Analysis of Worldwide Greenhouse and Carbon Monoxide Gas Emissions: Which Countries Exhibit a Special Pattern? A Closer Look via Twitter

Abstract

There is significant global concern about the harmful effects of greenhouse gas and carbon monoxide emissions (deforestation, air pollution, global warming, etc.). The 2015 Paris Agreement on climate change aspires to reduce global warming by achieving a climate-neutral world. Research has been carried out to calculate and diminish the aforementioned emissions in waste, power industry, transport, building, in addition to other areas. The aim of this paper is to analyse the carbon and greenhouse gas emissions across countries around the globe in order to find patterns and correlate them to socio-economic indicators [gross national income (GNI), industrial production (IPI) and human development indexes (HDI)] as well as Twitter interactions regarding climate change. For this purpose, time series and socio-economic data have been downloaded from different repositories including EDGAR (Emissions Database for Global Atmospheric Research), World Bank and UNDP (United Nations Development Programme). Although classical clustering algorithms have already been used in the examination of some environmental issues, we use a non-parametric time series clustering method, which has been suggested in certain scientific literature as a more flexible approach, since any ad hoc parametric assumptions are required. The chosen socio-economic indicators have also demonstrated their relevance in pieces of research related to various fields. With respect to Twitter, which is one of the most popular social networks nowadays, significant analysis has also been performed on the basis of capturing citizens’ perceptions on a multitude of matters. We found that several countries such as Brazil, India, China, Nigeria, Russia, United States, Spain, Andorra, Greece, and Qatar show differences in carbon and greenhouse gas emissions patterns. Besides, there does not seem to be a correlation between GNI, IPI and HDI as well as the above mentioned emissions $(\text{correlation}<0.16).$ Regarding Twitter interactions, a dissimilarity in the distribution of hashtags was detected between the aforementioned countries and the rest of the world. This research can help to identify countries in which more governmental measures are needed to reduce the type of emissions analysed in certain industrial sectors. In addition, it points out the topics related to climate change that seem to generate the most debate on Twitter for countries with an unusual pattern.

Supplementary Information

The online version contains supplementary material available at 10.1007/s41742-023-00510-4.

Introduction

The Problem of Greenhouse Gas Emissions

A recent United Nations report on climate change (United Nations (w. dc.)) explained that harmful carbon emissions during the period from 2010 to 2019 had been the highest in the history of mankind. As a result, some catastrophic effects could happen in a few years, such as cities becoming submerged under water, major heat waves and storms as well as water shortages and the extinction of many species (United Nations (w. dd.)). The top ten greenhouse gas emitting countries generated $>60$% of total emissions in 2015, with China (21.1%), The United States of America (14.1%) and India (5.2%) registering as the greatest contributors (Althor et al. 2016). A more recent United Nations document entitled: “The Emissions Gap Report 2020” (Capstick et al. 2020), expressed concern about global warming. The document concluded that, despite a slight drop in carbon dioxide emissions motivated by the COVID-19 pandemic, an increase of more than $3{}^{\circ }\text{C}$ was expected this century (Capstick et al. 2020). International concern about climate change and its effects prompted the 2015 Paris Climate Change Agreement.1

The 2015 Paris Climate Change Agreement aims to restrict global warming, which requires that a climate neutral world2 is achieved (United Nations (w. db.)). In 2022, European Union foreign affairs ministers approved in 2022 conclusions taken from the 26th United Nations Climate Change Conference of the Parties, COP26, which had been held in Glasgow in 2021, and agreed on priorities for the work of the European Union on this matter. They were: (i) secure global net zero by mid-century and restrict global warming to well below 2, preferably 1.5 ${}^{\circ }\text{C}$. (ii) make modifications and adaptations as required to safeguard natural communities and habitats. (iii) Require private and public sector funding to achieve global net zero emissions. (iv) Governments, businesses and civil society must work together in order to solve the climate catastrophe (UNFCCC COP26). Various actions have also been taken in The United States in recent years. The United States rejoined the Paris agreement in 2020,3 creating a National Climate Task Force, where several cabinet-level leaders from across agencies are working together in order to carry out the following actions: (i) lower the gas emission 50–52% by 2005 and before 2030 (ii) achieve 100% carbon pollution-free electricity by 2035, (iii) reach a net-zero emissions economy by 2050 (iv) delivering 40% of the benefits from federal investments in climate and clean energy to the most affected communities (The White House (w. d.)). At COP21, India also proclaimed its purpose to minimise the emissions intensity by 33% to 35% by 2030 compared with the 2005 levels, and to have 40% of electric power capacity deployed from non-fossil-fuel-based energy sources Economic and Political Weekly (2020). China supported the Paris Agreement, strengthening its domestic policy goals for a fast and early diminution of its greenhouse gas emissions (Ye et al. (w. d.)).

Background

Analysis of Green Gas House Emissions

There is research that studies the greenhouse gas emissions trying to come up with solutions to reduce them in various economic areas (agriculture, waste, power industry, transport, etc.). Regarding the agricultural sector, (Aguilera et al. 2021) computed the emissions and potential carbon reductions related to crop and livestock generation systems in the Mediterranean region. Israel et al. (2020) detected the factors that determine the smallholder farmers’ participation in high-emitter activities of greenhouse gases. Impacts of climate smart agricultural activities on minimisation of the aforementioned emissions were also estimated. The emissions of pre-and post-production processes of agri-food systems have been rigorously studied (Tubiello 2019; Rosenzweig et al. 2020; Tubiello et al. 2022).

The waste area has also been thoroughly analysed (Lawin et al. 2012; Nordahl et al. 2020). Johnson et al. (2022) carried out an end-to-end study to compute emissions from all phases of the sanitation-service chain. Such a study concluded that from the emissions associated with long periods of storage of faecal waste a high percentage of greenhouse gas emissions are derived. Campos et al. (2016) showed that wastewater treatment directly produced a relevant volume of greenhouse gas emissions, and proposed modification of operational conditions, handling of the gaseous flows, and to perform new configurations and processes in order to minimise them.

In the power industry field some solutions have also been proposed in order to achieve reductions in greenhouse gas emissions, among others Quintero et al. (2014) suggested using compressed air power enhancement technology in gas turbines. Hammons (2016) examined the effects of electric power production on greenhouse gas emissions in Europe (including the Asian part of Russia), with special focus on Russia, Greece, and Italy. Ahmed et al. (2022) tested five benchmark systems, in conjunction with conventional thermal power plants and renewable energy sources, to achieve an optimum solution for both cost and pollutant emission. An astute black widow optimization (ABWO) approach was utilised.

