Classification of food security status in disaster-prone areas: a comparative analysis of machine learning models using Support Vector Machine, Random Forest, and Logistic Regression

Classificação do status de segurança alimentar em áreas suscetíveis a desastres: uma análise comparativa de modelos de aprendizado de máquina utilizando Máquina de Vetores de Suporte, Floresta Aleatória e Regressão Logística

Abstract

Food security assessment in disaster-prone areas requires accurate and adaptive approaches capable of capturing dynamic production-consumption relationships. This study aims to develop and evaluate machine learning-based classification models for determining regional food security status using a Food Security Index (FSI), defined as the ratio of local food production to annual consumption requirements with an added safety buffer. Three algorithms were compared: Support Vector Machine (SVM), Random Forest (RF), and Logistic Regression (LR). Model performance was evaluated using k-fold cross-validation and assessed through accuracy, precision, recall, F1-score, and confusion matrix analysis. The results show that SVM achieved the highest and most consistent performance, with accuracy ranging from 92% to 96%, precision of 0.93, recall of 0.94, and an F1-score of 0.93. Random Forest demonstrated moderately strong performance (accuracy: 88-92%; F1-score: 0.87), while Logistic Regression exhibited the lowest performance (accuracy: 80-85%; F1-score: 0.80), indicating its limitation in modeling non-linear relationships. Confusion matrix analysis reveals that SVM maintains stable classification across all validation folds with minimal misclassification, whereas errors in all models are predominantly concentrated in regions with FSI values near the classification threshold (FSI ≈ 1), highlighting the challenge of boundary-sensitive cases. Substantively, the integration of production-consumption ratios with buffer reserves in the FSI effectively captures both food availability and system resilience, distinguishing surplus regions (FSI ≥ 1) from structurally vulnerable regions (FSI < 1). These findings confirm that SVM is the most suitable approach for food security classification in disaster-prone areas due to its ability to model non-linear patterns and optimize decision boundaries. This study provides a robust and scalable framework for data-driven food security monitoring and early warning systems. The results support the use of SVM-based classification to enhance the accuracy of vulnerability detection and to improve the targeting of policy interventions in disaster-prone regions.

Keywords:
Food Security Index; machine learning; Support Vector Machine; Random Forest; disaster-prone areas; classification; early warning system

Resumo

A avaliação da segurança alimentar em áreas suscetíveis a desastres requer abordagens precisas e adaptáveis, capazes de capturar as relações dinâmicas entre produção e consumo. Este estudo visa desenvolver e avaliar modelos de classificação baseados em aprendizado de máquina para determinar o status de segurança alimentar regional, utilizando um Índice de Segurança Alimentar (ISA), definido como a razão entre a produção local de alimentos e as necessidades de consumo anual, com uma margem de segurança adicional. Três algoritmos foram comparados: Support Vector Machine (SVM), Random Forest (RF) e Regressão Logística (RL). O desempenho do modelo foi avaliado por meio de validação cruzada k-fold e mensurado com base em acurácia, precisão, sensibilidade (recall), pontuação F1 e análise da matriz de confusão. Os resultados mostram que a SVM alcançou o desempenho mais alto e consistente, com acurácia variando de 92% a 96%, precisão de 0,93, recall de 0,94 e pontuação F1 de 0,93. O Random Forest obteve um desempenho moderadamente elevado (precisão: 88-92%; pontuação F1: 0,87), enquanto a Regressão Logística apresentou o desempenho mais baixo (precisão: 80-85%; pontuação F1: 0,80), indicando sua limitação na modelagem de relações não lineares. A análise da matriz de confusão revelou que o SVM manteve uma classificação estável em todas as validações com erros mínimos de classificação, enquanto as falhas em todos os modelos concentraram-se em regiões com valores de ISA próximos ao limite de classificação (ISA ≈ 1), destacando o desafio de casos sensíveis a limites. Do ponto de vista substantivo, a integração das taxas de produção-consumo com reservas de segurança no ISA mostrou-se eficaz para captar tanto a disponibilidade de alimentos quanto a resiliência do sistema, distinguindo regiões com excedente (ISA ≥ 1) de regiões estruturalmente vulneráveis ​​(ISA < 1). Essas descobertas confirmam que o SVM é a abordagem mais adequada para a classificação da segurança alimentar em áreas suscetíveis a desastres, devido à sua capacidade de modelar padrões não lineares e otimizar os limites de decisão. Este estudo fornece um quadro robusto e escalável para o monitoramento da segurança alimentar e sistemas de alerta precoce baseados em dados. Os resultados corroboram a utilização da classificação baseada em SVM para aumentar a precisão da detecção de vulnerabilidades e melhorar o direccionamento das intervenções políticas em regiões suscetíveis a desastres.

Palavras-chave:
Índice de Segurança Alimentar; aprendizado de máquina; Support Vector Machine; Random Forest; áreas suscetíveis a desastres; classificação; sistema de alerta precoce

1. Introduction

Food security is an integral component of sustainable agricultural development and a fundamental issue that significantly affects the socio-economic stability of communities, particularly in Indonesia and more specifically in West Java. Studies on food consumption are motivated by diverse objectives and closely interact with food security dynamics (Ioannidis, 2018; Penders et al., 2017), and constitute an integral part of broader dynamic processes within food systems (Doherty et al., 2022). Food security remains a critical global challenge, with approximately 735 million people worldwide experiencing chronic hunger in 2022, representing a substantial increase compared to pre-pandemic levels (FAO, 2024). This condition is especially severe in disaster-prone regions, where natural hazards such as floods, droughts, earthquakes, and volcanic eruptions can rapidly disrupt food production and distribution systems (Béné et al., 2024). The intersection of climate change and the increasing frequency of disasters has imposed unprecedented pressure on regional food systems, thereby necessitating more advanced approaches for food security monitoring and early warning systems (Westerveld et al., 2021).

Historically, food consumption forecasting has relied heavily on conventional statistical approaches, which often struggle to capture complex temporal patterns and the variability of factors influencing consumption behavior (Fattah et al., 2018). In this context, food security classification provides a critical foundation for the development of data-driven early warning systems designed to detect and anticipate food insecurity (Reddy et al., 2025).

Food security encompasses four fundamental dimensions: availability, access, utilization, and stability (FAO, 2024). Among these dimensions, food availability—defined as the physical sufficiency of locally available food quantities—serves as a core pillar upon which the other dimensions depend. Food availability is particularly vulnerable to production disruptions, supply chain interruptions, and sudden demand surges in disaster-prone areas (Busker et al., 2024). Consequently, assessing regional food security, especially in West Java, Indonesia, requires accurate methodologies capable of capturing the dynamic relationship between local food production capacity and population consumption needs.

Conventional approaches to food security assessment predominantly rely on survey-based methods, expert consensus, and statistical indicator systems such as the Integrated Food Security Phase Classification (IPC) framework (IPC, 2026). While these approaches provide valuable insights, they are often constrained by significant time lags between data collection and analysis, high implementation costs, and limited spatial granularity (Martini et al., 2022). Moreover, the complex and non-linear relationships among food production, consumption, and diverse socio-economic factors pose substantial challenges for traditional statistical methods (Deléglise et al., 2022). In Indonesia, food security is a strategic priority within the National Medium-Term Development Plan (RPJMN) 2020-2024, which targets reductions in undernutrition prevalence and improvements in the Desirable Dietary Pattern (DDP) score (BPN, 2023). Nevertheless, substantial inter-provincial disparities in production capacity and food insecurity levels persist, particularly in regions highly exposed to natural disasters (Saragih and Wibowo, 2025).

