Trends in Agroforestry Research from 1993 to 2022: A Topic Model Using Latent Dirichlet Allocation and HJ-Biplot

Abstract

Background: There is an immense debate about the factors that could limit the adoption of agroforestry systems. However, one of the most important is the generation of scientific information that supports the viability and benefits of the proposed techniques. Statistical analysis: This work used the Latent Dirichlet Allocation (LDA) modeling method to identify and interpret scientific information on topics in relation to existing categories in a set of documents. It also used the HJ-Biplot method to determine the relationship between the analyzed topics, taking into consideration the years under study. Results: A review of the literature was conducted in this study and a total of 9794 abstracts of scientific articles published between 1993 and 2022 were obtained. The United States, India, Brazil, the United Kingdom, and Germany were the five countries that published the largest number of studies about agroforestry, particularly soil organic carbon, which was the most studied case. The five more frequently studied topics were: soil organic carbon, adoption of agroforestry practices, biodiversity, climatic change global policies, and carbon and climatic change. Conclusion: the LDA and HJ-Biplot statistical methods are useful tools for determining topicality in text analysis in agroforestry and related topics.

1. Introduction

At present, several investigations have shown that there is a high degree of pollution and environmental degradation [1,2,3,4,5,6,7,8,9,10,11] in ecosystems, mainly generated by human activities that are focused on meeting the needs of a constantly growing population [12]. Within this conglomerate of activities, it is possible to identify the relentless use of pesticides and chemical synthesis inputs in the agricultural sector, the excessive use of land, the deforestation of forests, the constant change of land use for agricultural exploitation, and the constant emissions generated by the industry that considerably strengthen the greenhouse effect, among others.

Therefore, society has seen the need to generate mechanisms that help mitigate these undesirable effects, leading the world to hold events such as the Rio de Janeiro summit (or the United Nations Conference on Environment and Development) (1992), the Program or Agenda 21, the agreement on climate change and the agreement on biological diversity, the declaration on relative principles of the forest, the World Conference on Social Development (1995), the United Nations Framework Convention on Climate Change (1997), the Millennium Summit in Geneva (2000), the United Nations Conference on Sustainable Development, Rio +10 (2002), and the annual United Nations Conference on Climate Change (1995), among others, whose objectives are to try, through holistic approaches, to evaluate possible mitigation actions.

As a result of these events, based on scientific research, it has been established that the greatest threat in the medium term is the accumulation of greenhouse gases generated during human activities. In this aspect, the gas that causes the greatest concern, due to its abundance, is CO₂ [13,14], which can be stored by forests [15,16,17,18,19,20,21,22] or in the structure of trees in production systems [23,24,25,26,27].

In relation to this concept, and in order to mitigate unwanted effects, the inclusion of trees in productive systems or agroforestry techniques is proposed as an environmentally friendly alternative. While agroforestry research has provided the necessary scientific foundations for it to be considered a sustainable practice, there is still much to be done to efficiently adopt it by producers through extension and decision makers.

Within this order of ideas, there is an immense debate about the factors that could limit the adoption of agroforestry systems; many authors attribute this fact to the uniqueness and complexity of the technology [28]. On this subject, Ref. [29], after reviewing various studies regarding the adoption of agroforestry technology in 21 countries, found that preferences and resource endowments are the most frequent factors. However, it is undeniable that one of the elements that influences the adoption of production systems is the prior generation of scientific information [28,30,31] that supports the feasibility and benefits of the proposed techniques.

In any case, it is important to highlight that the development of research on agroforestry is being highly promoted by governmental and nongovernmental extension agencies that work together with farmers, which is an encouraging forecast for the future. However, the research agenda, until now, has given high priority to some issues and neglected others, which together could generate technologies for the benefit of the earth and its users [32].