The transport area has been exhaustively examined in Lewis et al. (2018), Kellner and Igl (2015), Miklautsch and Woschank (2022) and Albuquerque et al. (2020), which explored the existing literature related to greenhouse gas emissions generated by construction, operation, and maintenance stages of road projects. Certain assessment tools to calculate the road project-specific greenhouse gas emissions have also been presented. Kopfer et al. (2014) integrated into Dantzig’s classical vehicle routing model the option of choosing vehicles of different sizes for route fulfilment.

This paper studies, considering 202 countries, whether commonalities and differences exist in the greenhouse and carbon monoxide gas emission patterns in the field of agriculture, building, power industry, other industrial combustion, waste, and transport fields. Carbon monoxide (CO) is perceived as the silent killer gas of the twenty-first century (Dey and Dhal 2019) and the greenhouse gases are responsible for significant environmental damage (Darkwah et al. 2018).

Time Series Forecasting and Clustering Applied to Climate Change Data

Several research studies have carried out projections of the effects of climate change, applying several types of machine learning models such as generalised linear models (Jahn and Hertig 2022), random forest (van Straaten et al. 2022; Gibson et al. 2021), XGBoost (Gibson et al. 2021), as well as K-Nearest Neighbours Regressor, Decision Tree Regressor, Lasso Regression, among others (Isaev et al. 2022).

Clustering algorithms have also been utilised in the analysis of regional climates (Aliaga et al. 2017) and air pollution (Gao et al. 2011; Wang et al. 2015). The level of differentiation of the European nations in terms of emissions per capita of greenhouse gas emissions (carbon dioxide, nitrogen oxide, methane, and nitrous oxide) has also been studied utilising the k-means method (Kijewska and Bluszcz 2016). The study detected, based on annual Eurostat reports in the 1990–2014 period and analysing the full series, that the countries with the highest greenhouse gas emissions were Germany, United Kingdom, France, Turkey, Poland, Italy and Spain. A special clustering method to categorise sets of time series with a relevant volume of missing values has also been proposed. The structure is similar to those used in the k-means algorithm, but with some modifications in order to be able to select and attach of records with different temporal lengths (Carro-Calvo et al. 2021).

In this paper, we use a non-parametric clustering series method (instead of classic k-means), which was proposed in Vilar et al. (2009) as a very flexible approach, since it does not require any ad hoc parametric assumptions, and can be applied in a more general manner. Additionally, the classical multivariate procedures are not always useful in all time series circumstances (Vilar and Pértega-Díaz 2004).

Examination of the Relations with Economic Indicators

In addition to the examination of the connections between renewable energy consumption, economic growth and human development in some regions (Wang et al. 2020), the interconnections between economic indicators and other relevant social aspects (Batool and Liu 2021; Anker et al. 2008; Miladinov 2020; Ouedraogo 2013) have been studied.

The relationships between greenhouse gas emissions and economic indicators such as: gross national income (GNI), industrial production (IPI) and human development indexes (HDI) have been analysed. Each of the HDI components has been studied in depth (knowledge and education, life expectancy at birth, standard of living) (UNDP (w. d.)). The HDI has also been used in studies on deprivations and freedom of individuals in several regions (Sajith and Malathi 2021; Mariano et al. 2021). The IPI, which covers aspects such as manufacturing, mining, and utilities, has been thoroughly examined as well as had its importance accentuated (Ejaz and Iqbal 2019; Banikhalid and Oshaibat 2021). The GNI is a key economic indicator, as is evidenced by several pieces of the research (Nolan 2020; Cingano 2014).

Social Networks Analysis

Social networks are an interesting object of study, because they are a means of communication in which thoughts and opinions are built and shared between different communities. Twitter is one of the most used social networks, it had 217 million daily active users by February 2022 (Twitter by the numbers: stats, demographies and fun facts), (BACKLINKO 2022). Twitter data can also be obtained through robust Application (Weller et al. 2014).

Considering the above, Twitter can be an appropriate place to examine the existing interactions on certain current issues. The interesting role of the Twitter hashtag as a mechanism of managing a distributed discussion among many users, who do not need to be linked through existing follower networks, has been emphasised by various pieces of research (Bruns and Burgess 2011; Ferragina et al. 2021). Twitter hashtags enable users to increase the value of the posted content by restructuring, publishing, and disseminating the content (Chang and Iyer 2012). We analyse whether the distribution of discussions indexed on Twitter differs according to where the tweet was posted. OpenStreetMapping (OMS) Nominatim geolocation service was used (Kumar et al. 2014) in order to obtain the origin country of the tweet from longitude and latitude information.

Motivation and Objectives of the Research

The purpose of this research is to gain a better understanding of the global greenhouse and carbon monoxide4 gas emissions. We aim to answer the following research questions:

• Are there countries that present a particular pattern of greenhouse gas emissions according to the sector considered: agriculture, building, industrial combustion, etc.? Which are they? Is this also true for carbon monoxide emissions?

• Is there any relationship between the aforementioned emissions and certain socioeconomic variables such as: Human Development Index (HDI), the Industrial Production Index (IPI) and the Gross National Income (GNI)?

• Are there any differences between the distributions of discussions indexed on Twitter corresponding to countries with a particular pattern related to the rest of the nations?

The goals of this research are derived from the answers to the questions previously exposed.

The novelty of our study is that the proposed method in Vilar and Pértega-Díaz (2004) was used to analyse the carbon monoxide and greenhouse gas emissions in 202 countries and several sectors. We examined both the full series and their components (random5 and trend6 elements), which allowed us to identify nations with similar and different features. Additionally, we explored through various statistical tests the relationships of the full series and their components with socioeconomic indicators.

Furthermore, we checked, if there were differences in the hashtags used in those countries that exhibited a different pattern of the studied emissions, through a statistical analysis.

To our knowledge, all of the above had not been carried out before. A software tool, which is available for public use, was used to download the tweets to be analysed (Congosto et al. 2017). The utilised statistical tests were Anderson–Darling (LLC 2010), Fligner–Killeen (Beyene 2016), Kruskal–Wallis (Ostertagova et al. 2014), and Shapiro–Wilk (Das 2016) tests.