Advances in information technology and artificial intelligence have created new opportunities for food security analysis. Machine Learning (ML) techniques have emerged as promising tools for food security assessment and prediction. ML algorithms are capable of processing large volumes of heterogeneous data, identifying complex patterns, and generating accurate classifications without requiring explicit specification of underlying relationships (Kshirsagar et al., 2020). Recent studies have demonstrated the potential of various ML approaches in food security applications, including crop yield prediction, food price forecasting, and vulnerability mapping (Nakalembe et al., 2021).

Several ML algorithms have shown strong potential for classification tasks relevant to food security assessment. Support Vector Machines (SVM) have demonstrated outstanding performance in both binary and multi-class classification problems, particularly when dealing with high-dimensional data and non-linear decision boundaries (Kuri-Cervantes et al., 2020). Their ability to construct optimal separating hyperplanes that maximize class margins makes SVMs well-suited for distinguishing between food-secure and food-insecure regions (Amaya-Tejera et al., 2024). Random Forest (RF), an ensemble learning method that combines multiple decision trees, offers robust classification performance with inherent resistance to overfitting and the ability to handle both continuous and categorical variables (Alu’datt et al., 2024). Logistic Regression (LR), although computationally simpler, provides interpretable probability estimates but assumes linear relationships between predictors and outcomes (Hosmer Junior et al., 2013).

Despite these advances, the application of ML for regional food security classification—particularly in disaster-prone areas of West Java—remains relatively underexplored. Most existing studies focus on national or global scales, often overlooking sub-national heterogeneity that is critical for targeted interventions (Martini et al., 2022). Furthermore, comparative evaluations of multiple ML algorithms for food security classification using production-consumption-based indices are still scarce in the literature.

In response to these gaps, this study aims to evaluate the effectiveness of machine learning algorithms in classifying the food security status of regions in West Java based on a Food Security Index derived from local food production and consumption data. Specifically, the study seeks to identify which ML algorithm provides the most accurate and reliable classification for disaster-prone areas and to examine how key factors influence classification performance and food security monitoring strategies. The main contributions of this study include: (1) the development of a Food Security Index methodology that incorporates appropriate buffer reserves for disaster-prone regions; (2) a comprehensive comparative evaluation of three ML algorithms for food security classification; and (3) practical recommendations for implementing machine learning-based food security early warning systems.

2. Research Methods

This study employs local food commodity production and consumption data collected from multiple disaster-prone districts in West Java. The dataset was sourced from the ‘Production and Consumption Data of Disaster-Prone Regions’ database, comprising annual records of food production volumes, population consumption levels, and detailed commodity-specific attributes at the administrative unit level.

2.1. Food security index estimation

Σ (Commodity consumption i in the year of t)

• Safety Reserve= 10% of Annual Consumption (for low vulnerability areas)

• Safety Reserve = 25-30% from Annual Consumption (for high disaster prone areas)

The safety stock component is critical for disaster-prone areas because it takes into account potential production failures, distribution disruptions, and sudden increases in demand during emergency situations (Béné and Devereux, 2023). The classification threshold is defined as:

• Food Security (IKP ≥ 1,0): Local production meets or exceeds consumption needs plus reserves

• Food Vulnerability (IKP < 1,0): Local production is not enough to meet consumption needs plus reserves.

3. Machine Learning Algorithms

3.1. Support Vector Machine (SVM)

Support Vector Machine (SVM) is a supervised machine learning method used to classify data into two or more distinct groups. The approach operates by identifying an optimal separating boundary, referred to as a hyperplane, which maximizes the margin between different classes. This hyperplane is constructed such that observations from one class lie on one side of the boundary, while observations from the opposing class lie on the other. The data points located closest to the separating hyperplane are known as support vectors, and these critical observations play a decisive role in defining the position and orientation of the hyperplane (Figure 1).

Figure 1
Support Vector Machine Classification.

SVM non-linear classification hyperplane (RBF) is denoted (Equation 1):

$f\left(x\right)=W^T.x+B$ (1)

so that according to Cortes and Vapnik (1995), the equation obtained is (Equations 2 and 3):

$\left[\left(W^T.x_i\right)+b\right]\ge 1for{y}_i=+1$ (2)

$\left[\left(W^T.x_i\right)+b\right]\le -1for{y}_i=-1$ (3)

Let 𝑥_i denote the training dataset, i = 1,2,...n, and y_i = epresent the class label associated with 𝑥_i. The optimal hyperplane is obtained by identifying the hyperplane that lies equidistant between the two class-separating boundaries. Determining the optimal hyperplane is equivalent to maximizing the distance, or margin, between two sets of observations belonging to different classes. As demonstrated in the study by Oktaviani and Putrada (2022), the margin can be derived using the following formulation (Equation 4):

$\frac{2}{\left|\left|w\right|\right|}$ (4)

annotation: w is a weight vector that determines the orientation of the hyperplane; ||w|| is the norm or length of the vector w (i.e. the magnitude of the vector).

Margin is the distance between two support vectors from two different classes. Where the smaller the ||w||, means margin$\frac{2}{\left|\left|w\right|\right|}$getting higger, and that's what SVM wants. To find the best hyperplane in a Support Vector Machine, a method called the Quadratic Programming (QP) Problem is used. This method aims to minimize the value of $\frac{1}{2}{W}^{T}W$, yis half the length of the square of the weight vector w. This w vector determines the direction and slope of the dividing line (hyperplane) between two data classes. The smaller the value of ∥w∥, the greater the margin or separation distance between the two data groups, which means the separation between the classes becomes clearer and safer from errors. Therefore, by minimizing $\frac{1}{2}{W}^{T}W$, SVM automatically finds the hyperplane that separates two classes with the maximum margin. This process also includes constraints that ensure each data point falls on the correct side of the dividing line. In this way, QP helps SVM produce a robust classification model that performs well on new data.

SVMs are highly effective when applied to high-dimensional data (referring to data with many features or variables) and complex classification patterns. This method excels at clearly separating two classes, especially when there is a large margin between them. Furthermore, SVMs can handle high-dimensional data and can function well using kernels to handle data that is not linearly separable. However, using SVMs also requires an important parameter called the regularization parameter C, which serves to regulate the trade-off between larger margins and misclassifications on the training data. The C value controls how much the model penalizes misclassifications. If the C value is too large, the model will focus more on reducing misclassifications, thus attempting to separate the data very precisely. This can make the model more complex and susceptible to overfitting. Conversely, if the C value is small, the model will prioritize wider margins and accept some misclassifications, resulting in a simpler model that is more resistant to overfitting, but may be less precise in separating the data. Selecting the appropriate C value requires experimentation and adjustment based on the characteristics of the data used. Although SVM can be applied to multi-class classification, this makes its implementation more complicated and requires additional approaches such as one-vs-one or one-vs-all methods (Xiao et al., 2019).

