A predictive Model for analysing Chad’s food security

ABSTRACT

Forecasting food security is important to achieve the UN’s second sustainable development goal”,Zero Hunger”. The UN identifies climate change, conflict, and COVID-19 as the three current elements of food security. So using these three dimensions to forecast food security can assist countries and organisations in making decisions and implementing humanitarian action. However, there has not been much work done on forecasting food security for specific regions in some of the most vulnerable countries in the world, like Chad. This study addresses this gap by collecting data on the climate, conflict, and COVID-19 in Chad and by utilising machine learning methods to build predictive models. We propose a feasible predicted model for food security and the study’s results offer insight into (i) food security for Chad, (ii) demonstrate how this can be achieved with a relatively small dataset and, (iii) the different features that can be used from publicly available datasets to create a reliable model.

KEYWORDS:

• Food security
• Sahel region
• Chad
• machine learning
• predictive model

1. Introduction

Achieving global food security is one of the most important goals (Griggs et al., 2013; UNSD, 2021) as food insecurity is a threat to human development. In 2020, between 7.2 and 811 million people experienced hunger in the world, an increase of 161 million from the previous year (UN, 2022). The sub-Saharan region of Africa has the highest levels of food insecurity in the world (66.2%) (UNSD, 2021). The recent COVID-19 crisis has only increased the challenge of achieving global food security and has worsened poverty and food insecurity levels due to weak political, economic, and social conditions in poorer countries. In particular, countries faced a disruption of food production and distribution chains leading to reduced access to nutritious food from other countries (Pereira & Oliveira, 2020).

There has been diverse research done in different domains towards this cause. Researchers in biotechnology have investigated how crops grow in particular geographic areas (Biber-Freudenberger et al., 2020). Researchers have also looked at funding for hunger elimination and urbanisation from a business perspective (Mason D’Croz et al., 2019). Some academics have combined their studies of food security and climate action (SDG 13) (Hansen et al., 2022). To help achieve the aim of ”Zero Hunger”, researchers have also merged technological advancements into agriculture (Ezzy et al., 2021). However, there is not been much work done on forecasting food security for specific regions in some of the most vulnerable countries in the world, like Chad.

In this paper, we explore and construct a predictive model of food security in Chad by including data on conflict, climate change (especially extreme weather), and COVID-19. We evaluate our predictive model by hindcasting and validating the model’s viability and reliability. Hindcast, also known as a historical re-forecast can integrate the model forward just like a forecast does. The food security situation in Chad is reflected by analysing the dataset along with the results of our analysis of other organisations’ data. Specifically, we include an analysis of the prices of Chadian staple foods (millet, sorghum, rice, and maize) prices, the number of conflicts in different regions, the refugee trends, the trends in annual temperature and annual precipitation, and the impact of the COVID-19 epidemic on epidemiological and economic growth in Chad.

Our work has multiple contributions to both research and practice: (i) we showcase how one can create predictive models for food security in specific countries, (ii) we demonstrate how this can be achieved if we have a relatively small dataset, and (iii) specifically, we analyse the features that are important for Chad given the public datasets we have used.

We follow the data-driven model-building methodology, as specified in (Shmueli & Koppius, 2011). In terms of algorithms, we employ five single and three ensemble models to build our predictive model. The rest of the paper is structured as follows; Section 2 describes our detailed review of published literature, Section 3 describes our Research Design including the data collection, Section 4 the Results of our building our predictive model, Section 5 Discusses the results and finally, Section 6 concludes the paper.

2. Literature review

Previous studies have broken down the issue of food security into four components (accessibility, availability, use, and stability) (Ashley, 2016). For a general literature overview of the antecedents and causes of food insecurity, please see (Su & Amrit, 2024). However, our current literature review focuses on Chad’s food security from a temporal, global and international perspective. From a chronological perspective, we summarise the different historical contexts of Chad and illustrate their relevance to food security in Chad. We then explore the literature on the current food security in Chad.