Materials and Methods

Overview of Resources Used

Repositories

The following repositories were used:

• Emissions Database for Global Atmospheric Research (EDGAR). It contains data series relative to the emission of all greenhouse gas emissions per industrial sector (buildings, non-combustion, other industrial combustion, power industry, etc.) from 1970 until 20157. It also incorporates data on the total greenhouse gas emissions per capita as well as data series related to carbon monoxide emissions. The information covers 202 countries/regions.

• United Nations Development Programme. This repository (European Commission (w. d.)) stores information on the HDI, which can take a value in the range [0, 1]. The repository contains information related to 189 countries and covers the period from 1990 to 2019.

• World Bank. This repository contains information on both IPI and GNI. Regarding IPI (WORLD BANK (w. da.); OECD (w. d.)). The data refers to the period 1919–2018 and is related to 40 countries. With respect to the GNI (WORLD BANK (w. db.)) data corresponds to 263 countries and regions, including the period from 1960 to 2019.

The countries considered in the studies are described in the Supplementary Material document.

Software Program: T-Hoarder Kit

The data extraction process was performed through the T-Hoarder kit (Congosto et al. 2017). It is a basic set of tools used to obtain Twitter information and process it. The tool runs in UNIX or Windows as an operating system and uses Python programming (Python (w. d.)). Using T-Hoarder, one can retrieve both: user information (profiles, followers, following, relations, tweets and H-index) and tweets containing a set of keywords. The latter can be done in two ways: making a query or capturing them in real time. In both cases, the output format is equal. By making a query on Twitter one can obtain tweets as far back as the previous week (Rest API (Application Programming Interface)) whereas if tweets are taken in real time the data is collected continuously from the moment it is requested (Streaming API).

In this work, data was crawled through the two aforementioned methods. The set of keywords selected to extract only tweets related with climate change are8climate change, global warming, greenhouse effect, climate emergency, climate crisis, climate disaster, and climate action with and without spacing (to include hashtags present in tweets). The search produces a csv (Comma-separated values) file containing the following information per tweet: id_tweet, timestamp, author, text, app, id_author, followers_author, following_author, statuses_author, location, url, geolocation, name, bio, url_media, type_media, and lang.

The description of all these fields has been included in the Supplementary Material Document.

Other Software Programs

Various programs in R language (The R Project for Statistical Computing. https://www.r-project.org/) were implemented in order to perform the following functionalities:

• Extraction of the random and trend components of the time series. Realisation of the series clustering. The locpol, cluster, purrr, NbClust, clValid and sm packages were used.

• Processing data downloaded from repositories. Estimation of correlations. corrr package was used.

• Execution of statistical tests.stats, nortest, and tidyverse packages were utilised.

• Creation of graphic representations.

Also, several programs in Python were developed for the purpose of carrying out the following functionalities:

• Handling of the data was performed using pandas library, through DataFrames (2D data structure with columns of potentially different types)

• During the process of data preparation the following packages are utilised: NLTK, Demoji module, and re (regular expressions operators).

Overview of Methods Used

Extraction of Tweets

A total of 926,494 tweets were downloaded using the T-Hoarder tool from February 2021 to May 2021. This period was selected because the project under which this research was supported was granted in the first quarter of 2021, with the results having to be presented in the last quarter of that year.

The downloaded information includes a collection of tweets and messages produced from Twitter users and their interactions such as retweeting (RT), replying and quoting. Among all tweets, we found 589,272 RTs, 94,084 replies and 156,612 quotes. The number of RT found is the largest because it is produced when a user forwards another user’s tweet to their followers. This interaction is the most common because it easily allows the sharing of interesting findings and trending topics with which one can relate. However, similarly to Mouronte-López and Subirán (2022), in this work all interactions are neglected because they do not offer any personal opinion towards climate change. Humans generate much more new, sensitive content, while bots produce many more retweets. Bots tweet more URLs (Uniform Resource Locator) and multimedia components (Gilani et al. 2017; Himelein-Wachowiak et al. 2021). Considering the above, we do not carry out any specific processing to detect bots.

Additionally, analogously to Mouronte-López and Subirán (2022), the data was filtered to remove possible duplicates which may occur from how the data was acquired (streaming API and rest API). Then, non-English tweets are removed using the lang category which identifies the tweet’s language. After this, the remaining number of tweets (92,474), which will be used for later analysis, are posted by 56,866 unique users. Figure 1A depicts the count of unique users per tweet frequency.

Fig. 1.

A Histogram showing the count of unique users per tweet frequency. B Histogram representing the amount of tweets with a specific number of hashtags per tweet

Tweet Processing

In this stage, the purpose is to process the text in tweets to normalise the data so that later experimentation is facilitated. Firstly, similarly to Mouronte-López and Subirán (2022), punctuation is removed along with links, user mentions and hashtags. The latter is stored for later usage. Moreover, analogously to Mouronte-López and Subirán (2022), emojis are also removed from the text and replaced by their translation into text (provided by the demoji module Solomon 2020). Then, the NLTK library (Bird 2021) is used for two purposes. First, to remove non-discriminant features (keywords used during data crawling and the list of stop-words provided by nltk). Second, the text is tokenized, stemmed and lower-cased, while words containing less than two characters are deleted.

From the 92,474 tweets analysed, it was found that 33,251 (approximately 36%) tweets contain hashtags related to climate change. Moreover, the total number of hashtags is 73,544 which is greater than the number of tweets containing hashtags because one tweet can include several hashtags. To discover more insights into the data, the distribution of the number of hashtags per tweets is analysed and it is found that predominantly the tweets contain a single hashtag, with several tweets including up to ten hashtags (see Fig. 1B). These hashtags will later be analysed to find relations or coincidences in topic focus between countries with special emissions patterns.

GeoPy

Similarly to Mouronte-López and Subirán (2022), to identify where the collected tweets were posted, OpenStreetMapping (OMS) Nominatim geolocation service from GeoPy was used. Two different methods are used to extract this information. First, the geolocation field is inspected. On Twitter, this setting has to be manually activated by the user and provides the precise coordinates (longitude and latitude) from which the tweet was posted. Using reverse geocoding, the location can be inferred. Second, since geotagged tweets represent only around 1% of the total tweets, other user profile information such as the field location must be used. Using the geocode method the location of the tweet can be resolved from a string which can be the name of a city, state, province, etc. Taking advantage of these two options, the location retrieval rate is maximised. Unfortunately, not every user specifies location or geotags their tweets. From the data under analysis, location could be identified for 47,471 tweets (approximately 50% of all tweets).

