Supervised learning refers to a paradigm where an algorithm is provided with a labeled dataset containing input data and corresponding target values (correct answers). Through iterative training on these samples and their associated labels, the algorithm establishes relationships between input features and output mappings [3] - [5] . By learning from labeled data, the machine aims to predict outcomes for new, unseen data [6] . A simplified workflow of supervised learning is illustrated in Figure 1 .

Supervised learning primarily encompasses regression and classification. Regression predicts continuous numerical outcomes, while classification predicts discrete categorical labels. Regression algorithms are mainly divided into logistic regression and linear regression, whereas classification tasks are categorized into binary classification (two classes) or multiclass classification (multiple classes) based on the number of target categories. Regression models are typically applied to problems where the output variable is a real-valued quantity, whereas classification models categorize the output variable. The three core components of supervised learning are models, strategies, and algorithms.

In real-world applications, supervised learning is widely used in scenarios such as spam SMS filtering on mobile devices and spam email detection in email systems. These tasks rely on training models using historical human-labeled data (e.g., user-marked spam messages) to predict whether newly received messages or emails are spam

Unsupervised learning utilizes unlabeled datasets for training, where sample data lacks predefined “correct answers.” Unlike supervised learning, the machine does not analyze input-output mappings but instead identifies inherent patterns within the data to classify or cluster samples. The simplified workflow of unsupervised learning is illustrated in Figure 2 .

In contrast to supervised learning, which relies on labeled data with clear target outcomes, unsupervised learning operates without labels. Supervised learning has well-defined objectives and measurable performance criteria (e.g., accuracy), whereas unsupervised learning lacks standardized evaluation metrics due to its exploratory nature.

Unsupervised learning primarily addresses dimensionality reduction and clustering tasks. Dimensionality reduction aims to simplify data complexity while preserving critical structural relationships. Classic dimensionality reduction algorithms include:

Semi-supervised learning is a hybrid approach that bridges the gap between supervised and unsupervised learning. One of its primary objectives is to optimize the strengths of both paradigms by enabling machines to leverage unlabeled samples for enhanced learning performance. The simplified workflow of Semi-supervised learning is illustrated in Figure 3 . Semi-supervised learning can be classified into two categories:

Between 2003 and 2024, a total of 1,472 articles were retrieved for analysis. The annual publication trend is illustrated in Figure 4 , which reveals an overall upward trajectory in research output. Specifically:

This growth aligns with advancements in machine learning technologies and the rising global prioritization of emergency management, driving heightened scholarly engagement in this interdisciplinary field.

An analysis of contributing authors among the retrieved literature reveals the top five authors by publication volume, as shown in Table 2 . The most prolific author is Ong Marcus Eng Hock, whose research team [13] - [15] predominantly applies machine learning to emergency medicine, focusing on integrating machine learning with medical emergency response systems to improve treatment success rates. The team led by Li Nan [16] developed a novel random forest-based method to screen the most relevant variables contributing to chest pain and constructed a geometric distance-based machine learning scoring system. The results demonstrated that this machine learning-driven variable selection approach exhibits significant potential in identifying key predictors. Their research has laid a foundation for applying machine learning in clinical performance audits and screening critical risk variables, thereby providing strategic direction for advancements in related fields.

Keyword analysis of academic literature identifies research hotspots within a field, enabling the assessment of its developmental trends. By consolidating synonymous terms and analyzing keywords from the retrieved articles, Table 3 lists the top 10 high-frequency keywords, while Figure 5 visualizes their co-occurrence. Key findings include: Machine learning, artificial intelligence (AI), and deep learning dominate as the most frequent keywords in studies on machine learning applications for emergency management.

Advances in technology have facilitated interdisciplinary integration of machine learning, addressing critical challenges such as: forest fire early warning systems [17] , liquid nitrogen leakage detection [18] , post-disaster resource allocation [19] . Similarly, the application of deep learning in emergency contexts has enabled advancements like: emergency vehicle signal prioritization [20] , optimized emergency communication network planning [21] , indoor/outdoor evacuation route planning for fire emergencies [22] .

Data Analysis involves selecting appropriate statistical methods to analyze collected massive datasets, uncovering inherent relationships within the data to enhance interpretability and maximize its utility. In the era of big data, it is a purposeful and organized process of collecting, analyzing, and transforming raw data into actionable insights. Data analysis records objective events as identifiable symbols and finds applications across diverse fields. At its core lies the data itself, and modern analysis requires clarifying data variables and understanding their dynamics [23] .

A critical application of machine learning in emergency management is data analysis. Disasters and accidents often generate vast amounts of data, including meteorological, geographical, and demographic information. Machine learning processes and analyzes these datasets to identify patterns and trends, providing actionable references for emergency management agencies. For instance, prior to natural disasters, machine learning can analyze historical weather data to forecast future conditions, enabling proactive preparedness.

To mitigate leakage incidents, Zhou Dehong et al. [24] conducted a study using multiple LNG (liquefied natural gas) leakage accident reports. They applied Principal Component Analysis (PCA) to the dataset, first extracting the first two principal components to evaluate their classification impact, followed by analyzing the top five components to rank feature weights under each principal component. Subsequently, a Random Forest algorithm was employed to classify the data and assess the contribution of each indicator to classification accuracy. By comparing results from both methods, they systematically identified root causes of leakage.