Chad is an oil-exporting and food-importing country. Therefore, the economic and food security situation in Chad is closely linked to the international oil price market and food price market. Chad is a landlocked country in north-central Africa with a population density of 54 people per square kilometre in the south to 0.1 in the vast northern desert region. Despite the start of oil production in 2003, 40% of Chad’s population continues to live in poverty. More than 65% of the population is under the age of 25, and the country has a high fertility rate and a large youth population. Even with high mortality rates and low life expectancy, the population continues to grow rapidly. As it lies at the crossroads between the Sahara desert and the Sahel region, the influx of thousands of refugees into Chad has strained the country’s resources. As a landlocked country, Chad has high transportation costs for imported goods and a high reliance on its neighbours. Chad’s economy is based on oil and agriculture. Oil accounts for roughly 60% of Chad’s export revenue, while cotton, cattle, livestock, and Gum-arabic account for most non-oil revenue. While high oil prices and local bumper harvests have supported the economy, low oil prices have put pressure on Chad’s fiscal position and significant government spending cuts. The National Development Plan (NDP) emphasises the importance of private sector participation in Chad’s development and the need to improve the business environment, particularly in high-priority sectors like mining and agriculture.

The agroecological zone of Chad is separated into three sections. From north to south, they are the Desert area, Sahelian zone, and Sudanian zone. The desert and agriculture occupy the vast majority of the country’s Sudanian zone, which has a climate wetter than the rest. The rainy season is crucial for large-scale agriculture. Farmers start preparing the soil in April and plant from June to August, a period that corresponds to more rain. Harvesting can begin in September and last until November. Despite significant domestic sorghum production, Chad’s domestic demand is met primarily through imports and food aid. Dependence on food imports and food aid explains Chad’s inability to achieve food self-sufficiency at this time. Furthermore, Chad has also increased imports and food aid since the 2000s due to the influx of refugees from the Central African Republic, Libya, Nigeria, and Sudan. As a multi-ethnic and multi-religious country, thousands of refugees have undoubtedly exacerbated Chad’s food security issue.

Conflicts in the Lake Chad region, particularly those caused by Boko Haram, have impacted 17.4 million people, worsening food security (Riebe & Dressel, 2021). In turn, the problem of food insecurity worsens instability. Many unemployed and incensed young people and adults join terrorist organisations such as Boko Haram, exacerbating the cycle of instability and food scarcity.

Chad, also suffers from climate change, particularly in the Lake Chad region. Lake Chad is Africa’s fourth-largest lake. Lake Chad has shrunk by nearly 90% since the 1960s (UNEP, 2018)due to climate change, reduced rainfall, and human activities. The Chari-Logone River, which flows through Lake Chad, is increasingly used for irrigation (Zhu et al., 2019). Burning for agricultural purposes produces dust, which contributes to desertification and reduces the amount of rainfall available to the lake. With many farmers, fishers, and pastoralists living and working on the lake, it is an essential source of food. Lake Chad was a major food export centre until 2014, with large quantities of cereals, vegetables, meat, and fish being grown, fished, or processed. However, the lake is under strain due to the region’s rapid population growth as well as political issues.

COVID-19 has had repercussions for Chad’s agriculture and food systems, including direct effects on input and labour availability, transportation and market disruptions, etc. The COVID outbreak also caused a spike in Chad’s food prices. Agyei et al., (2021) observed that the number of COVID-19 cases impacted maize, sorghum, imported rice, and domestic rice prices. However, the embargo only affected maize prices and did not affect sorghum and imported as well as local rice prices (Agyei et al., 2021).

Lastly, Chad is a typical oil-exporting country. COVID-19 has had an impact on the oil industry sector around the world. Economic stagnation, transportation restrictions, and political factors lower oil demand. Oil price declines have led to instability in countries where oil export revenues account for a significant portion of GDP. As a result, COVID-19 has further complicated Chad’s already difficult food security situation.

3. Research design

We follow the data modelling process as described in (Shmueli, 2010). Our predictive model, therefore, involves Exploratory Data Analysis (EDA), choice of variables, choice of methods, model evaluation, validation, and selection while conducting the data analysis process (Shmueli & Koppius, 2011). The data collection method employed in this study is based on secondary data.

We focus on Chad in our research for the following reasons: (i) The climate is hot and dry, as Chad is a landlocked country in the Sahel region of Africa, (ii) Chad’s agricultural output could have benefited from Lake Chad, climate change and unjustified exploitation have led to the ecological deterioration of the lake, and (iii) Politically the situation in Chad is complex and not stable. Moreover, the outbreak of COVID-19 has had a direct impact on development. For countries like Chad that depend on food imports, COVID-19 has disrupted its food supply chains, thereby worsening the food shortages.

