable Development Goals	Record Count	% of Articles
13 Climate Action	807	80.94%
15 Life on Land	530	53.15%
14 Life Below Water	508	50.95%
02 Zero Hunger	282	28.28%
06 Clean Water and Sanitation	281	28.18%

). With this classification, obtained from the Web of Science database, Climate Action is more prominent due to the keywords used in the search. However, the Zero Hunger goal, which is related to the topic of food security, and goal 14 (Life Below Water) also stand out.

As each article can be related to more than one objective, the total value is higher than the base number of articles analyzed, 997. The SGDs assessed also relate to the trend themes; in particular, the trends related to crop yields correlate with the Zero Hunger and Water Quality goals. Increasing crop yields not only contribute to greater food security, but also reduce the need to use resources and land for cultivation. As such, they have an impact on these United Nations goals. Due to the relevance and urgency of these issues, they should continue to be highlighted in the coming years through the application of ensemble models that help to increase crop yields and model land use to adopt more sustainable practices.

4. Conclusions

Through the research carried out here, it can be concluded that ensemble models have a wide range of potential applications in research into the impacts of climate change on agriculture. Although most of the papers focus on assessing crop yields and classifying land use, articles on modeling soil’s organic carbon content, assessing susceptibility to climate change, modeling future scenarios, and obtaining plant physiological characteristics have also been published.

Among the models used, RF was the most cited due to some of its characteristics, such as its ability to deal with a small database and its low amount of overfitting. Some other models, such as Adaboost and XGBoosting, have also featured more prominently in more recent articles.

There is a concentration of articles published on wheat, corn, and rice. This is also due to the greater number of articles published by developed countries, which leads to the under-representation of papers on crops that are grown more in underdeveloped countries (such as sugar cane and cassava).

Topics more closely related to mitigating the impact or influence of agriculture on climate change have been more prevalent in recent publications. Assessments of more sustainable practices, modeling of soil and atmospheric carbon, and other factors have appeared in both recent topics and trend topics. Thus, such applications of ensemble models should become more common in future work, demonstrating that the current phase of research is more focused on reducing the impact of agriculture on climate change than on studying the effects of these changes on agricultural production.

5. Future Directions

Despite the potential application of these models, the articles analyzed showed certain disadvantages. The necessity of substantial computing capacity for data processing, the elevated energy costs of computer systems, and the paucity of data availability, among other factors, impede the broader implementation of these models. This could have ramifications for their use in decision-making or in supporting public policies to mitigate climate change in agriculture.

This phenomenon is especially evident in developing countries, which, despite their substantial agricultural output (e.g., Brazil, Argentina, and Mexico), exhibit a paucity of publications on the subject. The implementation of policies designed to enhance the computational infrastructure and provide financial support for research in this domain could potentially contribute to an increase in the number of publications from these countries. Furthermore, the necessity for enhanced international collaboration on this matter must be emphasized.

The ensuing results illustrate the efficacy of ensemble models in addressing a range of subjects pertaining to climate change in agriculture. However, given the utilization of a single database (Web of Science), it is imperative to expand the search to other sources to increase the number of articles included in the analysis. The studies also found minimal mention of public policies or decision-making by legislators and public projects, a topic that requires further study.