AI–Social Media Integration for Crisis Management: A Systematic Review of Data and Learning Aspects

Abstract

As natural disasters and crises increase globally in both frequency and severity, researchers have been exploring innovative technological solutions to manage them effectively. This systematic review examines the integration of artificial intelligence (AI) with social media platforms for crisis management, identifying and categorizing key components of AI-driven systems into data and learning aspects. It introduces a dual-aspect analytical taxonomy that provides a structured framework for analyzing how data and learning dimensions interact in AI-driven crisis management solutions. Following the PRISMA methodology, the review analyzed 30 high-impact, peer-reviewed journal articles published in English between 2020 and 2024 across major academic databases. The quality of the studies was assessed based on journal ranking and methodological rigor to ensure reliability and minimize bias. The analysis revealed several interconnected trends: text remains the dominant data modality (60%), while multimodal analysis (33%) and image-based analysis (7%) are gaining traction. Throughout these studies, deep learning models consistently demonstrated superior performance compared to traditional machine learning approaches, with hybrid methodologies significantly enhancing overall model efficiency. Notably, the majority of research (73%) concentrated on during-disaster phases, highlighting the critical need for real-time intervention solutions. Twitter/X emerged as the overwhelming primary data source (73%), creating potential platform dependency issues. Despite considerable advancements, the field continues to face persistent challenges, including an over-reliance on single platforms, insufficient real-time AI models, and complexities in multimodal data fusion. To advance crisis management capabilities, future research directions should address cross-domain generalizability, enhance real-time processing capabilities, and develop improved fusion techniques that can ultimately lead to more effective and timely disaster response systems.

1. Introduction

The severity and frequency of human-made and natural disasters have been increasing globally, resulting in devastating impacts that include harmful physical, social, and economic consequences for individuals and communities [1,2]. For the past decade (2011–2020), about 1.6 billion people worldwide were affected by natural disasters, which caused 188,583 fatalities and over USD 1.7 trillion in financial losses [1]. These staggering figures underscore the urgent need for effective crisis management strategies.

The impact of disasters extends far beyond immediate physical damage, affecting various aspects of human existence and community infrastructure. Vital systems such as transportation networks, power grids, communication systems, and water supply are often compromised, disrupting the operational integrity of entire communities. Moreover, the after-effects of disasters can lead to a cascade of social and economic issues, including bodily harm, unemployment, homelessness, psychological distress, and financial adversity. To mitigate these risks and reduce the overall impact of disasters, communities must continuously explore and implement innovative techniques for disaster preparation, response, and recovery [1].

In recent years, the widespread adoption of social media platforms has emerged as a game-changer in crisis communication. These nontraditional information sources have become practical tools for quickly acquiring relevant information and supporting various disaster risk reduction activities [1]. During crises, a substantial amount of user-generated content is shared across various platforms, including Twitter (now X), Facebook, YouTube, WeChat, and Weibo. This digital ecosystem facilitates multi-level communication, including person-to-person, person-to-government agencies, and government-to-people interactions [3].

The general public actively uses social media during disasters to post real-time situational updates, including reports of casualties, infrastructure damage, and urgent needs [4]. These social media messages, enriched with various data types, have proven invaluable in coordinating aid efforts and enhancing situational awareness. Numerous examples highlight the critical role of social media in disaster response. For instance, during Hurricane Harvey in 2017 (https://hdl.handle.net/1911/105229, accessed on 15 November 2024), a woman was rescued after tweeting for help when traditional emergency contact methods failed [5]. Similarly, research has shown the positive impact of social media usage during wildfire disasters in mobilizing first responders, demonstrating how rapid and immediate responses facilitated by these platforms can save lives and prevent critical injuries [6].

Recognizing the potential of social media in crisis management, governments and humanitarian organizations now actively seek relevant information on these platforms to prevent crises or assist victims promptly [4]. Many crisis management agencies have begun incorporating social media data into their workflows [2,4]. However, the sheer volume of data generated during disasters presents a significant challenge in rapidly understanding and responding to evolving situations.

To address this challenge, the integration of artificial intelligence (AI) techniques with social media data analysis has emerged as a crucial approach. In this paper, the term “AI techniques” encompasses computational approaches such as machine learning (ML), deep learning (DL), natural language processing (NLP), and computer vision (CV), which are used to process and interpret social media data during crises. AI’s capability to process and analyze massive amounts of social media data efficiently is essential for effective crisis management [6]. This integration allows responsible organizations to quickly comprehend crisis situations, coordinate rescue and aid efforts, and make informed decisions for rapid problem management and response.

While the timely, accurate, and effective use of social media information is vital for managing emergencies [7], extracting valuable insights using AI techniques presents several challenges [8]. These include near-real-time information processing, managing information overload, efficient information extraction, summarization, and verifying both textual and visual content [4]. In this review, “real time” refers to systems capable of analyzing and responding to social media data streams with minimal latency, encompassing both streaming and near-real-time processing approaches that support timely situational awareness and decision making during crises. To address these challenges, researchers have proposed various solutions combining social media and AI techniques. Most existing crisis management solutions share common essential factors that can be broadly categorized into two dimensions: the data aspect and the learning aspect.

1.1. Related Work and Motivation

Several review studies have been conducted on crisis management, each focusing on different aspects of the field. Martínez-Rojas et al. [7] and Dwarakanath et al. [9] concentrated on Twitter as a data source for emergency management, discussing methodologies, events, and disaster phases. However, their scope was limited to a single platform.

Ogie et al. [1] examined social media usage specifically in the disaster recovery phase, analyzing platforms, disaster types, and geographical focus. Saroj et al. [2] investigated the relationship between emergencies and social media data, covering impact, data processing, and organizational effects. Bukar et al. [10] discussed crisis management technologies, including social media, machine learning, big data, and social network analysis. However, the authors in this review provided limited case studies and practical applications demonstrating these tools in real-world scenarios.

Bukar et al. [11] focused on investigating existing theoretical models used in crisis communication in the context of social media and crisis informatics, but they lacked practical application analysis. Abboodi et al. [12] provided a multi-perspective view to identify the current body of knowledge on the adoption of social networks in crisis management, but they placed less emphasis on a systematic exploration of the topic.

Although some existing reviews have covered aspects of crisis management solutions as shown in

Table 1. Comparison of related reviews on crisis management using social media data and AI techniques. The table encompasses key aspects, including data types (unimodal, multimodal), data sources (single, multiple), labeling methods, learning models, learning domains, learning phases, learning types, and learning objectives.

Ref.	Year	Coverage	Unimodal Data	Multimodal Data	Single Data Source	Multiple Data Sources	Labeling Methods	Learning Models	Learning Domains	Learning Phases	Learning Types	Learning Objectives
[7]	2018	2009–2018	√	×	√	×	×	√	×	√	√	√
[2]	2020	2007–2019	√	√	√	√	×	√	×	√	×	√
[11]	2020	2014–2020	√	×	√	×	×	√	×	√	×	×
[9]	2021	2007–2019	√	×	√	×	×	√	√	√	×	√
[1]	2022	2008–2021	√	×	√	√	×	√	×	√	×	√
[10]	2022	2011–2021	√	×	√	√	×	√	×	×	×	√
[12]	2023	2018–2022	√	×	√	×	×	×	×	×	×	√
Ours	2025	2020–2024	√	√	√	√	√	√	√	√	√	√

√ indicates that the study meets the criterion, whereas × indicates that it does not.

, they often focus on isolated perspectives, such as particular phases of a disaster [1], limited data types [1,7,9,10,11,12], or single data sources [7,9,11,12]. Another noticeable gap is the limited consideration of labeling methods, learning types, and learning domains across various reviews. Indeed, only a few studies have explored these essential aspects, creating opportunities for a more in-depth analysis of how they impact crisis management solutions.

Based on our analysis of these reviews, it becomes evident that none of them discussed the complete spectrum of data types, learning models, labeling methods, learning types, learning domains, learning phases, and data sources for various learning objectives. These observed limitations underscore the importance of a comprehensive review that systematically identifies, compares, and synthesizes the key aspects of AI-driven crisis management solutions that integrate social media data.