3.1. Data collection

We began our data collection by searching The Humanitarian Data Exchange for data on food security (HDX). Food Security Data in West and Central Africa: Cadre Harmonise (CH) and Integrated Food Security Phase Classification (IPC) data are obtained in the HDX.

In this dataset, food security data for Chad from 2014 to 2022 was available. We could also find food prices in Chad, staple food prices, climate change, environmental indicators, COVID-19 vaccine doses, COVID-19 fatalities, and cumulative deaths (within African countries) using the keywords Chad and food security in HDX.

The problem with the datasets listed above is that they utilise data at the national level. However, in this research, we seek more specific predictions of the different regions of Chad to construct our prediction model. Upon searching further, we discovered CH data for the 22 Chadian regions (the CH dataset, however, excludes N’Djamena, Chad’s capital). Moreover, this dataset divided the year into three parts: January to May, June to August, and September to December. As a result, the dataset in this collection can be found with Chad’s provinces and three classifications per year.

In terms of climate change, HDX’s data is more focused on the impact of climate change on the industry in Chad. The features related to Chad are not available in detail in these datasets. Thus, we gathered data from visual Crossing (Weather Data & API, 2022), which contained weather information for specific regions of Chad. We used this website to collect data on the weather conditions for each province in Chad at the appropriate time. Natural disasters (particularly droughts and floods) caused by climate change can also significantly impact Chad’s food security. Therefore, we utilised EM-DAT (The International Disaster Database, 2022), an international disaster database, to collect geographical, temporal, human, and economic data on disasters at the national level. In terms of conflicts, HDX only has data on Chad’s conflicts up to 2019. Hence, we selected data from the Uppsala Conflict Data Program (UCDP) (Uppsala Conflict Data Program, 2022), which is systematically collected and has global coverage and comparability across cases and countries. The United Nations as well as the Food and Agriculture Organisation (FAO) utilise UCDP data as a reliable source of information.

However, concerning the COVID pandemic, the number of confirmed cases, COVID deaths, or vaccine coverage, the quantity and quality of data available were inadequate for Chad and the entire Sahel region. As a result, epidemic data was extracted and processed at the disaster level for the prediction model.

As the periods of our data were divided into three phases per year, the quantity of data was insufficient for creating our predictive model. The Chad data contained only 330 rows of tabular data and therefore training on this small dataset would most likely lead to a model that overfits .¹ We therefore included data from the G5 member countries’ sub-national regions from 2016 to 2021 (Burkina Faso, Chad, Mali, Mauritania, and Niger). These countries share the same challenges as well as climatic conditions and therefore have similar levels of food security (Benjaminsen, 2021; Dieng, 2021; Mbaye, 2020). Our final training dataset contained 1170 rows and 28 attributes in total. All variables were divided into six categories based on discipline. These were location, time, weather, disasters, conflict, and Cadre Harmonise, (2022) (food security phase classification).

3.2. Exploratory data analysis

The feature names and data types can be seen in Appendix Tables A2 and A3. In total, there were only 32 rows of missing data accounting for only 2.7 per cent of the total. We, therefore, removed these rows and we were left with 1138 rows of data.

The two main constraints in selecting features are their availability at the time of prediction and their measurement quality (Shmueli & Koppius, 2011). We sorted the variables using the information gain algorithm. We found phase5, max_sustained_wind, and total_precip information (Table A3) to be quite low. Based on the results of the two feature importance algorithms, we determined that the attributes phase5 and max_sustained_wind were of little significance, so these two variables were discarded.

Regarding disaster information, the information of disaster_group and disaster_type (Table A2) was similar. Therefore, only disaster_type, duration_months, and total_affected were retained. The type of conflict was less important than whether a conflict occurred, the number of times, and the scale of conflict regarding conflict information. Thus, these variables, including conflict, times_of_conflict, and scale_of_conflict, were retained (Table A2). We also noticed that the number of previous stages of food insecurity was significant. Phases 1 through 4 are crucial in terms of Cadre Harmonisé data. Although the Information Gain algorithm indicates that weather information is not that important, we still decided to include the variable max_daily_precip in the model. We did so as maximum daily precipitation is the highest level of precipitation in a given area over a given period. Increasing precipitation concentration and maximum daily precipitation can be expected to affect crop yield. The tables (Tables A3 and A2) are ordered by clustering the features based on the location, climate, disasters, conflict and the features of the cadre harmonise data