3.2. Random Forest (RF)

Random Forest (RF) is an ensemble method that combines multiple decision trees built from randomly selected data samples and features. The final prediction of this model is determined through voting for classification or averaging for regression. This technique has proven effective in improving accuracy and reducing the risk of overfitting, which often occurs with a single decision tree. To mitigate the problem of overfitting on small training data, a combination of bootstrap aggregation and random feature selection in a random forest can be used. Because Random Forest is a CART ensemble method, it has no assumptions and can be used in nonparametric cases (Mulyahati, 2020) (Figure 2).

Figure 2
Gambar 2. Random Forest (RF).

Classification is input data x, and there are N tree factors (Equation 5):

${}_y{}^{{}^{^}}=mode(T_1\left(x\right),T_2\left(x\right),\dots ,T_N\left(x\right))$ (5)

annotation: T1​(x), T2​(x),…, TN​(x): It is the prediction of each model (e.g., decision tree) regarding input x. So, N models each provide an opinion or prediction regarding the input; Mode (…): Mode function, namely the value that appears most frequently from the set of predictions; ${}_y{}^{^},$: It is the final prediction of the ensemble, namely the result of the majority voting of all models.

The final prediction in a classification ensemble (random forest) is the result of voting from all individual models, and the most voted value becomes the final result. Regression (prediction of numerical values), then (Equation 6):

${}_y{}^{^}=$​ $\frac{1}{N}$=${{\displaystyle \sum }}_{i=1}^{n}Ti\left(x\right)$ (6)

annotation: x: Feature data; y^{^}: Label; N: Number of models; Ti: every model to-i.

At the initial stage of constructing a Random Forest model, the algorithm identifies the optimal variable to be used as the first splitting criterion in the decision tree. This variable, referred to as the splitting variable, is selected based on its ability to partition the data into increasingly homogeneous subsets. In regression-based Random Forest models, one of the primary criteria employed for selecting the splitting variable is the Mean Squared Error (MSE). MSE measures the average of the squared differences between the observed (actual) values and the predicted values. A lower MSE indicates that the predictions more closely approximate the true values, implying a more effective data partition. Consequently, during the split selection process, the Random Forest algorithm chooses the variable that yields the minimum MSE, as this criterion provides the optimal data separation. Mathematically, the MSE is defined as follows (Equation 7):

MSE_n = $\frac{1}{N}$ ${{\displaystyle \sum }}_{i=1}^n\left(Y1-Yn\right)2$ (7)

annotation: MSE_n: Mean Squared Error (MSE) value of the n-th tree; N: Number of samples in the n-th tree; Y_i: Value of the i-th sample in the n-th tree; Y_n: Mean value of samples in the n-th tree.

The primary advantage of the Random Forest algorithm lies in its ability to effectively process datasets containing diverse feature types, including both numerical variables and categorical attributes (such as labels, classes, or groupings). This model is particularly well-suited for capturing complex and non-linear relationships among features within the data. As an ensemble method composed of multiple decision trees rather than a single classifier, Random Forest exhibits strong robustness and reduced susceptibility to overfitting, a condition in which a model fits the training data excessively and performs poorly on unseen data.

Despite these strengths, Random Forest also presents certain limitations. One notable drawback is its limited interpretability, as the ensemble structure involving numerous interacting decision trees makes the model more difficult to explain compared to a single decision tree. Additionally, when the number of trees is very large, prediction and computation processes may become relatively slow and demand higher computational resources than simpler models (Syukron and Subekti, 2018).

3.3. Logistic Regression

Logistic Regression is a statistical method used to estimate the probability of the occurrence of an event with two possible outcomes, such as “yes” or “no,” or “positive” or “negative.” This model is applied when the response (output) variable is binary in nature, meaning that it assumes only two possible states (Bimantara and Dina, 2019). Logistic Regression operates by modeling the relationship between a set of input variables (independent variables) and the outcome of interest (dependent variable) through a mathematical function known as the sigmoid or logit function. This function maps the linear combination of predictors to values bounded between 0 and 1, which are interpreted as the probability of the event occurring.

When the estimated probability approaches 1 (commonly exceeding a threshold of 0.5), the model classifies the event as occurring (denoted as Y=1). Conversely, when the probability approaches 0, the event is classified as not occurring (Y=0). Given the binary nature of the outcome, the response variable Y is assumed to follow a Bernoulli distribution, which characterizes random variables with exactly two possible outcomes (Oktaviani and Putrada, 2022). Mathematically, the relationship between the input variables and the predicted probability is expressed as (Equation 8):

$P\left(Y=1\right)=\frac{1}{1+e^{-z}}$ (8)

where $z={\beta }_0+{\beta }_1{x}_1+{\beta }_2{x}_n+\dots +{\beta }_n{x}_n$

annotation: $P\left(Y=1\right)$ is the probability that the response variable Y has a value of 1, that is, the observed event occurs; $e$ is an exponential number (approximately 2.718), which is the base of the natural logarithm; $z$ is a linear function of the predictor variables, which determines the shape of the logit function; ${\beta }_0$is the intercept or constant, namely the value of 𝒛 when all predictor variables are zero; ${\beta }_1,{\beta }_2,\dots ,{\beta }_n$​ is the regression coefficient, which shows the influence of each variable $x_1,x_2,\dots ,x_n$​ against the log-odds of events; $x_1,x_2,\dots ,x_n$​ are independent or predictor variables; $n$ is the total number of predictor variables in the model.

The above function is called the sigmoid function, which converts z-values ​​into probabilities. In contrast, the logit function writes this relationship as (Equation 9):

$logit\left(p\right)=\log (\frac{p}{1-p})$ (9)

annotation: $logit\left(p\right)$is the logarithm of the odds, namely the transformation of p probability into log-odds form; $p$ is the probability that an event occurs, namely P(Y=1), with a value between 0 and 1; $1-p$ is the probability that the event does not occur, namely P(Y=0); $\frac{p}{1-p}$ known as odds, namely the comparison between the probability of an event occurring and the probability of the event not occurring; $log$ is the natural logarithm (base $e$).

This model is well-suited for situations in which the data are linearly separable, meaning that a straight line (or a separating hyperplane in higher-dimensional spaces) is sufficient to distinguish between classes. However, logistic regression is less effective in modeling data with non-linear or complex patterns, as the decision boundary it produces is inherently linear (Bimantara and Dina, 2019). In addition, to achieve optimal model performance, the features within the dataset should be normalized or standardized, particularly when there are substantial differences in scale across variables. Although logistic regression is fundamentally designed for binary classification, it can also be extended to multi-class classification using strategies such as the one-vs-rest (OvR) approach, in which each class is iteratively compared against all remaining classes (Table 1, Figure 3).

Table 1
Comparison of Support Vector Machine, Random Forest, and Logistic Regression

Criteria	Support Vector Machine	Random Forest	Logistic Regression
Model Type	Margin-based classification model (best line separation)	Ensemble model based on multiple decision trees	Probability-based statistical models
Data Types	Suitable for numeric data, can be used for non-linear data	Suitable for both numeric and categorical data	Suitable for linear and binary (two-class) data
Overfitting Ability	Tends to overfitting if the data is not standardized and there is a lot of noise	Relatively resistant to overfitting due to the combination of many trees	Prone to overfitting if the data is too complex and non-linear
Excess	- Accurate on complex data	-Resistant to overfitting	- Easy to understand
			- Quick to practice
	- Effective for high-dimensional data	-Can handle large and mixed data	- Results can be interpreted
Weakness	- Difficult to interpret	- Results are difficult to explain	- Less suitable for non-linear data
	- Features require scaling
		- Prediction time can be long if there are many trees	- Sensitive to outliers

Figure 3
Logistic Regression Classification.