Addressing challenges in ensuring communication and data security for smart grid control systems, Wang Ning et al. [25] proposed a Support Vector Machine (SVM)-based framework integrated with real-time risk assessment and early warning capabilities. By enhancing SVM with Convolutional Neural Networks (CNNs), their system performs risk identification, severity classification, and dynamic alerts while accounting for multifactorial influences. This approach significantly improves the stability and reliability of grid control systems.

Another critical application of machine learning in emergency management is predictive analytics. By learning from and analyzing historical data, machine learning models can forecast potential future disasters or accidents. For example, analyzing historical traffic accident data enables the prediction of high-risk road segments and time periods, empowering transportation authorities to implement targeted interventions and reduce accident probabilities.

Yu Hong et al. [26] developed an emergency supplies demand prediction model using machine learning. They employed internal analytical methods to dissect demand management patterns and integrated visual analytics to geotag incident-prone areas, thereby enhancing the model’s architecture and improving prediction accuracy for emergency resource allocation.

Liu Fang et al. [27] proposed an improved ant colony optimization BP neural network algorithm (IACO-BP) to construct a prediction model for estimating the number of people requiring evacuation during flood disasters.

Liu Jianhua et al. [28] applied a BP neural network for urban flood disaster prediction. To enhance accuracy, they combined the BP network with clustering algorithms and introduced “inertia correction” and “S-function output clamping” techniques, achieving effective flood impact forecasting.

Machine learning can also assist emergency management departments in decision-making. In the aftermath of disasters and accidents, emergency management authorities need to make rapid decisions to allocate resources and conduct rescue and relief operations. Machine learning can provide decision support through real-time monitoring and analysis of on-site data. For instance, after an earthquake, machine learning can process image data from affected areas to automatically identify building damage levels, helping emergency management departments prioritize dispatching rescue forces to the most severely impacted regions.

He Peiyu et al. [29] proposed a genetic algorithm-based optimization model for emergency material reserve decision-making in transmission line ice-coating disasters. This model uses genetic algorithms to determine effective and rational strategies for emergency material storage during ice-coating events on power lines.

Wei Dong et al. [30] developed a novel method based on convolutional neural networks (CNNs) for fault identification and phase selection in transmission line internal/external fault discrimination. Their CNN-based approach simultaneously addresses both issues while resolving the shared weight problem in non-independent classification tasks. This method requires lower sampling rates, eliminates parameter dependencies, and delivers more accurate and reliable results.

Liu Wei et al. [31] introduced a deep reinforcement learning-based generator tripping control strategy for power grids. This strategy integrates deep convolutional networks with reinforcement learning, utilizes Q-Learning to compute Q-values, and validates its effectiveness through IEEE-standard verification methods. The control strategy enables autonomous environmental learning and continuous self-adjustment, reducing human error-related accidents while enhancing decision-making accuracy.

The primary task of machine learning involves training datasets with suitable algorithms, which imposes requirements on data quality. Machine learning applications rely heavily on large volumes of data, yet in emergency management, data quality and completeness are often constrained. After incidents occur, data uploads are frequently delayed, and data quality is difficult to guarantee. To enable machine learning to address emergencies promptly, coordination across departments and advanced equipment is needed to improve data quality and integrity. Additionally, machine learning requires representative datasets to ensure strong model generalization capabilities. However, post-incident data collection often yields chaotic data, necessitating additional steps to filter and organize it. For example, in earthquake disasters, accurately identifying and classifying building damage is critical, but uncertainties in data sources and quality limitations can compromise recognition accuracy [32] .

During machine learning implementation, multiple solutions may exist for the same problem. As model complexity increases, algorithmic interpretability often deteriorates [33] . Machine learning applications typically generate vast amounts of data and predictions, but explaining these results in an understandable and acceptable manner remains challenging. When inputting data, a machine learning model’s explanation mechanism provides an interpretation. It is crucial that similar inputs yield consistent or analogous explanations. For research purposes, models must produce interpretable results rather than functioning as “black boxes.” Emergency management departments need to understand the internal mechanisms and operational logic of machine learning algorithms and supply sufficient high-quality data to ensure accurate decision-making and resource allocation.

Machine learning involves two key phases—training and prediction—each posing security and privacy risks. In the training phase, privacy threats primarily stem from training data leakage. During prediction, risks include adversarial attacks, privacy breaches, and prediction data leaks [34] . Privacy protection must be prioritized in machine learning applications, focusing on safeguarding sensitive information through processes and rules. When applied to private data, machine learning risks exposing sensitive information, potentially harming affected parties. Privacy protection research in machine learning generally falls into three categories: equivalence class-based methods, data distortion-based methods, and cryptographic methods. In emergency management, large volumes of personal and sensitive data—such as medical records and census data—are involved. Applications must comply with privacy laws, ethical standards, and ensure data security. In the era of big data, where information is critical to productivity and daily life, securing data is paramount. Post-incident data collection must be strictly regulated to prevent leaks and protect personal information.