3.3. Creation of the predictive model

We chose phase_class as the target variable for our predictive model as it is the variable that indicates the phase of food insecurity in the particular district. As described in Tables A2 and A3, phase_class has five levels (1 indicating minimal food insecurity, and 5 indicating famine). As our target variable is nominal, we chose classification as our prediction task. The target variable, phase_class was not balanced, so we used random oversampling to balance it (Kaur et al., 2019). Among various algorithms, we chose to try typical ML tree-based models as they are known to perform well with tabular data (Grinsztajn et al., 2022; Shwartz-Ziv & Armon, 2022). We, therefore, selected Naive Bayes and k-NN representing the simple baseline models, and Decision Tree, Random Forest and gradient-boosted trees representing the tree-based models for our data analysis. To reduce the risk of overfitting the decision tree, we set the minimum leaf size to 10, which is close to 1% of the dataset. Furthermore, we used the maximum depth parameter of 10 and a minimal gain of 10.

For the Random forest algorithm, we set the minimum leaf size to 10, maximum depth to 10, and minimal gain optimisation to 10 in the same way as the decision tree algorithm. In addition, we set the number of trees to 100.

The maximum depth and number of trees in the Gradient Boosted Trees algorithm were initially optimised using similar values as the Random forest.

We also deployed ensemble algorithms by combining several learning algorithms into one prediction model. This method is generally assumed to produce better results than a single model. We used three ensemble techniques in this study: Bagging, AdaBoost, and Stacking. Just as in the regular models, we used stratified sampling for the ensemble models. Bagging, Bayesian Boosting, and AdaBoost were learned and integrated in the order of k-NN, Decision Tree, Random Forest, and Gradient Boosted Trees.

4. Results

We first evaluated and summarised each model’s performance, including accuracy, weighted average precision, and weighted mean recall. Based on these results we selected the best models that were further validated. We then performed hyperparameter optimisation on the best-performing model.

4.1. Evaluation

We selected the Naive Bayes model as our baseline model that we use to compare the other models. After parametrisation, the maximum depth of the decision tree was 10, the minimal gain was 0, and the leaf size was 10. The parameters for the best-performing random forest (RF) algorithm were 75 trees, a minimal leaf size of 10, a maximum depth of 10, and a minimal gain of 1. The best-performing parameters for Gradient Boosted Trees were a maximal depth of 10 and 100 trees. As a robust algorithm, GBT gave a favourable result.

In addition, we set the sample ratio to 0.7 and iterations to 10 for bagging. The AdaBoost iterations were also 10. Table 1 shows the specific performance of all models.

Table 1. Performance of models.

Name	Accuracy	weighted_mean_precision	weighted_mean_recall
Naive Bayes	55.43%	44.22%	43.48%
Decision Tree	67.50%	58.41%	50.89%
Random Forest	80.05%	90.55%	85.26%
k-NN	83.69%	63.17%	61.62%
Gradient Boosted Trees	92.10%	69.10%	68.93%
Bagging	88.71%	65.94%	66.92%
AdaBoost	82.69%	62.52%	61.87%
Stacking	84.46%	64.66%	62.97%

Based on the results of the models (Table 1), we can conclude that the accuracy of the ensemble algorithm was generally higher than that of a single model, except for the Gradient Boosted Trees. The Gradient Boosted Trees had the highest accuracy with 92.10%. Therefore, we selected the Gradient Boosted Tree algorithm for building our predictive model.

4.2. Validation and model selection

We found that Gradient Boosted Trees (GBT) with a maximum depth of 9 with the number of trees set to 200 was the best-performing set of hyperparameters for the algorithm. When compared to the previous run of the same algorithm, the hyperparameter optimisation showed a clear improvement. The accuracy went up from 92.10% to 95.23%, while the weighted mean precision increased from 69.10% to 71.33% and the weighted mean recall increased from 68.93% to 71.35%.

4.3. Model use and reporting

Having trained the GBT on a larger dataset, we then evaluated our algorithm by focussing on Chad’s data by using the hindcast method. The dataset for Chad contained 330 rows and 9 columns. The variables are country, region, reference_time_label, reference_year, millet, sorghum, rice(local), rice(imported) and maize. We, therefore, predict the food security in Chad in 2021 based on the food security data of Sahel countries from 2016 to 2020 (see Sec. 3.1) and then we compared the results with Chad’s actual food security in 2021.