1.2. Research Gap and Contributions

Understanding the essential aspects of AI-driven crisis management solutions is critical in today’s dynamic landscape. While numerous review papers have explored crisis management from various angles, to the best of our knowledge, no existing review has comprehensively discussed and identified the essential aspects of AI-driven, social-media-based crisis management solutions. This gap in the literature motivates our current review, which aims to identify the common essential aspects of state-of-the-art AI and social-media-based crisis management solutions.

This review examines proposed crisis management solutions that utilize social media and AI techniques, identifying key aspects frequently explored in these approaches. Our research endeavors to establish a foundation for further studies in this critical domain by synthesizing current research on crisis management solutions and highlighting existing gaps and challenges in the literature.

We focus primarily on identifying and discussing the common factors and essential aspects of the proposed crisis management solutions that integrate AI techniques with social media platforms. The review also addresses the following research questions:

• RQ1: What are the recently proposed solutions for crisis management using AI and social media platforms?
• RQ2: What are the common essential aspects of the crisis management solutions proposed through the integration of social media and AI techniques?
• RQ3: What taxonomy can effectively summarize the essential aspects of the proposed solutions in the field of crisis management?
• RQ4: What are the key findings derived from the reviewed studies?
• RQ5: What are the main challenges in current solutions related to the integration of social media data and AI techniques for crisis management?

To address our research questions and provide a comprehensive analysis of crisis management solutions, this review offers the following key contributions:

• A comprehensive analysis of recent crisis management solutions that integrate AI techniques with social media platforms.
• A comparative synthesis of existing approaches, highlighting the common essential aspects across studies and providing a unified understanding of the field.
• The development of a structured taxonomy of these aspects, forming the basis for a novel analytical framework for AI-driven crisis management research.
• The synthesis of key findings from reviewed studies, identifying emergent trends and evidence-based practices in the field.
• The identification of critical challenges in current solutions, establishing a foundation for future research directions in crisis management using AI and social media.
• Actionable insights and evidence-based recommendations to support researchers and practitioners in developing more effective and innovative crisis management solutions.

In addition to the aforementioned contributions, this review provides a novel conceptual advancement by introducing a dual-aspect analytical taxonomy that structures the analysis of AI-driven crisis management solutions. This taxonomy integrates two complementary dimensions:

• Data Aspect: encompasses key factors such as data types, data labeling methods, and data sources, which together define how information is collected, represented, and annotated across diverse crisis scenarios.
• Learning Aspect: includes learning models, learning domains, learning phases, learning types, and learning objectives, which determine how AI methods interpret, generalize, and adapt to diverse crisis contexts.

This unified framework provides a structured foundation for analyzing the interaction and alignment between data and learning dimensions. It elucidates how data characteristics influence the design of learning models as well as how learning strategies adapt to address crisis-specific challenges such as data heterogeneity, labeling inconsistency, and multimodal information fusion. This analytical perspective constitutes the principal theoretical contribution of this review. It offers a systematic framework for identifying research gaps in current studies and guiding future investigations in AI-based crisis management.

1.3. Paper Organization

The organizational structure of this systematic review is illustrated in Figure 1. Section 2 presents the research methodology employed for systematic article selection. Section 3 explores the common aspects of crisis management solutions from a data perspective, summarizing and critically analyzing the solutions proposed in the reviewed studies based on these data-related aspects. Section 4 discusses the relevant factors associated with the learning aspect and provides a comprehensive overview of recently proposed solutions in this context. Furthermore, Section 5 integrates the findings of this review by analyzing the interaction between data and learning aspects in recent AI-driven crisis management solutions. This section highlights emerging trends, identifies challenges, and proposes potential research directions. Finally, Section 6 summarizes the contributions and implications of this work.

2. Research Methodology

This section outlines the methodology employed to identify articles relevant to the research topic. Our methodological approach adheres to the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines [13], ensuring a rigorous and transparent process for article selection and analysis. Figure 2 presents the PRISMA flow diagram, illustrating the systematic process of article identification, eligibility assessment, screening, and final inclusion in our review.

As the review focuses on rapidly evolving AI and social media research domains, it was not registered in a formal registry such as PROSPERO, which primarily covers health-related studies. The review team conducted all stages of screening and data extraction independently to ensure consistency and minimize bias. Disagreements were resolved through open discussion and a consensus-based approach.

2.1. Source Material and Search Strategy

To identify candidate papers, we conducted a comprehensive search across four leading databases in computer science and artificial intelligence: Springer (https://link.springer.com/, accessed on 23 November 2024), ScienceDirect (https://www.sciencedirect.com/, accessed on 23 November 2024), Taylor & Francis (https://www.tandfonline.com/, accessed on 23 November 2024), and IEEE Xplore (https://ieeexplore.ieee.org/, accessed on 23 November 2024). All databases were systematically searched for studies published between 2020 and 2024.

To initiate our search for articles relevant to the research topic, we developed a comprehensive search query incorporating key terms from our areas of interest. The search query used in this research is as follows: (crisis OR disaster OR emergency OR “disaster management” OR “crisis management” OR “emergency management” OR “natural disaster” OR hazard) AND (“social media” OR “Social network”) AND (“artificial intelligence” OR “machine learning” OR “deep learning” OR “natural language processing” OR “computer vision” OR “data mining” OR “data analysis”).

Our search across the four selected databases yielded a substantial initial corpus of 64,375 records, which were distributed as follows:

• Springer: 15,784 records;
• ScienceDirect: 38,581 records;
• Taylor and Francis: 9073 records;
• IEEE Xplore: 937 records.

2.2. Study Selection Based on Eligibility Criteria

This phase rigorously evaluates the selected articles to ensure their relevance and significance to the review’s objectives. By applying carefully defined inclusion and exclusion criteria, we refine our corpus to include only the most pertinent and high-quality research.

• Inclusion Criteria:

  -

  Journal articles in English;

  -

  Published between 2020 and 2024;

  -

  Focused on computer science and technology;

  -

  Utilized AI techniques in crisis management contexts.

• Exclusion Criteria:

  -

  Non-article document types (e.g., books, book chapters, conference proceedings, reference works);

  -

  Publications before 2020;

  -

  Studies from fields outside computer science (e.g., engineering, business, medicine, public health) not employing AI techniques.

This systematic filtering process, implemented through an automated screening tool, resulted in the exclusion of 62,266 records. The remaining articles form a highly relevant and focused corpus for our review, ensuring that our analysis concentrates on contemporary AI applications in crisis management within the field of computer science.

After applying the inclusion and exclusion criteria, we further refined our selection process:

• Duplicate Removal: 217 duplicated records were identified and removed, ensuring each study was represented only once in our dataset.
• Quality Assessment: We evaluated the credibility and reliability of the remaining studies based on their journal’s quartile ranking in the Web of Science.
  In addition, the methodological quality and potential risk of bias of each included study were independently assessed by the review team to ensure transparency and rigor. The quality assessment followed two levels: (1) evaluating the credibility of journals based on their Web of Science quartile ranking and (2) assessing the methodological rigor and potential risk of bias of each study based on the consistency, completeness, and clarity of methodological reporting.
  • Priority was given to articles published in journals with an ISI Impact Factor in the Q1–Q2 range, recognizing their high standards and rigorous peer review processes.
  • Articles not indexed in the Web of Science or positioned in the third or fourth quartile rankings were excluded.

This stringent selection process aimed to enhance the integrity and credibility of our review findings while minimizing potential biases, resulting in the following decisions:

• 1700 records were excluded due to not meeting the quality criteria.
• 192 articles remained for further examination.

2.3. Titles and Abstracts Screening

In this phase, we performed a rigorous screening of the 192 remaining articles by examining their titles and abstracts. This crucial step was designed to identify and exclude studies that did not closely align with the specific focus of our research. The following categories of studies were filtered out:

• Articles exploring the use of social media in crisis management that do not incorporate AI techniques.
• Studies employing AI in crisis management that operate independently of social media data streams.