Time Series Analysis of Emission-Related Variables

As is well known, any time series can be decomposed in trend, seasonality and random terms. In the present research, various n univariate time series were used:

$$\begin{array}{c}\hfill S=\{S_1,S_2,\dots S_n\}\end{array}$$ 1

being:

$$\begin{array}{c}\hfill S_j=\{S_{j,1},S_{j,2},...S_{j,T}\}\end{array}$$ 2

where j is ranged from 1 to n, and T represents the number of observations. All time series were sampled at the same time and their number of time points were the same. Concerning the series components, such as only annual data were used, the seasonality term took a null value.

Time series clustering is an unsupervised learning method9 in which data varying over a period of time (those analysed time series) is categorised into groups with similar characteristics through a distance metric. Those time series that are most analogous are placed in an identical group. The procedure is substantiated by the establishment of a maximum distance, which allows us to decide the linking objects and their inclusion in an identical conglomerate.

In this research, in order to know whether common patterns in pollutant emissions between countries exist, a clustering analysis was applied to the time series. For greenhouse gas emissions time series, the clustering analysis was used on series related to non-combustion, buildings, power industry, transport and other industrial combustion sectors. The total greenhouse gas emissions per capita was also studied. For carbon monoxide gas emissions, the clustering analysis was carried out on series that referred to agriculture, build, other industrial combustion, other, power industry, and waste sectors. In particular, we use a non-parametric method proposed in Vilar et al. (2009), which is a model-based time series clustering approach. A summary of this algorithm is included in the Supplementary Material Document.

Additionally, the standard series clustering technique was also carried out on the full original series.10

The optimal number of clusters was calculated using the indexes Silhouette (Rousseeuw 1987) and Dunn (1973, 1974). For this estimation, the Euclidean distance (Euclidean Distance) was utilised as a distance metric.

The Dunn index (DI), which detects those cluster sets that are compact and properly segregated, can be defined as Ansari et al. (2011):

$$\begin{array}{c}\hfill \text{DI}=\underset{1\le i\le k}{min}\left(\underset{i+1\le j\le k}{min},\frac{\text{dist}(c_i,c_j)}{\text{max diam}(c_l)}\right)\end{array}$$ 4

• $\text{dist}(c_i,c_j)$ is the distance between i and j clusters, which is described as ${min}_{x_i\in c_i,x_j\in c_j}$ $d(x_i,x_j)$, where $d(x_i,x_j)$ represents the distance between data points $x_i\in c_i$ and $x_j\in c_j$.

• $\text{diam}(c_l)$ is the diameter of cluster $c_l$, which is defined as ${max}_{x_{l1},x_{l2}\in c_l}d(x_{l1},x_{l2})$.

DI takes a value in the range $[0,\infty ]$. The higher the Dunn index value, the better the clustering.

The Silhouette Coefficient (SC), calculates the Silhouette width for each data point, averaging out the Silhouette width for each cluster and the overall average Silhouette width for the full data set.

In the calculation of this coefficient, the average dissimilarity of a data point i to a cluster $C_m$, which is described as the mean of the distance from i to all points in $C_m$, is used as a metric.

The Silhouette width of ith data point is defined as Ansari et al. (2011):

$$\begin{array}{c}\hfill \frac{o{c}_i-i{c}_i}{max(i{c}_i,o{c}_i)}\end{array}$$ 5

• $i{c}_i$ is the mean dissimilarity of ith data point to all other points in an identical cluster.

• $o{c}_i$ is the minimum mean dissimilarity of ith data point to all data points in another cluster.

SC is in the range $[-1,1].$ A value equal to 1 symbolises that the data point is very compact within the cluster where it is located as well as far away from the rest of clusters. The worst value is $-1.$ A value close to 0 represents overlapping clusters.

We take the number of clusters that maximise both DI and SC as the optimal number of clusters.

Correlations with Socioeconomic Data

The correlations between both greenhouse and carbon monoxide gas emissions with the HDI, IPI and GNI were calculated. An examination of the normality of the distributions was made in order to establish whether they must be estimated using Spearman’s or Pearson’s methods. The Anderson–Darling test was applied, with a significance level equal to 0.05. The following hypotheses were used:

• $H_0$: “The sample is normally distributed”

• $H_a$: “The sample is not normally distributed”

If p-value $\le$ 0.05, $H_0$ is rejected and correlation is calculated using Spearman’s procedure. Otherwise $H_0$ is accepted and Pearson’s algorithm is applied. The Spearman’s and Pearson’s algorithms are thoroughly described in Statistics Solutions. A detailed analysis of the Anderson–Darling test can be found in Ramachandran and Tsokos (2021).

Hashtags Analysis

A hypothesis test was performed to determine whether there was a statistical difference between the hashtags used in the tweets according to the geographical location indicated therein, i.e. depending on whether their location corresponded to countries with a different pattern of emissions.

For hashtags in tweets whose locations are nations with a peculiar pattern of emissions, we calculate the total percentage of tweets in both types of nations (i.e. those that exhibit this unusual behaviour and those that do not). In order to establish the most appropriate hypothesis test to be performed, be it either parametric or non-parametric, the normality and homoscedasticity of the hashtags distributions were checked. For these two purposes, the Anderson–Darling and the Fligner–Killeen tests were applied with a significance level $\alpha =0.05$.

The hypotheses used in the Anderson–Darling test were:

• $H_0$: “The sample is normally distributed”

• $H_a$: “The sample is not normally distributed”

And, the hypotheses utilised in the Fligner–Killeen test were:

• $H_0$: “The variance is constant”

• $H_a$: “The sample is not constant”.

If both normality and homoscedasticity existed in the distributions, the analysis of variance (ANOVA) method would be used as a parametric hypothesis test. Otherwise, the Kruskal–Wallis non-parametric test would be utilised. A significance level $\alpha =0.05$ would be taken. The hypotheses utilised were:

• $H_0$: “The groups come from identical populations.”

• $H_a$: “The groups derive from different populations.”

In all cases, if p-value had a value $\le$ 0.05, $H_0$ would be rejected and $H_a$ would be taken.

Results

Worldwide Emissions Characterization

Temporal Evolution of Emission-Related Variables

The worldwide greenhouse and carbon monoxide gas emissions are depicted in Figs. 2, and 3, respectively. This will be explained later in the “Discussion”.

Fig. 2.