3.4. Confusion matrix

A confusion matrix is ​​a table that shows how a model or algorithm works. Each row represents the actual data class and each column represents the predicted data class, or vice versa (Table 2).

Table 2
Confusion matrix.

	Predicted Negative	Predicted Positive
Actual Negative	True Negative (TN)	False Positive (FP)
Actual Positive	False Negative (FN)	True Positive (TP)

Annotation:

1. True Positive = meaning how much of the class data is actually positive and the model prediction is also positive.

2. True Negative = meaning how much of the actual class data is negative and the model prediction is also negative.

3. False Positive = This shows how much of the class data is actually negative, but the model predicts positive.

4. False Negative = This shows how much of the actual class data is positive, but the model predicts negative.

Saputro and Sari (2020) propose four key performance metrics for evaluating machine learning models, which provide robust and reliable measures of model performance, including:

1. Accuracy

Accuracy is a performance metric that indicates how frequently a model correctly classifies the data, encompassing both positive and negative classes. It represents the proportion of correctly classified instances relative to the total number of evaluated instances. A higher accuracy value reflects a greater ability of the model to produce predictions that are consistent with the actual observed conditions. The accuracy metric can be mathematically expressed using the following Equation 10:

$\frac{TP+TN}{Total}$ (10)

1. Precision

Precision is a performance metric that measures how often the model’s predictions are correct when it assigns a data instance to a particular class (e.g., the positive class). Precision reflects the model’s level of confidence in labeling an instance as belonging to the predicted class. A higher precision value indicates that a larger proportion of instances predicted as belonging to a given class are indeed true members of that class. The precision metric can be computed using the following Equation 11:

$\frac{TP}{FP+TP}$ (11)

1. Recall (sensitivity/original positive rate)

Recall is a performance metric that measures the ability of a model to correctly identify positive instances among all actual positive cases. It reflects the model’s sensitivity and its effectiveness in detecting or capturing all existing positive cases. A high recall value indicates that the model is sufficiently sensitive and minimizes the number of missed positive instances, thereby fulfilling the objective of comprehensive detection. The recall metric can be formally expressed using the following Equation 12:

$\frac{TP}{FN+TP}$ (12)

1. F1-score

The F1-score is a performance metric that combines precision and recall into a single measure. It is computed as the harmonic mean of precision and recall, thereby providing a balanced assessment of both metrics. The F1-score is particularly useful in situations where a trade-off between the model’s ability to correctly identify positive instances (precision) and its capacity to capture as many relevant positive instances as possible (recall) is required. The F1-score can be formally expressed by the following Equation 13 (Figure 4):

Figure 4
Illustration of Machine Learning Model Metrics.

$\frac{precision*recall}{precision+recall}$ (13)

4. Result

4.1. Distribution of Food Security Index

The analysis of the Food Security Index (FSI) across all regencies in West Java reveals substantial heterogeneity in regional food security status. Regencies with FSI values significantly exceeding 1.0 demonstrate surplus production capacity, indicating their potential role as local food supply buffers for neighboring deficit areas. In contrast, regencies with FSI values below 1.0 exhibit production deficits, necessitating external food supplies to meet local consumption demands.

Temporal analysis further reveals year-to-year fluctuations in both food production and consumption, reflecting the combined influences of climate variability, market conditions, and population dynamics. These fluctuations underscore the critical importance of establishing adequate buffer stocks, particularly in disaster-prone regions, to enhance regional food system resilience (Figure 5).

Figure 5
Food Security Index for each district.

Based on the processing of annual local food production and consumption data by commodity, the Food Security Index (FSI) was calculated for each district as the ratio between total food production and the annual food requirement threshold. The requirement threshold was defined as the average annual consumption plus a 10 percent buffer reserve, which was incorporated to capture potential uncertainties such as production shocks, distribution constraints, and natural disasters. This formulation allows the FSI to more accurately reflect not only static food adequacy but also the resilience of each district’s food system.

The analytical results demonstrate a clear spatial disparity in food security levels across districts in West Java. Districts such as Sukabumi, Cianjur, Garut, and Subang exhibit FSI values significantly greater than one (FSI ≥ 1), indicating that local production exceeds consumption requirements even after incorporating the buffer reserve. This surplus condition suggests a structurally strong production capacity and relatively high resilience to short-term shocks. Moreover, these districts have the potential to function as regional supply buffers, supporting adjacent deficit areas, which highlights their strategic role within the regional food distribution network.

In contrast, districts with FSI values approaching one, such as Ciamis, show a marginal balance between production and consumption. Although these areas meet the minimum adequacy threshold, the results indicate a low margin of safety. This implies that even minor disturbances—such as climate variability, yield reduction, or distribution inefficiencies—could quickly shift these districts into a deficit condition. Therefore, the analysis suggests that food security in these areas is conditional and requires policy attention, particularly in strengthening food reserves, improving supply chain efficiency, and enhancing adaptive capacity to external shocks.

Furthermore, several districts, including Tasikmalaya, Karawang, Kuningan, Pangandaran, Cirebon, and Sumedang, record FSI values below one (FSI < 1), indicating that local production is insufficient to meet annual consumption needs. This deficit reflects structural limitations in production capacity and a high reliance on external food inflows. The results imply that these districts are more vulnerable to supply disruptions, price volatility, and disaster impacts, which can directly affect food availability and access at the household level. Consequently, the findings underscore the need for targeted interventions, such as increasing local production, improving inter-district distribution systems, and developing strategic food reserve mechanisms to reduce vulnerability and enhance overall food system resilience.

The regional food security analysis in this study is based on annual, commodity-level local food production and consumption data sourced from Sheet 1 of the file “Data Produksi dan Konsumsi Daerah Rawan Bencana.xlsx.” The dataset contains information on the production and consumption of multiple food commodities across districts over several observation years. To derive annual food security conditions, commodity-level data were first aggregated to obtain total food production and total food consumption for each district and year. Subsequently, the annual food requirement threshold was defined as the average annual consumption supplemented by a 10 percent reserve as a minimum benchmark to ensure food availability.

Based on these calculations, the Food Security Index was obtained as the ratio between total production and the annual food requirement threshold. This index served as the primary variable in the classification process to determine regional food security status, categorized as food-secure or food-insecure. The classification was conducted using three machine learning methods—Random Forest, Logistic Regression, and Support Vector Machine (SVM)—with model performance evaluated using confusion matrices across multiple cross-validation folds (Figure 6).

Figure 6
Visualization of the confusion matrix in the Random Forest method.

The confusion matrix visualization results for the Random Forest method aim to evaluate the model's accuracy and consistency in classifying food security status in each test fold. Based on Figure 6, in Fold 1, the model demonstrated excellent performance, with all data in classes L1 and L2 correctly classified (two data points each without errors). In Fold 2, there was a slight decrease in performance, with one data point in class L1 being incorrectly classified into class L2, indicating misclassification of data at the classification boundary. Meanwhile, in Fold 3, the model again demonstrated optimal performance, with all data points correctly classified without errors.