This methodologically rigorous screening process facilitated the systematic refinement of our corpus, thereby ensuring that only the most relevant articles advanced to full-text analysis and leading to the following results:

• 52 articles were excluded due to their failure to satisfy the eligibility criteria.
• 140 studies were retained for further analysis.

2.4. Data Extraction

During the data extraction phase, we categorized the 140 articles based on specific classifications derived from previous studies [14,15]. As illustrated in Figure 3, these categories included the following:

• Empirical studies;
• Comparative studies;
• Behavioral models;
• Theoretical models;
• Improvements on theoretical models;
• Frameworks, tools, models, and systems;
• General application and usage;
• Review papers;
• Other.

Given our review’s primary objective of identifying common characteristics across crisis management solutions that integrate AI and social media platforms, we strategically focused our analysis on the “frameworks, tools, models, and systems” category—a substantial subset comprising 66 studies from our initial corpus. To ensure a comprehensive yet focused examination, we conducted a meticulous evaluation process involving thorough content analysis and a critical assessment of relevance criteria. This systematic approach yielded the 30 most relevant articles (visually distinguished in green in Figure 3) that formed the foundation for our in-depth synthesis and critical examination of emerging patterns and technological approaches. The review team independently extracted data from the selected studies using a predefined template to ensure consistency and minimize bias.

2.5. Synthesis of Results

After selecting the most relevant 30 articles, we conducted a comprehensive analytical examination to identify common aspects and construct a taxonomy that summarizes their fundamental components. To ensure the reliability and objectivity of the synthesis, we prioritized the most studies demonstrating exceptional methodological quality and substantial scholarly impact in the field, thereby minimizing potential biases in our results. This synthesis directly addresses Research Question 1 (RQ1), which undergoes further detailed examination in Section 3 and Section 4 through a detailed analysis of recent AI and social-media-based crisis management solutions.

Our analysis revealed that most crisis management solutions leveraging social media and AI techniques share common aspects frequently examined in proposed solutions within this domain. We categorized these aspects into two primary categories: the data aspect and the learning aspect, as illustrated in Figure 4. The data aspect captures how social media information is collected, labeled, and sourced for crisis management tasks, while the learning aspect encompasses the models, phases, learning domains, learning types, and objectives. These aspects collectively guide proposed solutions in transforming raw information into actionable insights for crisis response. The identification of these aspects addresses Research Question 2 (RQ2), highlighting the key factors that are commonly examined in AI-driven crisis management solutions.

Additionally, the data and learning aspects highlight the foundational elements necessary to construct robust, adaptable, and effective crisis management frameworks. These aspects are detailed in Section 3 and Section 4, respectively, which form the basis for the dual-aspect analytical taxonomy proposed in this review. This categorization forms the foundation for addressing Research Question 3 (RQ3), which focuses on developing a taxonomy that summarizes the essential aspects of AI-driven crisis management solutions.

3. Data Aspect

To address Research Question 3 (RQ3), this section examines the first dimension of the proposed taxonomy, the data aspect, which forms the foundation for analyzing AI-driven crisis management solutions. The data aspect encompasses several key elements necessary for performing efficient analysis, including data types, data labeling methods, and data sources, as illustrated in Figure 5. Each of these elements plays a critical role in shaping the effectiveness of crisis management solutions.

In the following subsections, we thoroughly examine these essential aspects, and the solutions proposed in the reviewed studies are discussed and summarized in relation to these aspects.

3.1. Data Type

Different data types are usually present in social media, including text, images, videos, and audio. Each type has its own characteristics and requires a different method of analysis. The analysis of messages in social media can be categorized as unimodal, which deals with one type of available information, or multimodal, which deals with messages that contain more than one type of data.

3.1.1. Unimodal Data

Unimodal data analysis primarily focuses on examining a single data type, such as textual or visual data.

For textual data, several studies [6,16,17,18,19,20,21,22,23,24,25,26,27,28,29] have been conducted that focus on analyzing social media textual data for various objectives in crisis management. For instance, studies [6,16,17,18] have developed multi-label and binary classification schemes to analyze disaster relevance, damage, and crisis-specific tweet classification using textual tweets, which play a crucial role in classifying and identifying critical pieces of information during emergencies. Other studies on the topic [30,31,32] have addressed detecting sub-events and enhancing the semantic representation of disaster-related tweets by utilizing hashtag semantic refinement or term frequency-inverse document frequency (TF-IDF)-based feature vectors. These approaches have indeed shown improvement in the classification of disaster-specific phenomena.

As for visual data, social media image data are analyzed using image-based models. This analysis can provide significant information for disaster response. Since images can contain useful information, various studies have examined the value of analyzing social media images during different types of disasters [33].

For instance, Chaudhuri et al. [33] addressed the problem of a lack of insightful information on disaster sites by utilizing image data obtained from smart infrastructure and social media platforms. They demonstrated the capability of image analytics in aiding resource prioritization decisions for disaster response. Another study by Zhang et al. [34] utilized streaming visual data posted on social media to automatically assess the severity of damage in affected areas during a disaster event.

3.1.2. Multimodal Data

Multimodal content encompasses a range of elements, including text, images, video, geolocation data, and metadata. These different data types are known as modalities, and when analyzed collectively, they may provide a comprehensive understanding of disaster situations [4].

Several studies have been conducted using multimodal data, demonstrating that multimodality effectively provides a comprehensive understanding of disaster events. For instance, Wang et al. [35] developed a framework that analyzes both text and images embedded in social media data to capture disaster-related information, thereby improving situational awareness. Wu et al. [36] exploited the correlation and independence relationships between text and image data to identify humanitarian information for emergency response. They found that social media posts with multimodal data can provide better and more apparent insight into ongoing emergencies. Similarly, Refs. [37,38,39,40,41,42] have empirically investigated various multimodal techniques. According to these studies, the performance of multimodal models that combined text with images outperformed unimodal models.

The current technology for generating crisis event summaries is primarily influenced by unimodal bias, which overlooks the diversity of information in both text and images. To address these problems, Wang et al. [43] proposed a hierarchical multimodal crisis event summary generation model based on modal alignment and hierarchical thinking. Similarly, Saini et al. [44] generated summaries using image and tweet data from disaster events.

3.2. Data Labeling Methods

While machine learning systems have improved tremendously in recent years, the most crucial factor behind their effectiveness is the quality of the labeled training data [45]. The process of assigning appropriate labels to data is known as data annotation. To train the models, we need labeled datasets. The labeled data may take the form of images, text, or video, and data annotators must label them as accurately as possible, as this step is critical for training the model and evaluating its performance. As explained below, the labeling methods could be manual, automatic, or hybrid.

3.2.1. Manual Labeling

One approach used for manual labeling is crowdsourcing. It has been particularly beneficial in times of crisis, as demonstrated by various studies [6,21,22,27,30,33,34,35,36] that have employed crowdsourcing methods for the manual labeling of data. For example, Fan et al. [22] used tweets collected during the 2017 Hurricane Harvey in Houston, labeling them with annotations for humanitarian categories. Zhang et al. [34] manually classified social media images from Typhoon Hagupit and Hurricane Harvey into damage severity levels using domain experts. Wu et al. [36] constructed an annotated dataset designed explicitly for the multimodal humanitarian information identification task to evaluate the effectiveness of their proposed model.

Other studies [16,17,19,25,28,38,39] have utilized existing manually labeled datasets such as CrisisLex (https://crisislex.org/, accessed on 18 December 2024), CrisisNLP (https://crisisnlp.qcri.org/, accessed on 20 December 2024), and CrisisMMD (https://crisisnlp.qcri.org/crisismmd, accessed on 26 December 2024), which include tweets related to various types of crises, providing a rich source of annotated data for disaster-related research. Studies like those by Yao and Wang [28] and Yang et al. [38] leveraged existing annotated datasets and combined them with manual classification to ensure consistency.