Worldwide greenhouse gas emissions by sector in the 1970–2015 period. A Buildings (tonnes/year), B other industrial combustion (tonnes/year), C other sectors (tonnes/year), D power industry (tonnes/year), E transport (tonnes/year), and F total GHG tonnes/capita/year. Graph has been elaborated with data from EDGAR. Note: GHG is equivalent to greenhouse gas emissions

Fig. 3.

Worldwide carbon monoxide emissions by sector in 1970–2015 period. A Agriculture (tonnes/year), B buildings (tonnes/year), C power industry (tonnes/year), D transport (tonnes/year), E waste (tonnes/year), F other industrial combustion (tonnes/year) and H other (tonnes/year). Graph has been elaborated with data from EDGAR

Time Series Clustering of Emission-Related Variables

The results obtained in the clustering series for carbon monoxide and greenhouse gas emissions are depicted in Tables 1, 2, 3, 5, 6, and 7. The values of Silhouette (Rousseeuw 1987), Dunn (1973, 1974) indexes for the number of clusters presented in the aforementioned Tables are included in the Supplementary Material document. Tables 4 and 8 depict several statistical parameters for the full series. These results will be examined in the section labelled as “Discussion”.

Table 1.

Clustering analysis of trend components of carbon monoxide emissions time series (expressed as tonnes/year)

Sector	Cluster	Country	Sector	Cluster	Country
Agriculture	2	Brazil, India	Other	2	China
	1	Rest of countries		1	Rest of countries
Build	3	India, Nigeria	Power industry	2	China, Russia, United States
	2	China		1	Rest of countries
	1	Rest of countries
Other industrial combustion	2	China	Waste	2	Spain, Andorra
	1	Rest of countries		1	Rest of countries

Table 2.

Clustering analysis of random components of carbon monoxide emissions time series (expressed as tonnes/year)

Sector	Cluster	Country	Sector	Cluster	Country
Agriculture	2	Brazil	Other	2	China
	1	Rest of countries		1	Rest of countries
Build	2	China	Power industry	3	United States
	1	Rest of countries		2	Russia
Other industrial combustion	2	China		1	Rest of countries
	1	Rest of countries	Waste	2	Spain, Andorra
				1	Rest of countries

Table 3.

Clustering analysis of full carbon monoxide emissions time series (expressed as tonnes/year)

Sector	Cluster	Country	Sector	Cluster	Country
Agriculture	2	Brazil, India	Other	2	China
	1	Rest of countries		1	Rest of countries
Build	3	India	Power industry	3	Russia, United States
	2	China		2	China
	1	Rest of countries		1	Rest of countries
Other industrial combustion	2	China	Waste	2	Spain, Andorra
	1	Rest of countries		1	Rest of countries

Table 5.

Clustering analysis of trend components of greenhouse gas emissions time series (expressed as tonnes/capita/year)

Sector	Cluster	Country	Sector	Cluster	Country
Non-combustion	2	Greece	Power Industry	2	Greece
	1	Rest of countries		1	Rest of countries
Buildings	2	Greece	Transport	2	Greece
	1	Rest of countries		1	Rest of countries
Other industrial combustion	2	Greece	Total greenhouse gas emissions per capita	3	United Arab Emirates
	1	Rest of countries		2	Qatar
				1	Rest of countries

Table 6.

Clustering analysis of random components of greenhouse gas emissions time series (expressed as tonnes/year)

Sector	Cluster	Country	Sector	Cluster	Country
Non-combustion	3	Greece	Power Industry	3	Greece
	2	China		2	China
	1	Rest of countries		1	Rest of countries
Buildings	2	Greece	Transport	2	Greece
	1	Rest of countries		1	Rest of countries
Other industrial combustion	3	Greece	Total greenhouse gas emissions per capita	2	Qatar
	2	China		1	Rest of countries
	1	Rest of countries

Table 7.

Clustering analysis of full greenhouse gas emissions time series (expressed as tonnes/year)

Sector	Cluster	Country	Sector	Cluster	Country
Non-combustion	2	Greece	Power Industry	2	Greece
	1	Rest of countries		1	Rest of countries
Buildings	2	Greece	Transport	2	Greece
	1	Rest of countries		1	Rest of countries
Other industrial combustion	2	Greece	Total greenhouse gas emissions per capita	3	United Arab Emirates
	1	Rest of countries		2	Qatar
				1	Rest of countries

Table 4.

Statistical parameters of clustering analysis for full carbon monoxide emissions time series (expressed as tonnes/year)

Sector	Cluster	Median	Maximum	Minimum
Agriculture	1	21,700.72	4,352,046	0
	2	5,015,170.35	13,594,249.61	1,486,573.71
Build	1	116,482.09	18,642,799.69	4.87
	2	45,936,969.04	55,838,351.31	29,703,285.64
	3	22,154,393.81	29,065,177.69	15,316,731.22
Other industrial combustion	1	36,920.89	12,856,887.98	0
	2	17,068,870.39	48,330,211.31	9,037,866.16
Other	1	30.6	4,248,178.75	0
	2	3,916,302.98	32,385,403.84	2,077,526.12
Power industry	1	2,168.29	680,145.13	0
	2	349,198.56	1,416,033.73	28,793.40
	3	850,112.02	1,501,543.04	268,666.23
Transport	1	137,218.54	31,876,969.43	0
	2	93,223,033.11	119,019,160.72	39,557,819.35
Waste	1	0	16,938.32	0
	2	40,779.57	44,273.79	167.69

Table 8.

Statistical parameters of clustering analysis for full series of greenhouse gas emissions time series (expressed as tonnes/year)

Sector	Cluster	Median	Maximum	Minimum
Buildings	1	1,438,739.75	918,119,861	0
	2	3,609,904,637	3,923,624,888	3,242,346,554
Non-combustion	1	12,892,273.25	3,731,463,751	9365
	2	12,029,922,761	16,064,340,602	9,932,195,022
Other industrial combustion	1	1,625,592.25	3,206,766,301	0
	2	5,059,481,042.50	7,901,009,513	4,279,057,646
Power Industrial	1	1,956,621.76	4,435,608,401	0
	2	8234,042,441.98	13,759,392,540	3,697,886,023
Transport	1	2,028,544	1,856,378,442	0
	2	4,884,459,973.50	7,862,869,272	2,868,487,131
Total greenhouse gas emissions per capita	1	3.93	306.13	0.22
	2	83.73	790.85	51.04
	3	44.23	431.04	24.83

Correlations with Socioeconomic Data

To estimate the correlations of the greenhouse and carbon monoxide gas emissions with socioeconomic data, the mechanisms explained in the “Overview of Methods Used” are used. The calculation of the correlations considers the values of the aforementioned emissions from 1990 to 2015, as well as the HDI, IPI and GNI.