In this work, we present a Latent Dirichlet Allocation (LDA) topic modeling method to identify and interpret scientific topics relating to existing topics or categories in a set of documents [33]. Thus, from a collection of documents called the corpus, LDA, based on Bayesian models, generates a probabilistic extension of the latent semantic analysis [33,34]. In addition, the LDA assumes an a priori sparse Dirichlet distribution on the topics in the document using Gibbs sampling [35]. Gibbs is able to determine the topic probabilities of the documents that mix various topics in different proportions. On the other hand, the HJ-Biplot method [36] is proposed in this work to obtain a more precise data evaluation, revealing the existing relationships between the analyzed data.

2. Materials and Methods

The procedure for the topic modeling analysis through Latent Dirichlet Allocation [37] was divided into four stages: (1) search and collection of articles; (2) preprocessing; (3) construction of the LDA model; (4) labeling of topics (Figure 1).

2.1. Data Collection

The search for articles was carried out through Scopus, utilizing the following query: “TITLE-ABS-KEY (agroforestry) AND (EXCLUDE (PUBYEAR, 2024) OR EXCLUDE (PUBYEAR, 2023)) AND (LIMIT-TO (DOCTYPE, “ar”)) AND (LIMIT-TO (LANGUAGE, “English”))”. We decided to select this database since it is one of the most used by researchers [39]. The inclusion criteria focused on selecting research articles and reviews published in the English language between 1993 and December 2022. The search was conducted on 10 March 2023.

The preliminary database with the documents obtained after executing the search query contained 9966 documents; this initial sample was subjected to a filtering process, where documents that were repeated or misclassified or that contained no abstract were eliminated. The final sample obtained consisted of 9794 documents.

The procedure for the identification of topics through LDA was divided into three stages: (i) preprocessing, (ii) creation of the LDA model, and (iii) labeling topics.

2.2. Preprocessing Texts

Data processing in this part of the study was carried out using LDAShiny [40], an open-source package of the R programming language [41], which contains the development of a tool that provides a web-based graphical user interface to perform a review of the scientific literature under the Bayesian approach of Latent Dirichlet Allocation (LDA) and machine learning algorithms.

To increase the coherence of the topics, each abstract was tokenized using bigrams, which are the combination of consecutive unigrams. Although this process seems trivial, since the text is downloaded to the computer in readable format, it must be converted to lowercase and all punctuation marks, dashes, brackets, numbers, space blanks, and other characters removed. In addition, a standard list of words called “stopwords” was identified and eliminated, since their main function is to make a sentence grammatically correct (i.e., articles and prepositions).

2.3. Creation Model Latent Dirichlet Allocation

Topic models are latent variable models of documents that use correlations between words and latent semantic themes in a collection of documents [42]. This definition assumes that the expected number of topics, k, (i.e., latent variables) must be established a priori. Thus, selecting the right number of topics for a given collection of articles is not trivial. This challenge has been approached in a variety of ways, with the goal of striking a balance between the requirement for a large number of topics to cover all of the documents in the collection and the need for a small number of topics to ensure that the findings are intelligible. Simulations were run with k ranging from 5 to 50 and an inference process, Gibbs sampling [35], with 500 iterations was used. A topic coherence measure [43] was used to determine the quality of the LDA model, which is a measure of a topic model from the standpoint of human interpretability and is considered a more suitable measure than computational metrics such as perplexity [44].

2.4. Labeling Topics

First, a naive labeling algorithm provided by the package textmineR [45] based on bigrams was employed (these naive labels are based on P(bi-gram|topic) − P(bi-gram)). However, because these algorithms have a very limited capacity for understanding the latent meanings of human language, we decided to use manual labeling, which is considered a standard in topic modeling [46]. Two agroforestry professionals, with more than ten years of experience, manually identified the topics using two sources of information: the most common word lists (most likely) and a sample of the titles. Then, they summarized the 10 most-loaded articles.