3.2.2. Automatic Labeling

Automatic labeling methods are crucial for managing the vast amount of data generated on social media, particularly during emergencies. These methods involve both supervised and unsupervised machine learning techniques. Supervised machine learning techniques classify new instances based on previously labeled datasets, while unsupervised machine learning techniques, such as clustering, find hidden patterns in unlabeled data to group similar items together, minimizing manual labeling effort [46].

Automatic labeling methods are employed when labeled data are scarce or manual labeling is inefficient. In the literature, most studies combine automatic labeling methods with manually labeled datasets, as presented in the following approach.

3.2.3. Hybrid Approach

Several studies have applied a hybrid approach that combines automatic labeling methods with manually labeled datasets [18,20,23,24,26,29,31,32,37,40,41,42,43,44]. These studies incorporate manual labeling with automatic methods to enhance the efficiency and accuracy of the labeling process.

For example, Bryan et al. [37] initially manually labeled their dataset before using automatic labeling through predefined categories. Hossain et al. [40] utilized a benchmark multimodal damage dataset that was initially filtered through relevancy classifiers before undergoing manual annotation. Humaira et al. [42] combined manual labeling with pretrained object detection models to annotate flooding images, aiming to reduce manual workload. Wang et al. [43] utilized news content from the Daily Mail’s mobile phone service to build a large-scale multimodal dataset with annotated summaries, using a pretrained Transformer model to classify crisis event news. Dhiman et al. [20] utilized existing annotated datasets and automatic labeling techniques for various Twitter datasets. In contrast, Kamoji and Kalla [41] employed both manual and automatic approaches, incorporating Bidirectional Encoder Representations (BERTs) for automatic labeling.

To address the time-consuming nature of manual labeling, Saini et al. [44] created a semi-supervised data annotation method. They used initial data from the CrisisMMD resource, which includes pre-annotated tweets, and employed human annotators to generate gold summaries from these data manually.

Active learning, a type of automatic labeling, enhances classification accuracy by selecting items for human labeling, thereby improving the overall model performance [46]. Kaufhold et al. [18] combined expert labeling with active learning techniques to improve labeling efficiency. Snyder et al. [23] developed an interactive learning framework that utilizes users to label tweets, thereby iteratively improving real-time relevance predictions.

Jagadeesan et al. [32] evaluated the performance of their model on benchmark datasets, using manual labeling for one dataset and NLP models for automatic labeling on the other. Similarly, Lu et al. [31] created datasets by manually labeling Brexit-related tweets and automatically labeling tweets related to the EgyptAir Flight crash.

3.3. Data Source

According to the reviewed studies, some proposed solutions utilized multiple data sources, while others employed a single data source.

3.3.1. Single-Sources

Social media data, especially from Twitter, has been widely utilized in research because it provides valuable insights into public opinions and behavior [23]. Most crisis management studies in this review have relied on Twitter for their experiments [6,16,17,18,19,20,22,23,25,26,28,29,30,32,34,35,36,37,39,40,41,44]. Researchers preferred Twitter for several reasons: it is widely used, allows individuals to express their opinions, generates significant volumes of real-time data, and is readily available through its API [29].

Other studies utilized single data sources beyond Twitter. For example, Wang et al. [43] utilized news content from the Daily Mail’s mobile phone service to build a large-scale multimodal dataset with annotated summaries. Furthermore, Tian et al. [21] and Wan et al. [27] conducted their experiments on data collected from Sina Weibo, which is a social media platform in China.

3.3.2. Multi-Sources

In contrast to single-source approaches, few studies used more than one data source in their proposed solutions for crisis management. For example, Chaudhuri et al. [33] addressed the problem of lacking valuable information during disaster events by utilizing image data from smart infrastructure and social media platforms, such as Twitter and Facebook. They obtained geotagged images from earthquake-damaged areas, offering a comprehensive understanding of the damaged areas.

Humaira et al. [42] collected images from social media platforms and web sources, including Google search engines and GitHub repositories, to perform image analysis and assess flood risks. Dou et al. [24] obtained datasets from social media platforms (Twitter and Weibo), disaster-related damage data, and typhoon-related geographic data. Yang et al. [38] utilized news media and social media platforms, especially Flickr, emphasizing the complementarity of different data domains for discovering and describing events. Another study by Lu et al. [31] utilized data from two social media platforms, Twitter and Weibo, and found that there has been a high reliance on user-generated content during the dissemination of real-time information in recent years.

As shown in

Table 2. Summary of data and learning aspects in AI-driven crisis management solutions.

Main Aspect	Sub-Category	Type	Representative Studies	Key Applications/Findings
Data Type (30 studies)	Unimodal (20 studies)	Textual (18 studies)	Tweet classification: [6,16,17,18] Situational awareness: [19,20,21,22,23,24] Information extraction: [25,26,30,31,32] Sentiment analysis: [27,28,29]	– Crisis-specific tweet classification – Sub-event detection – Sentiment analysis – Situational awareness – Real-time monitoring
		Visual (2 studies)	Damage assessment: [33,34]	– Damage assessment – Resource prioritization – Severity assessment
	Multimodal (10 studies)	Text + image, Text + geolocation, Multiple modalities	Situational awareness: [35,36,37] Event summarization: [43,44] Multimodal classification: [38,39,40,41,42]	– Enhanced situational awareness – Humanitarian information – Comprehensive event summaries – Outperforms unimodal approaches
Data Labeling (30 studies)	Manual (16 studies)	Crowdsourcing, Domain experts	Visual data: [33,34] Social media content: [6,21,22,27,30,35,36] Pre-existing datasets: [16,17,19,25,28,38,39]	– Uses datasets: CrisisLex, CrisisNLP, CrisisMMD – High quality but time consuming – Domain expertise required
	Automatic	ML techniques	Uses supervised and unsupervised machine learning techniques	– Clustering – Classifier-based labeling – Efficient for large datasets
	Hybrid (14 studies)	Active learning, Semi-supervised	Real-time processing: [18,23,29,37,42] Optimized labeling: [20,24,26,31,32,40,41,43,44]	– Combines manual and automatic – Reduces workload – Improves efficiency – Iterative improvement
Data Source (30 studies)	Single-source (25 studies)	Twitter (22), Sina Weibo (2), News sources (1)	Twitter: [6,16,17,18,19,20,22,23,25,26,28,29,30,32,34,35,36,37,39,40,41,44] Sina Weibo: [21,27] News sources: [43]	– Twitter predominant due to API – Public opinions and behavior – Real-time data availability – Focus on specific demographics
	Multi-source (5 studies)	Social media, Smart infrastructure, Geographic data, Web sources	Cross-platform: [24,31,33,38,42]	– Comprehensive understanding – Complementary information – Enhanced crisis assessment – Broader coverage
Learning Model (30 studies)	Traditional ML (5)	SVM, RF, NB, SMM, SVR	Classification: [6,17] Relevance filtering: [18] Early warning: [29] Event discovery: [38]	– Binary and multi-label classification – Feature weighting and selection – Relevance classification – Effective for limited datasets
	Deep learning (16)	CNN, BERT, LSTM, ResNet, Bi-LSTM, Transformer, GAT	Textual analysis: [16,20,21,23,25,26,27,31] Multimodal: [35,36,37,39,40,43] Real time: [34,42]	– Multimodal data processing – Sentiment analysis – Real-time classification – Enhanced feature extraction – Improved performance
	Hybrid (9)	ML+DL, MOEA, DANN, SVM+CNN, ESVM-ELM	Multimodal fusion: [44] Domain adaptation: [28] Damage assessment: [19,33,41] Enhanced classification: [22,24,30,32]	– Leverages multiple approaches – Better performance on complex tasks – Multi-objective optimization – Enhanced pipeline
Learning Domain (30 studies)	In-domain (23)	Same disaster event for training/testing	Disaster-specific: [6,16,18,29,32,35,36,37] Event-targeted: [19,20,21,22,23,24,25,27,31,39,40,41,42,43,44]	– Better performance metrics – Event-specific features – Requires new labels per event – Risk of overfitting
	Cross-domain (4)	Different events for training/testing	Transfer learning: [30,33] Generalization: [34,38]	– Cross-event generalization – Less need for new data – Lower performance than in-domain
	Hybrid (3)	Mix of in-domain and cross-domain	Domain adaptation: [17,26,28]	– Combines specificity/generalization – Effective with limited data – Improved adaptability
Learning Phase (30 studies)	Pre-disaster (1)	Early warning systems	Informative data: [29]	– Proactive disaster management – Early landslide warnings – Signal identification
	During-disaster (22)	Real-time monitoring, awareness	Real-time classification: [18,23,26,32,34,42] Situational awareness: [20,27,28,35,36,37,38,41] Event-specific: [6,16,17,22,24,30,39,44]	– Real-time classification – Situation assessment – Resource allocation – Public sentiment monitoring
	Post-disaster (7)	Damage assessment, recovery	Damage evaluation: [19,33,40] Recovery support: [21,25,31,43]	– Infrastructure damage analysis – Sub-event detection – Rumor prediction – Post-crisis planning
Learning Type (30 studies)	Batch (20)	Retrospective data analysis	Classification: [6,16,17,25,27,36,37,40,43] Analysis: [20,21,22,24,30,31,33,35,39,41,44]	– Complete data analysis – Timeline reconstruction – Model training on historical data – Higher accuracy but delayed
	Real time (2)	Data stream analysis	Stream processing: [34,42]	– Immediate processing – Streaming analytics – Real-time flood severity – Rapid response
	Hybrid (8)	Batch + continuous stream processing	Adaptive systems: [18,19,23,26,28,29,32,38]	– Initial training with history – Continuous updates – Adaptive learning – Balanced performance
Learning Objective (30 studies) *	Identifying informative data (5)	Relevance classification, info extraction	Filtering: [6,18,23,29,39]	– Relevance filtering – Enhance awareness – Filter noise
	Sentiment analysis (3)	Opinion mining, emotional analysis	Public opinion: [27,28,37]	– Sentiment tracking – Domain-specific emotion – Aid emergency response
	Event detection (6)	Anomaly detection, pattern recognition	Event discovery: [20,22,30,31,38,42]	– Sub-event detection – Cross-event generalization – Hashtag and pattern mining
	Event summarization (3)	Static, real-time summarization	Summary generation: [26,43,44]	– Info overload reduction – Multimodal summaries – Dynamic cluster updates
	Damage assessment (6)	Infrastructure impact analysis	Damage evaluation: [17,19,24,33,34,41]	– City-level evaluation – Visual/textual damage inputs – Correlation with loss
	Humanitarian info (3)	Multi-category classification	Aid-related: [22,25,36]	– Humanitarian type prediction – Actionable info extraction – Text + image usage
Learning Objective (30 studies) *	Disaster classification (3)	Multi-label, crisis type identification	Disaster typing: [16,32,40]	– Crisis category prediction – Visual–text fusion – Real-time multi-labeling
	Rumor prediction (1)	User behavior analysis	Rumor propagation: [21]	– Predicting fake news spread – Early rumor detection – Post-disaster safety
	Flood tracking (1)	Flood phase monitoring	Flood monitoring: [35]	– Stage-based flood analysis – Text-image fusion – Enhanced awareness