Overall, these results demonstrate that Random Forest has excellent and relatively stable performance in classifying data, especially in classes with clear patterns. However, the errors that appeared in Fold 2 indicate that the model still faces challenges in classifying data with characteristics close to the interclass threshold (Figure 7).

Figure 7
Logistic Regression confusion matrix.

In the Logistic Regression model, the confusion matrix results indicate a higher tendency for misclassification compared to the other methods. This outcome is primarily attributable to the linearity assumption inherent in Logistic Regression, whereas the relationships among food production, consumption, and food security in local food system data are inherently more complex and non-linear. Consequently, regions with Food Security Index values located near the annual requirement threshold are more likely to be misclassified, thereby reducing the overall predictive accuracy of the model (Figure 8).

Figure 8
Support Vector Machine (SVM) Confusion matrix.

In contrast to the other two methods, the Support Vector Machine (SVM) demonstrated the most stable and consistent performance across all validation folds. The confusion matrix of the SVM model indicates that the majority of observations were correctly classified, with only a minimal number of misclassifications. The superior performance of SVM can be attributed to its ability to construct an optimal hyperplane that maximizes the margin between classes, enabling a clearer separation between food-insecure and food-secure regions, even when food security index values are close to the annual requirement threshold.

The accuracy level achieved by the SVM-based index further confirms the effectiveness of this method in interpreting regional food security conditions using local commodity-level food production and consumption data on an annual basis. The incorporation of an annual food requirement threshold plays a critical role in enhancing the model’s discriminatory power, as it provides a well-defined reference point for determining whether a region is in a food surplus or deficit condition. Consequently, SVM not only delivers high classification accuracy but also offers stronger interpretability with respect to regional food security status.

Overall, the comparative analysis of the three models indicates that the Support Vector Machine (SVM) is the most optimal approach for classifying regional food security based on local food data. Random Forest ranks second with reasonably strong performance, while Logistic Regression exhibits limitations in capturing non-linear data patterns. These findings underscore that the application of SVM, combined with a food security index framework and an annual requirement threshold, is highly suitable for supporting decision-making in food security planning and management, particularly in disaster-prone regions.

4.2. Classification model performance evaluation

Model performance was evaluated using several key metrics, namely accuracy, precision, recall, and the F1-score. These metrics were employed to provide a comprehensive assessment of the model’s ability to classify regions into food-secure and food-insecure categories. Accuracy represents the proportion of instances that were correctly classified. Precision reflects the model’s reliability in predicting a specific class, indicating the extent to which predicted positive instances are truly positive. Recall measures the model’s capacity to correctly identify all instances that genuinely belong to a given class. The F1-score, defined as the harmonic mean of precision and recall, provides a balanced measure that accounts for both false positives and false negatives (Table 3).

Table 3
Comparison of Support Vector Machine (SVM) methods.

Classification Method	Accuracy (%)	Precision	Recall	F1-score
Random Forest	88-92	0.88	0.87	0.87
Logistic Regression	80-85	0.81	0.80	0.80
Support Vector Machine (SVM)	92-96	0.93	0.94	0.93

Based on the comparative results, the Support Vector Machine (SVM) method demonstrates the best overall performance compared to Logistic Regression and Random Forest. SVM achieves the highest accuracy, precision, recall, and F1-score, indicating its superior capability to classify regional food security status in a consistent and reliable manner. This performance can be attributed to SVM’s effectiveness in handling non-linear data structures and its ability to construct optimal classification margins.

Random Forest ranks second, exhibiting strong performance, particularly in capturing complex interrelationships among variables. However, variations observed across several cross-validation folds suggest that this method remains sensitive to data distribution. In contrast, Logistic Regression records the lowest performance among the three approaches, highlighting its limitations in modeling food security patterns that are not strictly linear (Figure 9).

Figure 9
Annual Food Production and Consumption Graph.

Based on the analysis of annual food production and consumption trends, it can be observed that local food availability for each commodity and across regions exhibits considerable year-to-year fluctuations. The disparity between production volumes and consumption levels reflects the capacity of each region to meet its food requirements independently. In certain years, local food production falls below consumption levels, placing regions in a food deficit condition and increasing their reliance on external supply sources. Conversely, in years when production exceeds consumption, regions experience food surpluses that can be utilized as strategic reserves. This analysis serves as a fundamental basis for defining annual food requirement thresholds by commodity, as actual consumption accurately represents the real food needs of local populations.

The establishment of annual food requirement thresholds for each commodity is conducted by adopting annual food consumption as the minimum benchmark that must be met by regional food availability. Each commodity is assessed separately based on consumption levels within individual regions, ensuring that the resulting thresholds reflect local characteristics such as population size, dietary patterns, and geographical conditions. In general, the annual requirement threshold for a given commodity can be set equal to its total annual consumption. However, to enhance food security—particularly in disaster-prone regions—this threshold should be augmented with a safety buffer, thereby ensuring sufficient food stocks in the event of production shortfalls or distribution disruptions.

Furthermore, the magnitude of the safety buffer incorporated into the annual food requirement threshold is adjusted according to each region’s level of vulnerability. For regions with low vulnerability, a reserve equivalent to 10-20 percent of total annual consumption is considered adequate. In contrast, disaster-prone regions require larger reserves, typically ranging from 25-30 percent of annual consumption. Accordingly, the annual food requirement threshold for each commodity is formulated as the sum of total annual consumption and an appropriate safety buffer tailored to regional vulnerability characteristics. This approach ensures that community food needs remain adequately met even under conditions of production decline or distribution disruption caused by natural disasters.

The Annual Food Requirement Threshold graph illustrates the outcomes of these calculations, showing that threshold values vary in response to fluctuations in commodity-specific consumption and regional conditions. Years characterized by higher consumption levels result in higher requirement thresholds, necessitating more intensive production and distribution planning. Conversely, in years with lower consumption, the requirement threshold decreases but must still be maintained above the minimum level required to ensure food security. By comparing actual production levels against these annual thresholds, the food availability status of each commodity in each region can be classified as secure, vulnerable, or critical.

Overall, the establishment of annual, commodity-specific food requirement thresholds at the regional level constitutes a crucial instrument for local food security planning. This approach not only facilitates the identification of priority commodities requiring increased production but also provides a robust basis for regional food reserve management and disaster mitigation strategy formulation. Utilizing production and consumption data presented in Sheet 1, this analysis offers a comprehensive, data-driven, and sustainable assessment of regional food security conditions.

5. Discussion

5.1. The superiority of Support Vector Machines for food security classification

Beyond food security, sustainable agriculture has become a fundamental component of resilient food systems. The ability to accurately predict food consumption patterns plays a critical role in supporting sustainable agricultural planning by aligning production with demand, reducing resource inefficiencies, and minimizing environmental pressures. In this context, predictive analytics contributes not only to improving food security but also to ensuring the long-term sustainability of agricultural systems. The findings of this study demonstrate that machine learning–based classification frameworks enable better synchronization between food production and consumption demand, thereby reducing inefficiencies, minimizing food waste, and promoting more sustainable resource management. In addition to strengthening food security, the predictive framework developed in this study supports sustainable agricultural development by providing an analytical basis for efficient resource allocation and environmentally responsible food production planning.