The application of Shapiro–Wilk test to the total greenhouse and carbon monoxide gas emissions as well as HDI, IPI and GNI variables show a p-value $\le 0.05$. For each variable, the obtained p-value is depicted in Table 9. None of the variables are normally distributed. Therefore, the correlation between the previously mentioned emissions, HDI, IPI and GNI has to be calculated using the Spearman’s method. The results are depicted in Table 10.

Table 9.

For greenhouse and carbon monoxide gas emissions, HDI, IPI, and GNI variables, obtained p-value in the Shapiro–Wilk test

Variable1-Variable2	p-value Variable1	p-value Variable2
GHG-HDI	2.30${\text{e}}^{-82}$	4.79${\text{e}}^{-23}$
GHG-IPI	4.98${\text{e}}^{-46}$	3.85${\text{e}}^{-08}$
GHG-GNI	4.08${\text{e}}^{-93}$	4.26${\text{e}}^{-84}$
CM-HDI	9.81${\text{e}}^{-78}$	5.10${\text{e}}^{-23}$
CM-IPI	4.62${\text{e}}^{-41}$	4.28${\text{e}}^{-08}$
CM-GNI	1.07${\text{e}}^{-88}$	9.46${\text{e}}^{-85}$

Only countries and years with values in each duo of variables are considered

Table 10.

Correlations of HDI, IPI and GNI with the total carbon monoxide and greenhouse gas emissions

Correlation	Value
GHG-HDI	− 0.03818
GHG-IPI	0.15592
GHG-GNI	0.01745
CM-HDI	− 0.05477
CM-IPI	− 0.04111
CM-GNI	− 0.01962

As is well known, a correlation between any two variables, can be categorised as: perfect correlation, in which its value is ± 1; high correlation, where its magnitude is in the range $(-1,-0.5]$ and [0.5, 1); moderate correlation, in which it is in the range $[-0.30,-0.5)$ and [0.3, 0.5); low correlation, where its value is in the range $(-0.30,0)$ and (0, 0.3); and finally, no correlation, where it takes the value 0. In Table 10, it can be observed that the correlations GHG-HDI, GHG-GNI, CM-HDI, CM-IPI and CM-GNI exhibit values very close to zero. The correlation GHG-IPI is low.

Hashtags on Twitter

A total of 4449 unique hashtags were found in 15,050 tweets in Andorra, Brazil, China, Greece, India, Nigeria, Qatar, Russia, Spain, and The United States of America. This information is shown, for the aforementioned countries, in Table 11. Table 12 shows those existing hashtags in tweets whose locations are included in both groups of countries, but which occur in more than 70 tweets and in those with a peculiar emissions pattern. Table 13 depicts the existing hashtags only in the latter countries. the above will be discussed later in the “Discussion”.

Table 11.

Number of tweets with unique hashtags in countries with a special pattern of emissions

Country	Number of tweets with unique hashtags	Country	Number of tweets with unique hashtags
United States	8385	Russia	56
India	5566	Qatar	55
Brazil	64	China	38
Spain	389	Andorra	5
Nigeria	350	United Arab Emirates	0
Greece	142

Table 12.

Hashtags present in all countries, which have more than 70 tweets, in Andorra, Brazil, China, Greece, India, Nigeria, Qatar, Russia, Spain, and United States of America, number of tweets and unique users are shown

Hashtag	Number of tweets	Unique users	Hashtag	Number of tweets	Unique user
Environment	392	297	Energy	94	81
Earth_Day	247	204	Air_Pollution	93	24
Art	196	9	Sustainable	91	74
Nature	167	118	C_O_P26	90	66
Earth_Day2021	137	127	R_T	84	10
C_O_V_I_D19	121	91	Afforestation	83	4
Earth	118	92	World_Water_Day	78	73
Water	104	90	Artist	77	4
Climate_Justice	98	46	Pollution	74	54

Table 13.

Hashtags present with more than 70 tweets, only in Andorra, Brazil, China, Greece, India, Nigeria, Qatar, Russia, Spain, and The United States of America, the number of tweets and unique users are shown

Hashtag	Number of tweets	Unique users	Hashtag	Number of tweets	Unique users
Surrealism	163	1	W_M_T_C_Blog	129	3
We_Mean_To_Clean	145	3	Climate_Changeart	114	1
Clean_Delhi	143	3

In Figs. 4 and 5 a diagram exhibiting the importance of the hashtags according to the emission pattern is represented. Using the concept of quartile, the data samples are divided into four equal parts, which are defined as:

• Quartile 1, $Q_1=[0,0.25\ast \text{MaxNumTweets}$]

• Quartile 2, $Q_2=(0.25\ast \text{MaxNumTweets},0.50\ast \text{MaxNumTweets}$]

• Quartile 3, $Q_3=(0.50\ast \text{MaxNumTweets},0.75\ast \text{MaxNumTweets}$]

• Quartile 4, $Q_4=(0.75\ast \text{MaxNumTweets},\text{MaxNumTweets}$]

where $\text{MaxNumTweets}$ is the total number of tweets including either common hashtags worldwide or only those in countries with a special emission pattern.

Fig. 4.

Diagram showing the significance of the common hashtags in countries with a special pattern of emissions $(\text{MaxNumTweets}=9744)$

Fig. 5.

Diagram showing the significance of the hashtags only in countries with an unusual pattern of emissions $(\text{MaxNumTweets}=5306)$

The dependence of hashtags on location was analysed according to the procedure described in the “Materials and Methods”. The obtained p-value was $\le$ 0.05 in the Anderson–Darling test. The statistical distributions showed no homoscedasticity either, since the p-value was $\le$ 0.05 in the Fligner–Killeen test. Therefore, the Kruskal–Wallis test was utilised, obtaining a p-value equal to 0.