* One study addressed two different objectives and is counted in both.

, AI-driven crisis management solutions utilize various data aspects, including different data types, labeling methods, and sources. This comprehensive overview highlights the predominance of single-source approaches, the effectiveness of hybrid labeling methods, and the superior performance of multimodal data analysis compared to unimodal approaches for complex crisis situations.

3.4. Analysis and Discussion of Data Aspect

To address Research Question 4 (RQ4), this subsection analyzes and interprets the key findings derived from the data aspect of reviewed studies, highlighting observable trends, relationships, and performance implications. As illustrated in Figure 6a, textual data analysis dominates the research landscape in AI-driven crisis management, constituting approximately 60% of the reviewed studies. This prevalence can be attributed to the abundance and accessibility of text-based data from social media platforms. However, Figure 7 reveals an emerging trend toward multimodal approaches in recent years, indicating a growing recognition of the value in integrating multiple data types for more comprehensive crisis assessment.

Figure 6b demonstrates a clear preference for hybrid labeling methods, combining the accuracy of manual annotation with the efficiency of automatic techniques. Interestingly, no studies relied solely on automatic labeling, highlighting the current limitations of fully automated annotation processes for crisis data.

The distribution of data sources, depicted in Figure 6c,d, reveals a strong dependence on Twitter, which accounts for approximately 73% of single-source studies. This quantitative evidence underscores the significant reliance on single-source approaches in AI-driven crisis management solutions.

Figure 8 illustrates the relationship between data types and labeling methods with textual data using manual labeling representing the most common approach in the literature. However, the increasing presence of multimodal data with hybrid labeling suggests a shift in research focus toward more sophisticated analytical frameworks.

The performance comparison in Figure 9 confirms the superior effectiveness of multimodal approaches, which achieve approximately 7.6% higher F1-scores compared to textual-only analysis. Finally, Figure 10 presents a multidimensional analysis of challenges and benefits across different data approaches, highlighting that while multimodal methods offer enhanced information richness and accuracy, they face increased processing complexity and implementation costs.

Figure 10 reveals the significant trade-offs among data processing approaches in disaster management systems. Multimodal approaches, while offering superior information richness (5) and accuracy (4), demand substantially higher processing complexity, implementation costs, and data volume requirements. This stands in stark contrast to textual approaches, which excel in real-time processing capabilities (5) compared to their visual and multimodal counterparts (both 2). The information richness dimension demonstrates a clear hierarchy, with multimodal providing the most comprehensive information (5), which is followed by visual (4), and textual approaches offering the least (2). Resource requirements follow a predictable progression across all three approaches with textual requiring the fewest resources, visual demanding moderate resources, and multimodal consuming the highest levels of processing complexity, implementation cost, and data volume. These relationships highlight the crucial decisions that disaster management teams must make when selecting suitable data processing methodologies, taking into account their specific operational constraints and information requirements.

Figure 6, Figure 7, Figure 8, Figure 9 and Figure 10 collectively provide a comprehensive, multi-perspective analysis of the data aspects in AI-driven crisis management solutions. Each figure presents a distinct analytical view, covering distribution, temporal evolution, cross-variable relationships, quantitative performance, and qualitative trade-offs. Collectively, these figures offer complementary insights that enhance the depth and clarity of the data aspects analysis.

These results suggest that both technological and operational constraints shape the evolution of AI-driven crisis data strategies. The dominance of text-based analysis stems from its low resource cost and the historical accessibility of Twitter data through public APIs. However, the growing trend toward multimodal integration indicates an effort to overcome the contextual limitations of unimodal inputs, particularly in detecting situational nuances such as infrastructure damage or humanitarian needs. The preference for hybrid labeling highlights the need to balance automation with human supervision, as crisis data often contain ambiguity and misinformation that automated systems alone cannot resolve. Overall, the transition toward multimodal and hybrid data frameworks reveals a field that is gradually prioritizing interpretability, inclusiveness, and cross-platform resilience, which represent critical characteristics for next-generation crisis management solutions.

4. Learning Aspect

Another critical dimension in AI-driven crisis management solutions is the learning aspect. To further address Research Question 3 (RQ3), this section examines the second dimension of the proposed taxonomy, which explores how diverse learning strategies structure and optimize AI-driven crisis management solutions. It focuses on various elements, as shown in Figure 11:

• Learning models: Traditional machine learning and deep learning;
• Learning domains: in-domain and cross-domain;
• Learning phases: pre-disaster, during-disaster, and post-disaster
• Learning types: batch and real time;
• Learning objectives: event detection, damage assessment, sentiment analysis, etc.

These elements collectively determine how well a proposed solution can adapt to, predict, and act upon dynamic crisis situations. In the following subsections, we explore these essential aspects, discussing and synthesizing proposed solutions from the reviewed studies in the context of each aspect.

4.1. Learning Model

Text and multimedia processing require practical AI algorithms to handle the overwhelming amount of data generated during crises [47]. Several studies have proposed crisis management solutions utilizing AI algorithms, including traditional machine learning and deep learning approaches. These algorithms are detailed below.