Using this sample we achieved an accuracy of 85.22%, while the weighted mean recall was 88.87% and the weighted mean precision was 83.80%. This performance demonstrates that the optimised Gradient Boosted Trees model can be harnessed to predict Chad’s food security.

Next, we applied the hindcast method. Hindcast (or Backtesting) refers to testing a predictive model on historical data. We divided the data from 2016 to 2020 into 70% for training and 30% for testing again using stratified sampling and a subset of 11 features listed in Table A1. Upon running the GBT algorithm, we got an accuracy = 81.25%, weighted mean recall = 76.09%, and weighted mean precision = 81.05%.We then analysed the top features through a SHAP analysis (see Figure B1 in Appendix). To simplify the analysis and provide a consistent evaluation criterion, we calculated the average value of the SHAP values for all categories. By evaluating the SHAP values, we identified phase1, phase2 and phase3 as particularly important variables. In addition, conflict also showed significant influence, mainly in the frequency and scale of its impact. Regarding the disaster type, we observe differences in the impact of different hazards on the model predictions. Specifically, epidemic has the highest impact, followed by flood, while drought has a relatively low impact.

We then selected only six characteristic features, namely; conflict, disaster_type, phase 1, phase 2, phase 3, and scale_of_conflict. The model performance results were as follows. The accuracy was 82.29%, the weighted mean recall was 76.87%, and the weighted mean precision was 80.24%.

Finally, we utilised all of Chad’s data from 2016 to 2020 as the training set and Chad’s data from 2021 as the test set to apply the model. The accuracy was 80.30%, the weighted mean recall was 80.66%, and the weighted mean precision was 77.09%. All of the results are in Table 2.

Table 2. Performance of Chad.

Name	Accuracy	weighted_mean_precision	weighted_mean_recall
Performance of Chad in 2016–2021	85.22%	83.80%	88.87%
Performance of Chad in 2016–2020 with 11 characteristic variables	81.25%	81.05%	76.09%
Performance of Chad in 2016–2020 with 6 characteristic variables	82.29%	80.24%	76.87%

The hindcast results prove that the final predictive model is viable.

5. Discussion

As stated briefly in the literature review (Su and Amrit, 2024), there are four components of food security: food accessibility, availability, use, and food stability. Chad relies on food imports and therefore, food prices are closely related to Chad’s food security. The staple food of people in Chad is millet, sorghum, rice, and maize. We collected data on the price of staple food in Chad from 2017 to 2021 on HDX. In selecting variables in the predictive model, conflict is mentioned as a significant variable for predictive modelling. As a result, we developed a chart (Figure 1) of Chad’s conflict data from 2016 to 2021. The chart shows the number of conflicts that occur in each location; as seen below, Lac has the highest number of conflicts. Lac is a region in western Chad that borders Niger, Nigeria, and Cameroon. A large portion of it is located near Lake Chad. The Lac region is only 22,320 square kilometres in surface area, but it has a population of 346,000 to 556,000. Lake Chad is a vital water source with a large population living and producing around it. Thus, conflicts over water are widespread. State-based violence and one-sided violence are prevalent in these types of conflicts.

Figure 1. Number of conflicts in various regions of Chad.

Furthermore, refugees are also a problem that cannot be ignored. Chad currently has 578,842 refugees as of 31 May 2022. Chadian returnees from Lake Chad Basin are 23,901 as of 30 April 2022. According to the information from the United Nations High Commissioner for Refugees (UNHCR), Refugee Trends in Chad, 2016 marked a turning point in the influx of refugees into Chad. As seen in Figure 2, after 2016, the number of refugees in Chad was essentially above 400,000.

Figure 2. Refugees trend in Chad.

UNHCR, Government

Conflict is the most important factor threatening food security. Conflict endangers people’s lives and health and has a direct or indirect impact on their productive lives. Non-state violence is also a type of violence. In conflicts with the government and unilateral violence against civilians, armed groups, rebel groups, and terrorist organisations can seriously threaten food security. As a result, international organisations are required to assist in resolving the Chadian conflict. The issue of food security for the people of Chad can be gradually tackled if the country achieves peace and political stability.