Discussion

Worldwide Emissions Characterization

Temporal Evolution of Emission-Related Variables

As can be noted in Fig. 2, in all sectors, the greenhouse gas emissions expressed as tonnes/year show an increase between years 1970 and 2015. The highest accretion occurs for the power industry. This increment was 456,140,817 tonnes/year for buildings, 6,132,145,580 and 3,238,727,657 tonnes/year for both non-combustion and combustion industrial sectors. The power industry field showed a growth of 9,937,664,287 tonnes/year and the total greenhouse gas emissions per capita also experienced an increase of 0.09 tonnes/year. Focusing on the European Union, greenhouse gas emissions were reduced by 31% between 1990 and 2020 going beyond the European Union’s 2020 goal by 11 percentage points. This skip was driven by steep emission cutbacks in 2019 and 2020. While the reduction in 2019 was mainly motivated by fossil fuel prices consequences and policy measures, the reduction in 2020 was linked to the COVID-19 pandemic (Total greenhouse gas emission trends and projections in Europe).

The power industry is the top greenhouse gas emitter in many countries (Timeline of major accomplishments in transportation, air pollution, and climate change; StatisticsCanada (w. d.)). According to the McKinsey report entitled: “Redefining the power industry” (McKinsey Sustainability (w. d.)), a relevant transformation in power industry networks is required in order to achieve significant emissions reductions through a more robust and resilient infrastructure. The usage of microgrids, with locally managed energy loads, as well as distributed plants with energy resources that can work independently are suggested. In addition to this, a greater use of battery storage, supplying local energy reserves (McKinsey Sustainability). Also, 5G technology can help companies decentralise their energy infrastructure and manage their network more efficiently by connecting devices and assets using Internet Of Things (IoT) (Deloitte 2022).

With respect to the carbon monoxide emissions, they grow in all areas except transportation,11 between the years 1970 and 2015. Specifically, this expansion was 32,352,380.37 tonnes/year in agriculture, 49,235,854.62 tonnes/year in buildings and 52,290.82 tonnes/year in waste. The rise was 53,040,454.35 tonnes/year in other industrial combustion sectors, and 29,716,378 tonnes/year in the rest of the fields. However, a reduction of 1,452,734.39 tonnes/year happened in the transportation area. In Europe, most of the policies and measures envisaged in the area of transport focus on promoting low-carbon fuels, the use of electric cars, as well as encouraging people to use public transport European Environment Agency (2021). The brusque recession in total carbon monoxide emissions during the COVID-19 pandemic, originated by government-mandated lockdowns, is expected to disappear by the end of 2021 (Tollefson J (w. d.)).

Catalytic transformation of carbon monoxide (CO) is one of the most relevant chemical processes to protect human health from the effects of this gas. Many CO oxidation catalysts have been tested Dey and Dhal (2019), but many of them exhibit difficulties with deactivation in the presence of moisture (Cai et al. 2012), high cost (Dey et al. 2018; Daté et al. 2004), lower thermal stability (Burch et al. 2002), and resistance (Marinoiu et al. 2015).

Time Series Clustering of Emission-Related Variables

The United States, which have not signed or ratified the Kyoto Protocol (Britannica 2020) yet,12 has a special pattern of emissions, as detailed below. Specifically, in the power industry for the carbon monoxide emissions, as shown in Table 2.

As can be seen in Tables 1, 2 and 3, carbon monoxide emissions per sector, in the trend and random components as well as in the full series, show significant differences in several countries. The clustering analysis has grouped them into different conglomerates. The countries displaying a significant difference to the rest are Brazil, India, China, Nigeria, Russia, The United States, Spain and Andorra. Particularly relevant is the case of China, with differences in 4 sectors both in the full and trend components. This country also presents dissimilarities in the random component in 3 sectors. By sector, for the full series, we observe that in the area of Build, China and India are in the clusters with the highest statistical values. China also shows this situation for other types of industrial combustion and for the remaining types of consumption. For the power industry, Russia, the United States and China are the countries that are located in clusters with the highest figures. In terms of waste, Spain and Andorra are the ones present in the cluster with the highest values. According to research, the high carbon monoxide emissions produced in China happen because of its high population, inefficient capital investment, heavy reliance on coal and poor urban planning (Mohajan 2014).

As can be seen in Tables 5, 6, and 7, greenhouse gas emissions also exhibit inequalities in the trend and random component analysis as well as in the full series. Greece, Qatar, and China show a dissimilar behaviour to the rest. Particularly noteworthy is the case of Greece, with differences in 5 sectors. Therefore, for the full series, Greece is in the clusters with the largest values for non-combustion, buildings, other types of industrial combustion, power industry and transport. Also, the case of China is remarkable, with an unusual pattern, in the random component for 3 sectors. Qatar is located in the cluster with the highest values for the greenhouse gas emissions per capita, while the Arab Emirates is in the one with the lowest values. In 2021, Qatar started a national climate change action plan aiming to reduce greenhouse gas emissions by 25% before 2030 (Qatar targets 25respect to The Arab Emirates, its largest contributor of greenhouse gas emissions (with a 90% over the total) in 2016 was the energy sector (Khondaker et al. 2016).

Greece achieved a 56.3% decrease in emissions in all sectors included in the European Union’s Emissions Trading System in 2020 in relation to 2005. This was the third best performance among the member countries of the European Union. Slightly better results were obtained by Denmark and Estonia. However, Greece had the highest percentage of reduction in emissions from coal among the European-Union coal-producing countries (Balkan Green Energy News (w. d.)). The Greek power system is under frequent development with respect to installed capacity, its composition and fuel mix. Total installed capacity is embroidered, as is the installed capacity of the power plants utilising Natural Gas or Renewable Energy Sources (Tsilingiridis et al. 2013). The situation of China seems to be due to the high utilisation of coal and of fossil fuels (Leggett 2011). This country is responsible for a relevant share of the global livestock emissions, particularly caused by pork production. The technological progress and structural modifications in animal husbandry reduced emissions somewhat (Dai et al. 2021). However, further technological advances and changes in the economic structure are required (Dai et al. 2021).

Correlations with Socioeconomic Data

As can be observed in Table 10, the correlation is very close to zero (or it is low) between the total amount of carbon monoxide and greenhouse gas emissions and the IPI, HDI and GNI indexes. These types of emissions seem to be unrelated to a country’s economic growth and human development. This is in line with (Fujii and Managi 2016), which showed by exploring an environmental Kuznets curve (EKC) (Zhang 2021), that the relationship between GDP per capita and sectoral environmental pollutant emissions vary both by industry and by types of air pollutants. There, the authors explain that sulphur dioxide (${\text{SO}}_2$) and nitrogen oxide (NOx) seem to have a link to the above-mentioned economic growth. However, this is not the case for carbon monoxide emissions or other pollutants related to global environmental issues. A relevant relationship between ${\text{SO}}_2$ emissions and economic growth is also demonstrated in Xu and Yang (2020).