4.1.1. Traditional Machine Learning Algorithms

Several studies have emphasized the effectiveness of ML algorithms in crisis management. For example, Madichetty et al. [17] proposed a binary and multi-label classification method to assess disaster damage. To identify features related to damage assessment, they applied low-level lexical features, the most frequent features, and syntactic features. They then weighted these features using simple linear regression (SLR) and support vector regression (SVR) algorithms, and they applied the random forest (RF) algorithm for classification. Similarly, Ghafarian et al. [6] employed a support measure machine (SMM) to identify potentially informative tweets through learning on distributions, which is an advanced machine learning concept.

Kaufhold et al. [18] used RF for the relevance classification of social media posts. Similarly, Demirbaga [29] proposed the HTwitt framework based on the Hadoop ecosystem with the naïve Bayes (NB) classifier and n-Grams language model using the Tf-Idf and log-likelihood ratio similarity to classify tweets related to landslides. Yang et al. [38] proposed an In-domain and Cross-domain Laplacian Regularization (ICLR) model to jointly learn data representations for both the news media domain and images shared on social media.

4.1.2. Deep Learning Algorithm

Recent research has demonstrated the effectiveness of deep learning algorithms for crisis management. Several studies have utilized DL to analyze textual data. For instance, Xie et al. [16] developed a supervised contrastive learning framework with data reconstruction for the multi-label classification of disaster information, thereby improving the model’s ability to classify disaster-related data. Tian et al. [21] proposed a Rumor-Convolutional Neural Network (R-CNN) model that uses various features derived from users’ historical tweets to make rumor predictions.

Wan et al. [27] proposed the Emotion–Cognitive Reasoning Integrated BERT (ECR-BERT) model, which combines an emotion model with DL and reliable auxiliary knowledge to improve BERT performance in sentiment analysis. Lu et al. [31] proposed a hashtag-based sub-event detection framework using the Text-Convolutional Neural Network (Text-CNN) model with an attention mechanism to obtain more features from hashtags. Zahera et al. [25] developed an approach known as I-AID for multi-label tweet classification. Their system involves BERT for text encoding, the Graph Attention Network (GAT) for capturing structural information, and Relation Networks to learn the similarities between tweet and label vectors. Dhiman et al. [20] proposed an approximate graph-based global event detection method utilizing Joint Spherical Embeddings (JoSEs) for feature extraction, capturing contextual information, and reducing computational time.

Other studies employed DL algorithms for multimodal analysis. Wu et al. [36] proposed a multimodal humanitarian information identification model, which used BERT for text processing and a Deep-Convolutional Neural Network (Deep-CNN) for image processing. Bryan et al. [37] introduced a new flood prediction and tracking model using the transformer network. It compared several DL techniques to analyze images and text data, such as Contrastive Language–Image Pretraining (CLIP), the Transformer, Long Short-Term Memory (LSTM), and Residual Neural Network (ResNet-50). Similarly, Wang et al. [35] focused on extracting information from visual and textual contents of social media for flood monitoring. They employed deep learning-based named entity recognition (NER), specifically NeuroNER, and TensorFlow-based classification schemes utilizing Convolutional Neural Network (CNN) and ResNet architectures.

Wang et al. [43] proposed a crisis event summary generation model based on hierarchical multimodal fusion. Text feature vectors and guidance vectors are obtained using bidirectional long short-term memory (Bi-LSTM) and LSTM, while the Transformer obtains visual feature vectors. Hossain et al. [40] proposed a multimodal disaster identification system that used ResNet-50 for the image processing of disaster scenes and Bi-LSTM with attention for text analysis. Khattar and Quadri [39] proposed a Cross-Attention MultiModal (CAMM) deep neural network to classify multimodal disaster data. The model utilized the pretrained Visual Geometry Group Network (VGG-16) to extract features from input images and a two-layer Bi-LSTM to learn hidden dimension representations.

DL algorithms have also proven effective in real-time processing. Snyder et al. [23] presented an interactive learning framework that utilizes a CNN to identify relevant tweets in real time, thereby supporting situational awareness. Humaira et al. [42] suggested an end-to-end pipeline called DX-FloodLine for real-time flood detection based on deep neural networks. The system’s primary component is a hybrid neural network that combines VGG-16 and LSTM for flood detection and object identification. Zhang et al. [34] developed a Crowd–AI Dynamic Neural Architecture Search (CD-NAS) framework, which dynamically adjusts neural network architectures for real-time disaster damage assessment by combining AI and crowdsourcing. Bansal et al. [26] introduced a deep learning classification model using CNN for situational tweet classification and real-time summarization.

4.1.3. Hybrid Algorithms

Some research has been conducted to incorporate both traditional ML and DL to leverage the strengths of both approaches. Saini et al. [44] implemented a multi-objective optimization-based evolutionary algorithm (MOEA) and dense captioning for images to extract and summarize information from multimodal data in disaster events. Yao and Wang [28] presented a domain-specific sentiment classification approach named DSSA-H, using RF for collecting hurricane-relevant tweets and the Domain Adversarial Neural Network (DANN) for sentiment classification.

Chaudhuri et al. [33] compared the performance of SVM, ANN, and three variants of CNN, including AlexNet, InceptionV3, and ResNet-50, to classify images based on the presence or absence of potential survivors. In a different approach, Jagadeesan et al. [32] proposed a novel ML model, applied Apache Spark for big data processing, and incorporated NLP with TF-IDF for feature extraction. To classify disaster events at the city council level, they employed the city council evolution (CCE)-optimized ensemble support vector machine-based extreme learning machine (ESVM-ELM).

Fan et al. [22] proposed a hybrid machine learning pipeline integrating NER for location identification, a fine-tuned BERT model for post-classification with humanitarian categories, and graph clustering for reliable situational information extraction. Belcastro et al. [30] extracted sub-events as secondary effects of a disaster event. They applied neural networks for textual classification and post filtering, geocoding services, and NLP for enrichment and sub-event identification, utilizing the density-based spatial clustering (DBSCAN) algorithm and text analysis.

Avvenuti et al. [19] proposed the HERMES system that relies on Recurrent Convolutional Neural Networks (RCNNs) for text categorization and SVM for the identification of witnesses to enrich disaster-related data collected from social media. Kamoji and Kalla [41] developed the Bidirectional Multilayer Perceptron (BMLP) and Sparse Denoising Autoencoder–Hyperbolic Hopfield Neural Network (SDAE-HHNN) approach for effective flood prediction. Dou et al. [24] proposed a method for estimating disaster losses using social media data. They employed a Continuous Bag-of-Words (CBOW) model from Word2Vec for word embedding, graph-based methods for clustering, and Latent Dirichlet Allocation (LDA) for topic detection.

4.2. Learning Domain

The proposed solutions conduct the experiment based on the following learning domains: in-domain and cross-domain approaches. Below, we outline the major differences between the in-domain and cross-domain approaches.

4.2.1. In-Domain Approach

In-domain refers to training and testing a model on datasets from the same disaster event [17].

In-domain experiments can leverage similarities within a disaster event but may overfit event-specific nuances [17].

Many studies have been evaluated using an in-domain approach [6,16,18,19,20,21,22,23,24,25,27,29,31,32,35,36,37,39,40,41,42,43,44]. This approach utilizes disaster-specific data, resulting in better performance, but it requires new labeled data for each subsequent disaster.

4.2.2. Cross-Domain Approach

A cross-domain approach exploits data from past disaster events to train a model and then tests the model on a new and unseen disaster event [17]. This approach aims to discover more general disaster-related features that transfer across disasters, but it is challenging to capture event-specific characteristics. Pretraining cross-domain models on multiple historical events may improve their generalization to new events [3].

Indeed, few studies have followed a cross-domain approach for their experiments [30,33,34,38]. For example, Chaudhuri et al. [33] applied transfer learning, in which a machine learning model is trained for one task and then reused for another task. This technique enhances the target network’s capacity, even without a large target dataset, thereby overcoming the limitations of small training sets and reducing overfitting.