The superior performance of Support Vector Machines (SVM) in classifying regional food security status observed in this study is consistent with theoretical expectations and recent empirical evidence in related fields. The fundamental strength of SVM lies in its margin maximization principle, which identifies an optimal hyperplane that maximizes the distance between the nearest training samples from different classes (Kuri-Cervantes et al., 2020). This characteristic is particularly advantageous for food security classification, where the distinction between food-secure and food-insecure regions often depends on subtle variations in production–consumption balances and socioeconomic indicators.

The results of this study are in line with previous empirical findings demonstrating the effectiveness of SVM in food security prediction tasks. Kshirsagar et al. (2020) reported that SVM outperformed alternative classifiers in predicting household food insecurity across several countries in Sub-Saharan Africa. Similarly, Busker et al. (2024) found that SVM-based models achieved higher predictive accuracy than tree-based methods in forecasting food security crises in the Horn of Africa. The consistent superiority of SVM across different geographic regions and methodological settings highlights its robustness and suitability for modeling complex food security systems characterized by multidimensional and heterogeneous data structures.

Furthermore, the kernel trick employed by SVM enables the effective modeling of nonlinear relationships inherent in food systems. Food production and consumption dynamics are influenced by multiple interacting factors, including climate variability, economic conditions, demographic changes, infrastructure availability, and market dynamics (Nakalembe et al., 2021). These factors rarely exhibit simple linear relationships, making nonlinear modeling approaches essential for capturing their interactions accurately. By projecting data into higher-dimensional feature spaces, SVM can model these complex relationships without requiring extensive manual feature engineering, thereby improving classification accuracy and enhancing the interpretability of decision-support systems for food security management.

From a sustainability perspective, the integration of predictive analytics with robust classification models such as SVM offers strategic advantages for strengthening sustainable agricultural systems. Accurate classification of regional food security status enables policymakers and agricultural planners to anticipate consumption trends, optimize production planning, and allocate resources more efficiently. Such predictive capabilities support the reduction of production mismatches, minimize environmental stress caused by overproduction or resource misallocation, and enhance the resilience of food systems against external shocks, including climate variability and socioeconomic disruptions.

Overall, this study contributes not only to the methodological advancement of food security classification using machine learning techniques but also to the broader goal of sustainable agricultural transformation. By providing a reliable predictive framework capable of identifying food security conditions with high accuracy, the proposed approach supports adaptive production strategies, sustainable resource utilization, and environmentally responsible agricultural development. Consequently, the integration of advanced machine learning models into food system planning represents a promising pathway toward achieving resilient, efficient, and sustainable food systems in disaster-prone regions.

5.2. Performance and limitations of Random Forest

Beyond strengthening food security assessment, this study also provides meaningful contributions to the advancement of sustainable agriculture as an integral component of resilient food systems. Sustainable agriculture has become increasingly essential in maintaining food system stability, particularly in disaster-prone regions where production uncertainty and environmental pressures are more pronounced. The ability to accurately predict food consumption patterns plays a critical role in supporting sustainable agricultural planning by aligning food production with actual demand. Such alignment reduces inefficiencies in resource utilization, minimizes food losses and waste, and alleviates environmental pressures associated with excessive or misallocated production. Therefore, predictive analytics contributes not only to improving food security outcomes but also to enhancing the long-term sustainability of agricultural systems through more efficient and environmentally responsible resource management.

In terms of model performance, the findings indicate that the Random Forest (RF) model demonstrated moderate yet reliable performance, with accuracy levels ranging from 88% to 92%. This result confirms RF’s capability as a robust classification tool, consistent with its widespread application in environmental and agricultural studies (Ibrahim, 2023). The ensemble-based structure of RF provides inherent resistance to overfitting and allows relatively effective handling of noisy and heterogeneous datasets, which are commonly encountered in food security and agricultural research. These characteristics make RF particularly suitable for supporting decision-making processes in dynamic agricultural environments where variability in socio-economic and environmental indicators is unavoidable. Although RF exhibited slightly lower performance compared to Support Vector Machine (SVM) in this study, its stability and resilience against noise remain valuable attributes for building reliable predictive frameworks.

The relatively lower performance of RF compared to SVM can be explained by several methodological considerations. First, the binary classification structure employed in this study—distinguishing between food-secure and food-insecure conditions—may not fully utilize RF’s strengths, which are typically more pronounced in multi-class classification tasks involving large and diverse feature sets. Second, the relatively compact feature space used in this study, primarily composed of Food Security Index (FSI) ratios, may not provide sufficient complexity for RF’s tree-based partitioning mechanism to outperform SVM’s geometric margin-based optimization approach. These findings are consistent with Alu’datt et al. (2024), who reported that the performance advantages of RF become more evident when applied to larger datasets with higher feature dimensionality and multi-class classification structures. In the specific context of FSI-based food security classification, SVM’s margin optimization strategy appears to be more effective than RF’s ensemble voting mechanism.

Importantly, the practical implications of these findings extend beyond methodological performance and directly support the development of sustainable agricultural systems. The predictive framework established in this study enables improved alignment between food production and consumption demand, thereby reducing mismatches between supply and demand. This alignment has the potential to decrease food waste, enhance distribution efficiency, and promote sustainable utilization of natural resources. Furthermore, the model’s ability to classify regional food security status provides an evidence-based foundation for policymakers to prioritize interventions, allocate resources more effectively, and design adaptive food production strategies tailored to environmental conditions. Such targeted planning is particularly critical in disaster-prone areas, where resource limitations and environmental risks require precise and data-driven management strategies.

Overall, this study contributes not only to the development of machine learning-based food security classification methods but also to the broader agenda of sustainable agricultural development. By integrating predictive analytics into food system planning, the proposed framework supports efficient resource allocation, responsible production planning, and environmentally conscious agricultural management. Ultimately, this integrated approach strengthens the linkage between food security and agricultural sustainability, providing a strategic foundation for building resilient and sustainable food systems in the face of increasing environmental and socio-economic challenges

5.3. Limitations of Logistic Regression

The inferior performance of Logistic Regression (LR), with accuracy levels between 80% and 85%, reflects its fundamental assumption of linear relationships between predictors and outcomes. Food security dynamics in disaster-prone regions exhibit substantial non-linearity: the impact of production shortfalls on food security status is not constant but depends on baseline conditions, buffer stock levels, and adaptive capacity (Béné et al., 2024).

Furthermore, while LR’s probabilistic framework provides interpretable odds ratios, it may be less suitable for classification tasks that require hard decisions (secure versus insecure) rather than continuous probability estimates. The concentration of misclassifications near decision thresholds suggests that LR struggles to make confident classifications in ambiguous cases, where SVM demonstrates a clear advantage.

Our findings contrast with some studies reporting competitive LR performance in food security prediction (Martini et al., 2022), indicating that algorithm effectiveness is highly dependent on specific data characteristics and problem formulations. When non-linearity is prominent and classification boundaries are complex, more advanced algorithms such as SVM offer a distinct advantage.

5.4. Implications for food security monitoring systems

The demonstrated effectiveness of SVM for food security classification has significant practical implications for the development of early warning systems. First, the high accuracy achieved (92-96%) suggests that ML-based approaches can provide reliable food security assessments at substantially lower data collection costs compared to traditional survey-based methods. Automated classification using routinely collected production and consumption statistics enables more frequent and timely monitoring cycles.