Hashtags in Twitter

Understanding what contributes to the popularity of a hashtag has been the goal of several investigations (Ma et al. 2013; Fedushko et al. 2019; Graciyal and Viswam 2018). Certain researchers have concluded that if a hashtag is included in a tweet together with other similar hashtags, the tweet popularity increases (Pervin et al. 2015). The authorship and content of posts on certain topics on social networks has also been examined using hashtags (Carlson et al. 2020).

Regarding our research, we found that the distribution of the hashtags in the tweets showed a dependence on the group where their location was, since the Kruskal–Wallis test yielded a p-value $\le$ than 0.05.

For the hashtags present in the tweets of all the analysed countries, Fig. 4 shows the importance they exhibit in territories with a special pattern of emissions. Environment, and Earth Day are in the Q4 and the Q3 quartiles, respectively. The Q2 quartile locates: nature, Earth_Day2021, COVID19, earth, and water.

Figure 5 shows the relevance of hashtags only used in countries with a special pattern of emissions. The Q4 quartile includes: surrealism, We_Mean_To_Clean, Clean_Delhi, W_M_T_C_Blog. Climate_Changeart is in the Q3 quartile; Colorado_Wildfires, Volunteer, and New_Orleans are in the Q2 quartile.

Among the common hashtags around the world, those more highly present in Andorra, Brazil, China, Greece, India, Nigeria, Qatar, Russia, Spain, and The United States of America, are shown in Table 12. In particular, environment and Earth_Day,13 which represent a 2.60% and 1.64% respectively over the total tweets, were placed in the highest positions. They also appeared among the top ten most frequent hashtags in the rest of the world. Regarding, Earth Day, it is a modern environmental movement born in 1970, whose mission is to diversify, educate and drive an environmental movement around the globe (Growing The Movement Since 1970 Earth Day Every Day).

As shown in Table 13, those hashtags that only exist in tweets with locations that have a characteristic emissions pattern, given the number of users using them, were of low significance.

Conclusions and Future Research

The conclusions obtained for each of the objectives are detailed below:

Ranking of countries according to their carbon and greenhouse gas historical emissions. Discovery of relationships with various socio-economic parameters.

A correlation very close to zero is observed (or it is low) between IPI, carbon monoxide and greenhouse gas emissions, which is also the case of the HDI and GNI. This suggests that carbon monoxide and greenhouse gas emissions seem to be more related to policies and industry-specific characteristics (technology, planning, etc.) rather than to the country’s level of development or its production volume and gross national income.

Brazil, India, Nigeria, China, Russia, The United States, Spain and Andorra show different patterns of evolution in carbon monoxide emissions compared to the rest of the countries. The same applies to greenhouse gas emissions. Some of these countries stand out for having a highly polluting industry, while others have made significant reductions in their emissions.

Detection of whether the distribution of discussions indexed on Twitter differs according to where the tweet is located.

The distribution of hashtags in tweets have different statistical distributions depending on whether or not tweets have a location with characteristic greenhouse and carbon monoxide gas emission patterns.

Identifications of hashtags only utilised in countries with a special pattern of emissions.

We find that some hashtags are exclusively used with high relevance in countries with a particular pattern of emissions.

This investigation could be continued:

With regards to used indicators:

A review of some of the most relevant environmental indicators at the country level could be carried out, which would be: indices of consumption of ozone depleting substances (ODS), pollution loads to water bodies, population exposure to air pollution (OECD 2008; Environmental Indicators).

An examination of environmental performance indicators (EPI) referred to the main organisations in a country would also be performed. This would allow us to know how firms impact on ecosystems, land, air and water, etc. In summary, EPI reveals how organisations are operating, and provides a baseline for future improvements (Environmental Performance Indicators).

With respect to the emissions analysis:

Predicting the greenhouse gas and carbon monoxide gas emissions through time series forecasting methods such as Autoregression (Zubaidi et al. 2021), Moving Average, Autoregressive Moving Average (Abdel-Rahman and Marzouk 2018), Autoregressive Integrated Moving Average (Manikandan and Ariarathinam 2020) and Vector Autoregression (Azme et al. 2018), an algorithm that uses Long Short-Term Memory Recurrent Neural Networks (LSTM-RNN) could also be utilised. The results obtained with each method would be evaluated through well-known metrics such as Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE). According to the existing scientific literature, all aforementioned algorithms have provided accurate projections of time series related to image processing, financial markets, climate analysis, among other fields.

An analysis of specific air and toxic pollutants emissions could also be accomplished. These gases would be black carbon (BC), metano $({\text{CH}}_4),$ carbon dioxide $({\text{CO}}_2),$ Ammonia $({\text{NH}}_3),$ nitrogen oxides $({\text{NO}}_x),$ organic carbon (OC), sulphur dioxide ${\text{SO}}_2,$ as well as volatile organic compounds (VOC).

In addition to the above, an examination of the emissions correlation of all the gases previously mentioned with some environmental parameters would be performed. These parameters could be the daily minimum/maximum near surface-air temperature, the air temperature, the surface air pressure, the sea ice area percentage on ocean grid, and precipitation. All these variables refer to the properties of the atmosphere determining the climatic characteristics.

Regarding the messages exchanged on Twitter:

A study of the existing connections between different hashtags, grasping their semantics through the examination of the main topics to which the tweets correspond could be accomplished. The modelling of topics in the tweets would be carried using Latent Dirichlet Allocation (LDA), Latent Semantic Analysis (LSA), or Probabilistic Latent Semantic Analysis (pLSA), among others (Mikolov et al. 2013). This would allow us to know which are the most debated topics on Twitter in those countries that according to our research exhibit a particular pattern of emissions.

Supplementary Information

Below is the link to the electronic supplementary material.

Author Contributions

MLM-L contributed in the literature review, in the time series analysis, in the study of the correlations, as well as in analysing the distribution of hashtags. Additionally, she developed plotting scripts. MS contributed in downloading and processing tweets, as well as in analyzing the relevance of the hashtags. All authors reviewed the manuscript.

Funding

Not applicable.

Availability of data and materials

Data repositories have been described in the “Materials and Methods”.

Declarations

Conflict of interest

The authors declare that they have no conflicts of interest.

Ethics approval

The research complied with all the relevant national regulations. We only used public tweets.

Consent to participate

There is consent from the authors to participate in the manuscript.

Consent for publication

There is consent to the publication of the manuscript.