4.2.3. Hybrid Approach

Some studies leverage the advantages of both in-domain and cross-domain approaches. Madichetty et al. [17] proposed a method for detecting damage assessment tweets by utilizing both in-domain and cross-domain approaches, which is applicable in situations where enough labeled tweets are not available and when specific disaster-type tweets are not available. Yao and Wang [28] applied the in-domain approach by using a domain-specific sentiment analysis of hurricane-related tweets. They also applied cross-domain learning using a domain-adversarial neural network (DANN) algorithm. These neural networks are used to generalize the model on different domains, such as various disaster contexts.

Bansal et al. [26] proposed a model that works on both in-domain and cross-domain classification. They found that the model showed better performance in cross-domain datasets because it had been enriched with varied disaster-related information during the training phase. The model was tested on tweets from other disaster events, proving its cross-domain classification ability.

In summary, while in-domain experiments usually perform better, they require new training data for each disaster event. Cross-domain experiments, on the other hand, aim for generalization but typically perform significantly worse than in-domain experiments. Nevertheless, the results obtained from both in-domain and cross-domain settings demonstrate the ability of the proposed approaches to perform well with and without labeled data from specific disaster datasets, respectively [17].

4.3. Learning Phase

Based on the reviewed studies, we observed that the proposed solutions in the crisis management field focused mainly on the following disaster phases.

4.3.1. Pre-Disaster Phase

The pre-disaster phase involves proactive strategies in disaster management that are primarily focusing on early warning, signal identification, and severity prediction [7].

Among the reviewed studies, Demirbaga [29] focused on the pre-disaster phase specifically for early warning systems regarding landslides. This study proposed a Hadoop-based platform for analyzing and visualizing streaming Twitter data.

4.3.2. During-Disaster Phase

Real-time monitoring and situational awareness are crucial for effective crisis management [7]. Extracting relevant insights from massive amounts of streaming social media data has always been a challenge [2].

Several studies have addressed these challenges using real-time data analysis and situational awareness on social media. Snyder et al. [23] utilized an interactive learning approach to identify relevant tweets for real-time situational awareness. Zhang et al. [34] contributed to real-time analysis by designing a crowd-driven dynamic neural architecture that enables real-time damage assessment via social sensing, relying on data collected from the public. Bansal et al. [26] developed a framework for classifying and summarizing situational tweets during disasters to assist immediate response efforts. Kaufhold et al. [18] developed a Real-Time Relevance Classification Model using an active incremental online adaptation algorithm to enhance the efficiency of disaster responses. Jagadeesan et al. [32] designed a real-time social media analysis model to improve the categorization of disaster-related data.

Real-time flood prediction and monitoring have also been vital research areas. Humaira et al. [42] designed a pipeline that detected scenes of real-time floods utilizing multimedia images from social media. Kamoji and Kalla [41] developed a flood prediction model that utilizes text and image analysis from Twitter, enabling real-time updates during ongoing flood events. Wang et al. [35] developed a photo classifier and text geolocator to detect flooding phases, aiding disaster response teams.

Moreover, several studies [6,16,17,20,22,24,27,28,30,36,37,38,39,44] have focused on utilizing social media data to achieve other objectives during the disaster phase such as improving event detection, assessing public sentiment for situational awareness, identifying humanitarian information, and correlating social media topics with damage severity. These studies utilized techniques such as cross-domain representations, domain-specific sentiment analysis, and multimodal data fusion to propose classification and summarization models tailored for social media data during disaster scenarios.

4.3.3. Post-Disaster Phase

Community recovery from significant disasters is often challenging without knowledge about the locations and severity of damage to critical public infrastructure, such as electricity, roads/transportation, drinking water, buildings, and telecommunication networks [1].

Various studies [19,21,25,31,33,40,43] have attempted to provide better response and recovery for the post-disaster phase. For instance, Chaudhuri et al. [33] utilized image data obtained from social media and smart infrastructures to identify damage to objects more reliably immediately after disasters. Lu et al. [31] proposed a framework to identify sub-events within the broader post-disaster scenario, aiming to improve situational awareness. Tian et al. [21] developed a method to predict the spread of rumors on social media following public emergencies. Zahera et al. [25] identified that filtering posts from social media data is crucial for finding actionable information for disaster relief.

4.4. Learning Type

For crisis management and making decisions during disasters and emergencies, emergency services need to obtain a comprehensive and rapid overview of crisis events [18]. Therefore, the studies conducted in this context aim to propose a solution that supports the following learning types: batch learning and real-time learning [46].

4.4.1. Batch Learning

Retrospective data analysis begins with a batch of data relevant to an event, which may comprise messages from the entire period of interest. For example, we could start by recreating a timeline of events in the aftermath of an earthquake by examining all messages from the moment the earthquake occurred to two weeks later [46]. Several works [6,16,17,20,21,22,24,25,27,30,31,33,35,36,37,39,40,41,43,44] have been conducted by performing their experiments and model training on pre-collected datasets and did not involve real-time processing.

4.4.2. Real-Time Learning

Real-time learning is a critical approach, especially during crises, to analyze data streams as they arrive. Although this approach allows for the timely and rapid analysis of incoming data, it typically operates with an incomplete understanding of events and their potential consequences [46]. For example, Humaira et al. [42] used social media streaming services to determine real-time flood severity and object detection in flooded areas. Their system combines incremental model training with real-time image analysis, enabling a near-real-time response by analyzing streaming flood images and updating the model with incoming data.

Similarly, Zhang et al. [34] introduced a CD-NAS framework for real-time disaster damage assessment. They aimed to integrate AI and human intelligence on a crowdsourcing platform by streaming neural networks with dynamically related architectures for each new image in the stream.

4.4.3. Hybrid Approach

In the hybrid approach, models are initially trained on historical data to provide reliable learning while being continuously updated through streaming data from social media platforms to detect real-world events as they unfold. These models can process data streams in real time or work in a retrospective analysis fashion, depending on the urgency of the situation [46].

Many studies have applied a hybrid approach in their proposed solutions [18,19,23,26,28,29,32,38]. For example, Kaufhold et al. [18] combined batch and online learning for real-time relevance classification on social media platforms. Their system integrates active, incremental, and online learning to adapt to data streams and enhance the efficiency of disaster response.

Snyder et al. [23] proposed an interactive learning framework within the Social Media Analytics and Reporting Toolkit, SMART 2.0, to iteratively classify streaming social media data in real time, providing higher situational awareness during crisis events. Bansal et al. [26] introduced a framework that uses DL to classify and summarize situational tweets in real time. Initially, tweets are clustered using a multi-objective optimization algorithm. As new tweets arrive, the clusters are updated dynamically, allowing for real-time summarization.

Some studies have used Big Data technology for real-time learning. For example, Jagadeesan et al. [32] reduced the hardware and communication overhead in big-data classification tasks by utilizing parallel processing with the Apache Spark framework. Similarly, Demirbaga [29] developed a Hadoop-based framework for data collection and real-time visualization.

4.5. Learning Objective

When using social media for crisis management and emergency analysis, several objectives have been identified in state-of-the-art research as the most frequently investigated in currently proposed solutions, as shown in Figure 12. It is worth noting that an article may have multiple objectives. The objectives that have received the most attention in the literature are detailed below.

4.5.1. Identifying Informative Data

During disasters, obtaining or identifying relevant and valuable information from the large social media data stream is an essential task [48]. Informative data contain the most up-to-date information about the affected area, such as providing the number of victims or other vital information. Non-informative data express sympathy, feelings, and analysis after the disaster [44].

Several studies [6,18,23,29,39] have proposed methods to identify informative data from social media messages. For example, Khattar and Quadri [39] proposed a framework that integrates complementary information provided by multimodal social media messages to classify disaster tweets into ‘informative’ and ‘non-informative’ classes. Kaufhold et al. [18] presented a system that can immediately categorize social media messages as relevant or irrelevant during disasters. Similarly, Snyder et al. [23] developed a real-time, interactive learning framework that enhances the classification of relevant tweets to improve situational awareness. This framework enables users to iteratively correct the relevance of tweets in real time, thereby training the classification model for immediate predictive improvements.