2.5. Quantitative Indices Used to Analyze the Trend of Topics

We used some quantitative indices proposed by Xiong et al. [38], which were obtained by adding document–topic and topic–word distributions, in order to make the results and findings clear. The description of the indexes is as follows: the distribution of topics over time is obtained by:

$$\left({\theta }_k^y\right)=\frac{{\mathsf{\Sigma }}_{m\in y}{\theta }_{mk}}{n^y}$$

(1)

where $m\mathsf{ϵ}y$ represents articles published in a given year, ${\theta }_{mk}$ represents the proportion of the k-th topic in each item, and $n^y$ represents the total number of articles published in the year [38].

We used simple regression slopes for each topic, with the year as a dependent variable and the fraction of topics in the corresponding year as the response variable, to make it easier to characterize the topics in terms of their tendency. [33]. The slopes obtained for each topic were classified as positive or negative at a level of statistical significance of 0.01 and were classified as positive or negative trends, respectively.

Additionally, the VOSviewer software [47] was used to generate the co-authorship network, using countries as the unit of analysis, and employing the full counting method. Only countries with at least 20 documents were considered.

2.6. HJ-Biplot

Finally, the HJ-Biplot [36] was used. It is a multivariate statistical technique that allows a graphic representation of the information contained in the rows (individuals) and columns (variables) of a matrix of data [36]. This technique was chosen since it offers a more precise data evaluation, highlighting the relationships between the parts, years, and topics. For this analysis, the Multbiplot software [48] was used, which allowed us to have a fast and easy way to incorporate our tables from a *.xls format.

3. Results

The summary generated included basic statistics about the analyzed dataset; it is presented in

Table 1. Statistics about the analyzed dataset.

Description	Results
Main information about data
Timespan	1993:2022
Sources (journals, books, etc.)	1564
Documents	9794
Annual growth rate %	7.63
Document average age	9.8
Average citations per doc	22.48
References	1
Document contents
Keywords plus (ID)	19,986
Author’s keywords (DE)	20,415
Authors
Authors	25,174
Authors of single-authored docs	654
Authors collaboration
Single-authored docs	771
Co-authors per doc	4.41
International co-authorships %	37.99
Document types
Article	9173
Review	621

. Documents stemmed from 1564 different journals and were published over the course of three decades. A total of 25,174 authors were involved in the scientific production on agroforestry. Among the papers, 771 were single-authored papers, whereas the overall collaboration index of the sample was 2.61.

Table 2. Most influential sources of the documents analyzed.

Sources	Articles
Agroforestry Systems	1549
Agriculture, Ecosystems and Environment	307
Sustainability (Switzerland)	197
Forest Ecology and Management	188
Forests	149
Plant and Soil	118
Forests Trees and Livelihoods	112
Science of the Total Environment	82
Agricultural Systems	81
Land Use Policy	76
Biodiversitas	75
Small-Scale Forestry	75
Journal of Sustainable Forestry	74
Land	72
Biodiversity and Conservation	68
Land Degradation and Development	64
Journal of Environmental Management	62
Plos One	60
Journal of Forestry Research	56
Range Management and Agroforestry	53
Agricultural and Forest Meteorology	52
Agronomy	50
Current Science	50
Environmental Management	50
Agronomy for Sustainable Development	49

shows the most influential journals in terms of article count. These journals were distributed in different subject areas, such as environmental science, environmental stewardship, ecosystem services, economic investment for research, policies on sustainable land use, and the participation of entities, specialized policy, etc. The journal with the most published articles was Agroforestry Systems, which contributed about 16% of the documents analyzed.

In terms of the annual number of publications in the evaluated period, we observed an annual growth rate of 7.63%. A low oscillation in the number of publications could be observed in the first decade, while in the last year the number of publications diverged from an exponential trend (Figure 2).

Figure 3 represents the countries that have published the most research on agroforestry in the world. The United States (1817), India (1114), Brazil (947), the United Kingdom (809), Germany (807), and France (638) were the five countries that appeared as the main sources of high-impact research available in the world scientific literature. They were followed by China (629), Indonesia (591), Kenya (587), Australia (464), Canada (456), and finally Spain (451), with a lower percentage of publications. Other countries did not show the same frequency of publications or the same impact.