5.1. Climate change

Climate Action and Zero Hunger are Sustainable Development Goals (SDGs). They are both long-term and complex challenges. According to data from the World Bank, the average annual temperature in Chad has been rising since 1901. As seen in Figure 3, from 1901 to 2020, the average maximum temperature in Chad occurred in May, with temperatures close to forty degrees Celsius. Chad’s minimum temperature is also above ten degrees Celsius. The average temperature ranges from twenty degrees Celsius to thirty degrees Celsius. As seen in Figure 4, in terms of precipitation, Chad’s precipitation is concentrated between June and September. June is the month with the most precipitation in Chad, but the average precipitation does not exceed 120 mm.

Figure 3. Observed average annual mean - temperature of Chad for 1901 –2020.

World Bank Group

Figure 4. Monthly climatology of min -temperature, mean - temperature.

Max – Temperature and Precipitation 1991 –2020 in Chad

World Bank Group

Climate change puts Chad’s food security in jeopardy. As a result of climate change, Chad will experience higher average temperatures, a gradual increase in temperature differences, and lower and more concentrated precipitation totals. We have mentioned the issue of maximum precipitation in a single day in choosing the variables for the prediction model. There are places where the single-day maximum precipitation is close to the monthly precipitation in this area. This abnormally high single-day maximum precipitation can cause flooding and bring disaster to residents. Droughts and floods are intertwined in Chad. These disasters not only affect the food security situation in Chad but also directly affect the people’s lives.

In addition, we also mentioned that Chad needs to gradually transform the imported food situation. Under the climatic characteristics and the influence of Chad’s unstable political environment, it is already challenging for the people to carry out agricultural activities. Compared to many other regions’ climate conditions in the world, the geographical characteristics of Chad imply that it is more vulnerable to climate change. As a result, the impact of climate change on Chadians and their food security is likely to be far more severe than what we can predict through our model.

5.2. COVID-19

World Health Organisation (WHO) calculated that within two years and a half, from 2020 to 2022, Chad has reported 7,420 confirmed cases of COVID-19, including 193 deaths. As of 5 June 2022, a total of 2,355,126 doses of vaccine had been administered (WHO, n.d.). However, considering the national context of Chad, the actual number of people affected by COVID-19 may be higher than reported.

In constructing the prediction model, we employed epidemiological data for Chad from 2016 to 2021 because of the lack of COVID-19-related data. Nevertheless, unlike the previous epidemics in Chad, COVID-19 is a global pandemic. COVID-19 has a significant impact on the productive lives of people all over the world. The export of oil is the main source of income for Chad and the country imports most of its food. Both of these economic activities are dependent on the world supply network. However, Coronavirus adversely impacted the global supply chain. As a result, Chad’s food security situation has become more severe.

Moreover, the global economic recession was triggered by the pandemic. As a result of the pandemic, international oil prices plummeted. It thus hurt Chad’s oil exports. On the other hand, COVID-19 affected Chad’s domestic production activities. Declining demand for oil exports, border closures, and social restriction measures affected the macroeconomic situation in Chad.

6. Conclusion

Food security is a complicated issue. It requires various countries to work together to face this challenge. Poverty, inequality, war, social instability, rapid population growth, and discrimination are issues that contribute to food insecurity. In analysing the food security situation in Chad, we primarily utilise Chad’s staple food prices data, conflict data, climate change data, and COVID-19 information. All of these aspects are specific to Chad’s national context and the Chadian people’s food security.

According to the hindcast in the model construction, our predictive model is reliable. Based on the data analysis and factual information about the conflict, climate change, and COVID-19 in Chad, the variables used in the prediction model can be considered closely related to Chad’s actual situation.

Most current food security prediction is based on statistical data analysis rather than predictive models. Our modelling has taken an integrated and interdisciplinary approach that comprehensively considers the multiple factors affecting food insecurity. We emphasise analysing the historical levels of past food insecurity in the region, the frequency and scale of conflict, and the potential impact of disasters such as epidemics, floods and droughts. Through this approach, we gained a deeper understanding and forecast of which factors pose the greatest threat to food insecurity, thus providing a scientific basis for relevant decision-making. In a practical sense, our predictive model can be harnessed to forecast Chad’s food security situation. This is especially important during pandemics, as international organisations face more complex challenges when evaluating food security. Our predictive model can also assist in obtaining a general understanding of Chad’s food security situation. Based on the results of the forecasting model, international organisations can also provide early assistance to Chad to improve the population’s food security.