Second, the incorporation of buffer reserves into the FSI calculation proved critical for accurate classification in disaster-prone contexts. Buffer levels of 10% for low-vulnerability areas and 25-30% for high-vulnerability areas provide essential protection against production shortfalls and distribution disruptions. This methodological approach should be considered for application in other disaster-prone regions.

Third, the identified performance hierarchy (SVM > RF > LR) offers clear guidance for algorithm selection in similar applications. When computational resources permit, SVM should be prioritized for binary food security classification tasks. RF may be preferable in cases where interpretability—through feature importance rankings—is emphasized, despite its slightly lower accuracy.

6. Conclusion

This study confirms that machine learning provides a robust and reliable framework for classifying regional food security status in disaster-prone areas, with the Support Vector Machine (SVM) consistently outperforming other models by achieving the highest accuracy (92–96%), precision (0.93), recall (0.94), and F1-score (0.93), followed by Random Forest (accuracy: 88–92%; F1-score: 0.87) and Logistic Regression (accuracy: 80–85%; F1-score: 0.80). Confusion matrix results confirmed the stability of SVM across validation folds, while most classification errors occurred near the Food Security Index (FSI) threshold (≈1), highlighting the importance of models capable of handling non-linear and boundary-sensitive conditions. The integration of production–consumption ratios with buffer reserves effectively captured both food availability and resilience capacity, enabling clear differentiation between surplus and structurally vulnerable regions. Beyond methodological contributions, the proposed predictive framework provides valuable scientific and practical implications for food vulnerability zoning, adaptive agricultural system development, and the efficient utilization of land, water, and agricultural inputs, thereby supporting climate-responsive and sustainable food system management in disaster-prone regions. However, the study is limited by its reliance on production–consumption indicators and region-specific data, suggesting that future research should incorporate multidimensional food security variables, broader geographic coverage, and climate-related factors to enhance model generalizability and predictive capability.

Acknowledgements

The author team would like to express its gratitude for the support provided by the Sukabumi Regency Government, Muhammadiyah University of Sukabumi.

Data Availability Statement

The datasets used and analyzed during the current study are available from the corresponding author upon reasonable request. The data contain information related to food security indicators in disaster-prone areas and are not publicly available due to privacy and institutional restrictions.