Figure 4 shows the network of international coauthor relationships among 71 countries, the largest set connected. There are seven colors in Figure 4, indicating that these 71 countries are grouped into five clusters. There is a stronger cooperative relationship between countries in the same cluster than between countries in different clusters. Of course, this does not mean that there is no cooperation between countries in different clusters, but that there may be some common research topics among the countries of the same cluster, which makes their cooperation closer. More frequent connections (represented by the thickness of the line) could be observed between the United States and Brazil.

Topic Modeling Analysis

Figure 5 shows the coherence score for all LDA models that were examined. According to the findings, the LDA model with the best coherence score had 30 themes (k = 30).

The topic distribution from the document ${\theta }_m$ was added to calculate the average probability ${\theta }_k^y$ of all the articles published in a given year and to identify the trends (Figure 6). We found that the probabilities of some topics increased gradually over time (color red); these topics were t_3 (raw materials production), t_11 (shade tree), t_12 (land cover), t_16 (adoption of agroforestry practices), t_17 (remote sensors), t_19 (fungal communities), t_21 (carbon and climatic change), t_22 (ecosystem services), t_24 (secondary forests), t_27 (biodiversity), t_28 (soil organic carbon), and t_29 (cork oak). Topics that had a decreasing behavior were (color blue): t_1 (tree species), t_2 (leaf litter), t_10 (alley cropping), t_14 (silvopastoral systems), t_15 (fine root), t_20 (soil water), t_23 (soil fertility), t_26 (production systems), and t_30 (aboveground biomass). The topics where there was no observed trend (color black) were: t_4 (intercrop systems), t_5 (rubber plantations), t_6 (food security), t_7 (fruit tree), t_8 (local knowledge), t_9 (agroforestry systems), t_13 (climatic change global policies), and t_25 (plant species).

Our study found six groups through cluster analysis (Figure 7). The topics: t_16 (adoption of agroforestry practices), t_6 (food security), t_13 (climatic change global policies), t_28 (soil organic carbon), and t_27 (biodiversity) of group six obtained higher probability proportions for the years 1993 and 2022; the topics t_10 (alley cropping), t_23 (soil fertility), t_15 (fine root), t_1 (tree species), and t_30 (aboveground biomass) of group five obtained higher probability proportions for the years 1993 and 2008. The remaining topics had low probability statistics during the years of study.

In

Table 3. Frequent topics.