In summary, the contribution of our work is multifold: (i) we showcase how one can create predictive models for food security in specific countries, (ii) we demonstrate how this can be achieved if we have a relatively small dataset, and (iii) specifically, we analyse the features that are important for Chad given the public datasets we have used.

Disclosure statement

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

Notes

1. It might overfit as we do not have enough data to train and also test, and in our case, we cannot use transfer learning – as the data is specialised and because of the tabular nature of the data.

Appendix A.

Categorical and Numerical Features

Table A1. Target variable and characteristic features.

No.	Variable Name	Type
1.	phase_class	nominal (the only target variable)
2.	max_daily_precip	real
3.	disaster_type	nominal
4.	duration_months	real
5.	total_affected	integer
6.	conflict	nominal
7.	times_of_conflict	integer
8.	scale_of_conflict	nominal
9.	phase1	integer
10.	phase2	integer
11.	phase3	integer
12.	phase4	integer

Table A2. Categorical features.

No.	Variable Name	Description
1.	Country	Location information. G5 member countries name.
2.	Region	Location information. The names of the G5 member countries’ sub-national regions. Information on sub-national regions does not correspond to each country’s province. It includes 13 regions in Burkina Faso, 22 in Chad, 9 in Mali, 13 in Mauritania, and 8 Niger. There are 65 subnational regions in total.
3.	Region_pcod	Location information. Sub-national area codes.
4.	Reference_year	Time information from 2016 to 2021. The total is six years.
5.	Reference_time_label	The time information is divided into three parts: January to May, June to August, and September to December.
6.	Reference_time_code	Time information.
7.	Disaster_group	Disasters information. The five classifications are biological, Climatological, Hydrological, Meteorological, and N (no disasters).
8.	Disaster_type	Disasters information. It includes six categories: Drought, Epidemic Flood, Landslide, N (no disaster), and Storm.
9.	Conflict	The binary classification, where Y and N represent conflict occurrence and absence.
10.	Type_of_violence	Conflict information. There are four categories in total”.0” indicates no violence”,1” means state-based violence”,2” implies non-state violence, and”3” represents one-sided violence.
11.	Scale_of_conflict	Conflict information. It is divided into three categories”.0” denotes no conflicts”,1” represents skirmishes and clashes (deaths ranging from 10 to 99), and”2” implies minor conflicts (deaths between 100 and 999).
12.	Phase_class	Cadre Harmonisé information. There are a total of 5 levels from Phase1 to Phase5. This variable is the target variable It is significant to note that local food security under each sub-national area is different. According to Cadre Harmonisé map, the stage of food security in subnational areas usually coincides with the most severe areas in their subordinate regions. For instance, the region’s food security classification is not Phase 1 even though a significant portion of its population is in Phase 1. The departments in this region may have a high proportion of Phase 3 residents and these people experience more severe food insecurity. Then the Phase 3 is the final classification of this region.

Table A3. Numeric features.

No.	Variable Name	Description
13.	Max_temp	Location information. G5 member countries name.
14.	Min_temp	Weather information. The minimum temperature is given in degrees Celsius.
15.	Total_precip	Weather information. The total precipitation is measured in millimetres.
16.	Max_daily_precip	Weather information. Maximum daily precipitation is measured in millimetres.
17.	Rain_days	Weather information. It implies the number of days with rain.
18.	Max_sustained_wind	Weather information. The unit is kilometres per hour (kph)
19.	Duration_months	Duration of the disaster information.
20.	Total_affected	The total number of people affected by the disaster.
21.	Times_of_conflict	The number of conflict occurrences.
22.	Population	Cadre Harmonisé information. Total population.
23.	Phase1	Cadre Harmonisé information. Phase1 represents the number of people with minimal food and nutrition insecurity.
24.	Phase2	Cadre Harmonisé information. Phase2 represents the number of people with stressed food and nutrition insecurity.
25.	Phase3	Cadre Harmonisé information. Phase3 represents the number of people with crisis food and nutrition insecurity.
26.	Phase4	Cadre Harmonisé information. Phase4 represents the number of people with emergency food and nutrition insecurity.
27.	Phase5	Cadre Harmonisé information. Phase5 represents the number of people with famine food and nutrition insecurity.

Appendix B.

SHAP Analysis

Figure B1. SHAP analysis of the top features.