References

• ALU'DATT, M.H., RABABAH, T., AL-ALI, S., TRANCHANT, C.C., GAMMOH, S., ALROSAN, M., KUBOW, S., TAN, T.-C. and GHATASHEH, S., 2024. Current perspectives on fenugreek bioactive compounds and their potential impact on human health: a review of recent insights into functional foods and other high value applications. Journal of Food Science, vol. 89, no. 4, pp. 1835-1864. https://doi.org/10.1111/1750-3841.16970 PMid:38407443.
  » https://doi.org/10.1111/1750-3841.16970
• AMAYA-TEJERA, N., GAMARRA, M., VÉLEZ, J.I. and ZUREK, E., 2024. A distance-based kernel for classification via Support Vector Machines. Frontiers in Artificial Intelligence, vol. 7, pp. 1287875. https://doi.org/10.3389/frai.2024.1287875 PMid:38469159.
  » https://doi.org/10.3389/frai.2024.1287875
• BADAN PANGAN NASIONAL – BPN, 2023 [viewed 29 Jan 2026]. National food agency action plan [online]. Jakarta: National Food Agency. Available from: https://badanpangan.go.id/storage/app/media/2023/Dokumen%20PPID%20Berkala%202023/Rencana%20Aksi%20Badan%20Pangan%20Nasional%20Tahun%202022.pdf
  » https://badanpangan.go.id/storage/app/media/2023/Dokumen%20PPID%20Berkala%202023/Rencana%20Aksi%20Badan%20Pangan%20Nasional%20Tahun%202022.pdf
• BÉNÉ, C. and DEVEREUX, S., 2023. Resilience, food security and food systems: setting the scene. In: C. BÉNÉ and S. DEVEREUX, eds. Resilience and food security in a food systems context. Cham: Palgrave Macmillan, pp. 1-29. https://doi.org/10.1007/978-3-031-23535-1_1
  » https://doi.org/10.1007/978-3-031-23535-1_1
• BÉNÉ, C., CHEGE, C.K., EVEN, B., HERNANDEZ, R.A., LUNDY, M., PRAGER, S.D. and WIEGEL, J., 2024. Why building participatory dashboards is key for sustainable food system transformation. Frontiers in Sustainable Food Systems, vol. 8, pp. 1405670. https://doi.org/10.3389/fsufs.2024.1405670
  » https://doi.org/10.3389/fsufs.2024.1405670
• BIMANTARA, A. and DINA, T.A., 2019. Klasifikasi Web Berbahaya Menggunakan Metode Logistic Regression.Annual Research Seminar, vol. 4, no. 1, pp. 173-177.
• BUSKER, T., VAN DEN HURK, B., DE MOEL, H., VAN DEN HOMBERG, M., VAN STRAATEN, C., ODONGO, R.A. and AERTS, J.C.J.H., 2024. Predicting food-security crises in the Horn of Africa using machine learning. Earth’s Future, vol. 12, no. 8, e2023EF004211. https://doi.org/10.1029/2023EF004211
  » https://doi.org/10.1029/2023EF004211
• CORTES, C. and VAPNIK, V., 1995. Support-vector networks. Machine Learning, vol. 20, no. 3, pp. 273-297. https://doi.org/10.1023/A:1022627411411
  » https://doi.org/10.1023/A:1022627411411
• DELÉGLISE, H., INTERDONATO, R., BÉGUÉ, A., MAÎTRE D’HÔTEL, E., TEISSEIRE, M. and ROCHE, M., 2022. Food security prediction from heterogeneous data combining machine and deep learning methods. Expert Systems with Applications, vol. 190, pp. 116189. https://doi.org/10.1016/j.eswa.2021.116189
  » https://doi.org/10.1016/j.eswa.2021.116189
• DOHERTY, B., BRYANT, M., DENBY, K., FAZEY, I., BRIDLE, S., HAWKES, C., CAIN, M., BANWART, S., COLLINS, L., PICKETT, K., ALLEN, M., BALL, P., GARDNER, G., CARMEN, E., SINCLAIR, M., KLUCZKOVSKI, A., EHGARTNER, U., MORRIS, B., JAMES, A., YAP, C., SUZANNE OM, E. and CONNOLLY, A., 2022. Transformations to regenerative food systems: an outline of the FixOurFood project. Nutrition Bulletin, vol. 47, no. 1, pp. 106-114. https://doi.org/10.1111/nbu.12536
  » https://doi.org/10.1111/nbu.12536
• FATTAH, J., EZZINE, L., AMAN, Z., EL MOUSSAMI, H. and LACHHAB, A., 2018. Forecasting of demand using ARIMA model. International Journal of Engineering Business Management, vol. 10, pp. 1847979018808673. https://doi.org/10.1177/1847979018808673
  » https://doi.org/10.1177/1847979018808673
• FOOD AND AGRICULTURE ORGANIZATION OF THE UNITED NATIONS – FAO, 2024. The state of food security and nutrition in the world 2024: financing to end hunger, food insecurity and malnutrition in all its forms Rome: FAO.
• HOSMER JUNIOR, D.W., LEMESHOW, S. and STURDIVANT, R.X., 2013. Applied logistic regression. Hoboken: John Wiley & Sons. https://doi.org/10.1002/9781118548387
  » https://doi.org/10.1002/9781118548387
• IBRAHIM, S., 2023. Improving land use/cover classification accuracy from random forest feature importance selection based on synergistic use of sentinel data and digital elevation model in agriculturally dominated landscape. Agriculture, vol. 13, no. 1, pp. 98. https://doi.org/10.3390/agriculture13010098
  » https://doi.org/10.3390/agriculture13010098
• INTEGRATED FOOD SECURITY PHASE CLASSIFICATION – IPC [online], 2026 [viewed 29 Jan 2026]. Available from: https://www.ipcinfo.org/
  » https://www.ipcinfo.org/
• IOANNIDIS, J.P.A., 2018. The challenge of reforming nutritional epidemiologic research. Journal of the American Medical Association, vol. 320, no. 10, pp. 969-970. https://doi.org/10.1001/jama.2018.11025 PMid:30422271.
  » https://doi.org/10.1001/jama.2018.11025
• KSHIRSAGAR, P.R., MANOHARAN, H., KASIM, S., KHAN, A.I., ALAM, M.M., ABUSHARK, Y.B. and ABERA, W., 2020. Expedite quantification of landslides using wireless sensors and artificial intelligence for data controlling practices. Computational Intelligence and Neuroscience, vol. 2022, pp. 3211512.
• KURI-CERVANTES, L., PAMPENA, M.B., MENG, W., ROSENFELD, A.M., ITTNER, C.A., WEISMAN, A.R., AGYEKUM, R.S., MATHEW, D., BAXTER, A.E., VELLA, L.A., KUTHURU, O., APOSTOLIDIS, S.A., BERSHAW, L., DOUGHERTY, J., GREENPLATE, A.R., PATTEKAR, A., KIM, J., HAN, N., GOUMA, S., WEIRICK, M.E., AREVALO, C.P., BOLTON, M.J., GOODWIN, E.C., ANDERSON, E.M., HENSLEY, S.E., JONES, T.K., MANGALMURTI, N.S., LUNING PRAK, E.T., WHERRY, E.J., MEYER, N.J. and BETTS, M.R., 2020. Comprehensive mapping of immune perturbations associated with severe COVID-19. Science Immunology, vol. 5, no. 49, eabd7114. https://doi.org/10.1126/sciimmunol.abd7114 PMid:32669287.
  » https://doi.org/10.1126/sciimmunol.abd7114
• MARTINI, G., BRACCI, A., RICHES, L., JAISWAL, S., COREA, M., RIVERS, J., HUSAIN, A. and OMODEI, E., 2022. Machine learning can guide food security efforts when primary data are not available. Nature Food, vol. 3, no. 9, pp. 716-728. https://doi.org/10.1038/s43016-022-00587-8 PMid:37118143.
  » https://doi.org/10.1038/s43016-022-00587-8
• MULYAHATI, I.L., 2020. Implementasi machine learning prediksi harga sewa apartemen menggunakan algoritma Random Forest melalui framework website Flask Python. Yogyakarta: Universitas Islam Indonesia.
• NAKALEMBE, C., BECKER-RESHEF, I., BONIFACIO, R., HU, G., HUMBER, M.L., JUSTICE, C.J., KENISTON, J., MWANGI, K., REMBOLD, F., SHUKLA, S., URBANO, F., WHITCRAFT, A.K., LI, Y., ZAPPACOSTA, M., JARVIS, I. and SANCHEZ, A., 2021. A review of satellite-based global agricultural monitoring systems available for Africa. Global Food Security, vol. 29, pp. 100543. https://doi.org/10.1016/j.gfs.2021.100543
  » https://doi.org/10.1016/j.gfs.2021.100543
• OKTAVIANI, I.D. and PUTRADA, A.G., 2022. KNN imputation to missing values of regression-based rain duration prediction on BMKG data. Jurnal Infotel, vol. 14, no. 4, pp. 249-254. https://doi.org/10.20895/infotel.v14i4.840
  » https://doi.org/10.20895/infotel.v14i4.840
• PENDERS, B., WOLTERS, A., FESKENS, E.F., BROUNS, F., HUBER, M., MAECKELBERGHE, E.L.M., NAVIS, G., OCKHUIZEN, T., PLAT, J., SIKKEMA, J., STASSE-WOLTHUIS, M., VAN ’T VEER, P., VERWEIJ, M. and DE VRIES, J., 2017. Capable and credible? Challenging nutrition science. European Journal of Nutrition, vol. 56, no. 6, pp. 2009-2012. https://doi.org/10.1007/s00394-017-1507-y PMid:28718015.
  » https://doi.org/10.1007/s00394-017-1507-y
• REDDY, T.R.K., JANGIR, P. and TAMILARASI, G., 2025. Sustainable agriculture-based food security analysis using healthcare data modelling and deep learning techniques. Remote Sensing in Earth Systems Sciences, vol. 8, no. 1, pp. 45-55. https://doi.org/10.1007/s41976-024-00165-5
  » https://doi.org/10.1007/s41976-024-00165-5
• SAPUTRO, I.W. and SARI, B.W., 2020. Uji performa algoritma naïve bayes untuk prediksi masa studi mahasiswa. Creative Information Technology Journal, vol. 6, no. 1, pp. 1-11. https://doi.org/10.24076/citec.2019v6i1.178
  » https://doi.org/10.24076/citec.2019v6i1.178
• SARAGIH, F.M. and WIBOWO, W.C., 2025. Analysis of Food Security Index predictions in Indonesia using machine learning approach. Agro Bali: Agricultural Journal, vol. 8, no. 2, pp. 377-393. https://doi.org/10.37637/ab.v8i2.2302
  » https://doi.org/10.37637/ab.v8i2.2302
• SYUKRON, A. and SUBEKTI, A., 2018. Penerapan Metode Random Over-Under Sampling dan Random Forest Untuk Klasifikasi Penilaian Kredit. Jurnal Informatika, vol. 5, no. 2, pp. 175-185. https://doi.org/10.31294/ji.v5i2.4158
  » https://doi.org/10.31294/ji.v5i2.4158
• XIAO, G., CHENG, Q. and ZHANG, C., 2019. Detecting travel modes using rule-based classification system and Gaussian process classifier. IEEE Access: Practical Innovations, Open Solutions, vol. 7, pp. 116741-116752. https://doi.org/10.1109/ACCESS.2019.2936443
  » https://doi.org/10.1109/ACCESS.2019.2936443
• WESTERVELD, J.J.L., VAN DEN HOMBERG, M.J.C., NOBRE, G.G., VAN DEN BERG, D.L.J., TEKLESADIK, A.D. and STUIT, S.M., 2021. Forecasting transitions in the state of food security with machine learning using transferable features. The Science of the Total Environment, vol. 786, pp. 147366. https://doi.org/10.1016/j.scitotenv.2021.147366 PMid:33971600.
  » https://doi.org/10.1016/j.scitotenv.2021.147366

Edited by

• Editor:
  Takako Matsumura Tundisi

Publication Dates

• Publication in this collection
  29 May 2026
• Date of issue
  2026

History

• Received
  02 Feb 2026
• Accepted
  31 Mar 2026