TOPIC	LAB. ALGORIT	LABEL ASIGNED	P	R	ART.	TOPICS TERMS
t_28	soil_organ	soil organic carbon	4.799	1	825	soil	organ	depth	soc	properti
t_16	agroforestri_practic	adoption of agroforestry practices	4.931	2	785	farmer	farm	agroforestri	practic	adopt
t_27	speci_rich	biodiversity	4.527	3	694	divers	landscap	speci	habitat	conserv
t_13	climat_chang	climatic change global policies	4.656	4	511	project	develop	polici	approach	base
t_21	climat_chang	carbon and climatic change	3.36	5	487	carbon	climat	chang	stock	climat_chang
t_6	food_secur	food security	4.594	6	485	develop	food	research	sustain	resourc
t_15	fine_root	fine root	3.265	7	483	root	plant	growth	seedl	treatment
t_12	land_cover	land cover	3.866	8	392	land	area	cover	agricultur	degrad
t_10	allei_crop	alley cropping	3.626	9	389	crop	yield	maiz	field	allei
t_17	remot_sens	remote sensors	4.147	10	384	model	data	method	estim	variabl
t_20	soil_water	soil water	3.508	11	382	water	season	temperatur	rainfal	dry
t_11	shade_tree	shade tree	2.691	12	315	coffe	shade	cacao	pest	shade_tree
t_3	raw_materi	raw materials production	2.597	13	303	wood	energi	potenti	product	pine
t_26	product_system	production systems	3.408	14	298	product	econom	cost	benefit	market
t_23	soil_fertil	soil fertility	3.081	15	298	fertil	nutrient	fallow	increas	soil_fertil
t_4	intercrop_system	intercrop systems	3.017	16	282	intercrop	plant	space	wheat	poplar
t_25	plant_speci	plant species	3.343	17	268	speci	plant	nativ	divers	woodi
t_22	ecosystem_servic	ecosystem services	3.51	18	265	manag	agricultur	ecosystem	servic	practic
t_30	aboveground_biomass	aboveground biomass	3.191	19	260	biomass	year	growth	height	stand
t_18	genet_divers	genetic diversity	2.568	20	227	popul	genet	select	variat	trait
t_2	leaf_litter	leaf litter	2.588	21	221	litter	leaf	rate	content	concentr
t_8	local_peopl	local knowledge	3.326	22	216	local	tradit	knowledg	region	import
t_14	silvopastor_system	silvopastoral systems	2.22	23	201	pastur	grass	livestock	product	forag
t_24	secondari_forest	secondary forests	3.048	24	177	forest	af	natur	tropic	restor
t_9	agroforestri_system	agroforestry systems	3.548	25	159	system	agroforestri	agroforestri_system	base	monocultur
t_7	fruit_tree	fruit tree	2.127	26	155	fruit	cocoa	africa	pollin	west
t_1	tree_speci	tree species	3.358	27	148	tree	tree_speci	densiti	canopi	busi_media
t_29	cork_oak	cork oak	2.124	28	78	term	forestri	long	agro	long_term
t_5	rubber_plantat	rubber plantations	1.782	29	58	plantat	rubber	monocultur	palm	oil
t_19	fungal_commun	fungal communities	3.192	30	48	effect	posit	factor	affect	influenc

, it is possible to observe the topic names that were generated from the words with the highest number of repetitions. Words were ranked for relevance. After the searches within the articles according to each topic, the 20 words with the highest number of repetitions generated the prevalence rankings. The five main ones were: t_28 (soil organic carbon), t_16 (adoption of agroforestry practices), t_27 (biodiversity), t_13 (climatic change global policies), and t_4 (carbon and climatic change).

Our study also carried out the analysis with the LDA method, where the relationships between topics were observed. The existence of a possible relationship was obtained by statistical analysis of PC according to the matrix generated in the LDA model (Figure 8).

The grouping of topics and years was analyzed using the multivariate analysis method HJ-Biplot, using the theta matrix from the results, which was composed of probability values. First and foremost, a high quality of representation can be observed in the initial factorial plane, with 71.39% of the inertia being effectively absorbed and explained (Figure 9).

The topics t_14 (silvopastoral systems) and t_18 (genetic diversity) had greater relevance in the years 1993, 1994, and 1995. The topics: t_10 (alley cropping), t_23 (soil fertility), t_30 (aboveground biomass), t_20 (soil water), t_26 (production systems), t_15 (fine root), t_1 (tree species), t_7 (fruit tree), t_4 (intercrop systems), t_2 (leaf litter), t_29 (cork oak), and t_25 (plant species) had greater relevance in the years 1996, 1997, 1998, 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, and 2010.

The topics: t_28 (soil organic carbon), t_27 (biodiversity), t_11 (shade tree), t_19 (fungal communities), t_17 (remote sensors), and t_12 (land cover) had greater relevance in the years 2009, 2010, 2011, 2012, 2013, 2014, and 2015. Finally, in the years 2016, 2017, 2018, 2019, 2020, 2021, and 2022, the investigation of the topics t_22 (ecosystem services), t_21 (carbon and climatic change), t_3 (raw materials production), t_24 (secondary forests), t_9 (agroforestry systems), t_16 (adoption of agroforestry practices), t_8 (local knowledge), t_6 (food security), and t_13 (climatic change global policies) was the most relevant in the last seven years.

4. Discussion

The results of this study were consistent with those established by [49], who mentioned the United States, India, Germany, Brazil, Kenya, France, Australia, and the United Kingdom as countries with the highest interest in agroforestry research. Furthermore, some of the most common topics in this study were consistent with this 2023 analysis.

It was possible to notice a gradual increase in interest in research on agroforestry issues; proof of this was the increase over time in the number of publications made, which went from 120 in 1993 to 1011 in 2022, with a total of 145 countries involved in the 9794 articles that made up the study sample. However, there could be other factors with significant impact, such as economic investment for research, policies on sustainable land use, and the participation of entities specialized in agroforestry. In this regard, the United States, for example, has key policies to promote agroforestry; this is the case of the USDA’s National Agroforestry Center, which seeks to promote agroforestry science and its adoption. Similarly, Europe has included policies aimed at agroforestry issues in the common agricultural policy (CAP).

In this same sense, multistakeholder partnerships and agroforestry working groups and associations play extremely important roles in the generation and promotion of agroforestry science at different scales [50]. Considering the results of this study, countries such as India (which for the last 25 years has had a solid track record in agricultural and forestry research [51]), which has the Indian Council of Agricultural Research, the Indian Society of Agroforestry, and currently has the landmark national agroforestry policy that provides a framework designed to improve agricultural livelihoods; Germany that has the German Association for Agroforestry (DEFAF); France that has the French Agroforestry Association; Kenya that has the headquarters of World Agroforestry (ICRAF) are making efforts to develop agroforestry practices and economic growth, alleviate poverty, and make significant contributions in terms of environmental quality.

In this way, agroforestry is defined as a form of land use where it is possible to increase productivity, diversify production, and improve sustainability [52]. It is widely known as an agricultural production system where trees are grown together with annual crops and/or with animals, resulting in better complementary relationships between the components and an increase in multiple uses [53]. Additionally, it can provide high quality habitats, which is of great importance for the conservation of biodiversity. It can provide multiple alternatives for the commercialization of products, cushioning, in some cases, market fluctuations. The high levels of biodiversity also provide better ecological services, increasing local functions such as pest control [54] and pollination [55,56] and can improve soil fertility and reduce erosion [57,58].

Therefore, the relationship that existed between the results of the relevance analysis of the words used in the articles that made up the corpus under study was remarkable; the ten words that were the most relevant were: tree, system, soil, agroforestry, species, forest, crop, production, and land. In addition, it was possible to determine that there were topics that were gaining importance during the analysis period in the investigations; these being: soil organic carbon, adoption of agroforestry practices, biodiversity, climatic change global policies, carbon and climatic change, and food security.

5. Conclusions

This study provides relevant information about the evolution of science around the agroforestry topic over time and by country, providing a theoretical basis for the development of the discipline and could significantly contribute to decision making by researchers and research centers linked to the sector.

We found that the countries with the highest number of publications in agroforestry are the United States, India, Brazil, the United Kingdom, and Germany. Although there is generally a broad network of international cooperation, the United States and Brazil have more frequent connections in terms of international co-authorship relationships in the field of agroforestry.

Based on the results, we can conclude that there are certain topics related to raw material production, soil cover, adoption of agroforestry practices, biodiversity, and climate change that have gained importance over time, while other topics related to soil fertility, aerial biomass, and tree species have decreased in relevance. Bootstrap analysis also identified groups of topics that had higher probabilities in certain years, suggesting that certain topics may have been more relevant at certain times. Adoption of agroforestry practices, food security, global climate change policies, organic soil carbon, and biodiversity appear to be important topics both in 1993 and 2022, while topics such as alley cropping, soil fertility, fine roots, tree species, and aerial biomass appear to have been more relevant in 1993 and 2008. Overall, these results may be useful in identifying trends and areas of interest in agroforestry research, as well as guiding future research and policies related to agroforestry and sustainable land management.

Finally, the LDA method employed in this review categorized each subject based on subjective observations of high probability words and therefore showed effectiveness in generating responses about the most common topics studied in agroforestry. In contrast, the HJ-Biplot method was able to group topics by year, identifying which agroforestry topics were most relevant and for which years.