4.5.2. Sentiment Analysis

Sentiment analysis is a method for systematically identifying, extracting, and analyzing people’s opinions, attitudes, and sentiments from a large amount of data [49]. By examining texts, punctuation, and emotions, sentiment analysis can automatically classify people’s sentiment polarity or place them on continuous scales [50]. This task is particularly beneficial during crises, as it serves multiple functions related to monitoring public sentiment and reactions, facilitating an understanding of public issues, and enabling the evaluation of the timeliness and effectiveness of governmental responses to these issues [4].

Several studies have focused on sentiment analysis. For instance, Bryan-Smith et al. [37] assess the severity of flooding situations based on the sentiment analysis of multimodal inputs. Yao and Wang [28] developed a domain-specific sentiment analysis approach specifically for tweets posted during hurricanes. This approach can efficiently classify sentiments in hurricane-related tweets, increasing situational awareness and aiding emergency response and disaster management. Wan et al. [27] developed a sentiment analysis model to analyze online public opinions during disasters.

4.5.3. Event Detection

With the increase in the availability of real-time online data streams, event detection methods have become the primary and fastest mode of identifying anomalous patterns or trends in different communities, often earlier than traditional media [20].

An event is generally defined as the occurrence of something significant at a specified time and location. In the context of social media, an event is specifically defined as “an occurrence causing changes in the volume of text data that discusses the associated topic at a specific time. This occurrence is characterized by topic and time and is often associated with entities such as people and location” [46].

Several crisis management solutions [20,22,30,31,38,42] have focused on event detection. For example, the authors in [30] proposed a new method that focuses on discovering sub-events that can occur as secondary effects of a disaster. This method can be integrated with existing systems to coordinate and enhance emergency response. Yang et al. [38] proposed a model for event discovery that utilizes data distributed across multiple media domains, including news media and social media. Another study by Humaira et al. [42] focused on real-time flood scene detection and analysis using social media data.

4.5.4. Event Summarization

Another objective for crisis management in the context of social media is generating event summaries. This approach aims to address information overload by providing concise overviews of evolving events. Although some tools are already available, the increasing volume of online information makes it challenging to generate meaningful and timely summaries [2].

Several studies have focused on event summaries, utilizing both image and text data simultaneously to generate summaries from microblog data during disaster events. Wang et al. [43] proposed a hierarchical multimodal crisis event summary generation model based on modal alignment and hierarchical thinking. Similarly, Saini et al. [44] aimed to generate summaries from multimodal data during disaster events. The benefit of a summary from multimodal data compared to a summary from only textual data is that it adds complementary information to textual tweets, making them more informative.

Existing research [43,44] focused primarily on developing static summarization techniques. However, real-time summarization during disaster events is crucial due to the continuously evolving nature of tweets. Bansal et al. [26] developed a novel framework using a deep learning-based classification model that filtered situational tweets from others and summarized them in real time.

4.5.5. Damage Assessment

The damage assessment task is one of the critical situational awareness steps for humanitarian organizations, concerned with determining the extent and severity of damage produced by a crisis event, which may include infrastructure damage, property damage, injuries, and environmental impacts [4,17].

Many research studies [17,19,24,33,34,41] have focused on exploiting social media data in damage assessment. For example, Avvenuti et al. [19] developed a system that improves the quality and quantity of social crisis data by combining opportunistic and participatory sensing for damage assessment. Dou et al. [24] proposed a method for disaster damage assessment by extracting fine-grained topics from social media data and analyzing their correlation with damage losses at the city level.

Streaming disaster damage assessment is another crucial task during disasters, aiming to automatically assess the severity of damage in affected areas in real time by leveraging streaming data posted on social media. Zhang et al. [34] developed a novel crowd–AI collaborative framework to accurately assess the damage severity of affected areas using streaming imagery data posted on social media.

4.5.6. Identifying Humanitarian Information

Social media plays a significant role in disaster management by providing valuable data about affected people, donations, and help requests.

Recent studies [22,25,36] highlight the need to filter information on social media into humanitarian categories, including infrastructure and utility damage, affected people, response, and others. Zahera et al. [25] developed a system that can automatically categorize tweets into multiple humanitarian information types to identify actionable or critical information relevant to disaster relief and mitigation. Similarly, Wu et al. [36] focused on extracting various humanitarian information categories from social media text–image pairs. They grouped humanitarian information into six categories: caution and advice, needs and offers, infrastructure and utility damage, affected people, response, and others. Each category is designed to meet the information needs of responders in different tasks.

4.5.7. Disaster Classification

Disaster classification is another crucial task addressed by proposed solutions in the field of crisis management. Xie et al. [16] utilized disaster text information to reconstruct the dataset, enabling multi-label disaster classification. Jagadeesan et al. [32] focused on real-time disaster prediction and classification. They efficiently implemented real-time disaster-related applications using big data technology to classify and predict disaster events for smart cities and civilizations. Hossain et al. [40] developed an effective computational model for identifying disaster-related information by integrating features from visual and textual modalities.

4.5.8. Other Objectives

Various studies have proposed solutions for other objectives in crisis management. For instance, Tian et al. [21] focused on predicting the rumor retweeting behavior of social media users during public emergencies. Wang et al. [35] attempted to enhance the situational awareness of flooding by tracking phase transitions and locating emergency incidents. Moreover, some studies address multiple objectives. For example, Fan et al. [22] proposed a hybrid machine learning pipeline to detect the unfolding of disaster events corresponding to different locations from social media posts. They also aimed to identify humanitarian information by classifying the posts into various humanitarian categories.

4.6. Analysis and Discussion of Learning Aspect

To further contribute to Research Question 4 (RQ4), this subsection synthesizes the main findings regarding learning strategies, model performance, and their implications for improving AI-based crisis management frameworks. Table 2 provides a structured overview of the reviewed studies categorized by learning aspect categories. It highlights the variety of techniques employed, ranging from traditional machine learning to advanced deep learning and hybrid models, as well as various learning objectives, including damage assessment, event detection, and sentiment analysis. The table also illustrates how different learning phases influence the choice of learning approach and objective. Our analysis of the learning aspects in AI-driven disaster management reveals several key trends and research directions. This section examines the distribution and relationships between various learning approaches, highlights gaps in the current research, and discusses promising future directions.

4.6.1. Distribution Analysis of Learning Models

Figure 13 illustrates the distribution of different learning models across the reviewed studies. Deep learning approaches dominate the current research landscape, accounting for 53% of the reviewed works. This trend reflects the increasing complexity of disaster data and the superior performance of deep learning in handling multimodal inputs. Traditional machine learning approaches continue to play a significant role (17%), particularly in scenarios with limited labeled data or where interpretability is crucial. Hybrid approaches (30%) demonstrate a growing recognition that combining the complementary strengths of different algorithms can yield more robust crisis management systems.

4.6.2. Learning Phase and Type Analysis

Figure 14 presents a radar chart illustrating the research focus across different learning phases and learning types. The concentration of studies in the during-disaster phase (73%) highlights the critical need for real-time situational awareness and decision support during crisis events. In contrast, the pre-disaster phase (3%) remains comparatively under-explored, suggesting opportunities for developing more proactive disaster management solutions. Regarding learning types, batch learning approaches (67%) remain prevalent, although hybrid learning approaches (26%) are gaining traction due to their ability to combine historical knowledge with real-time adaptability.

4.6.3. Learning Objectives Analysis

Figure 15 shows the distribution of primary learning objectives across the reviewed studies. Multiple studies addressed more than one objective, so the total exceeds 100%. Damage assessment (20%) and event detection (20%) emerge as the most common objectives, underscoring their critical role in crisis response. Identifying informative data (16%) also receives significant attention, which is followed by identifying humanitarian information (10%), disaster classification (10%), event summarization (10%), and sentiment analysis (10%). The relatively lower focus on rumor prediction (3%) and flooding tracking (10%) suggests an area for future research, particularly given its direct impact on relief operations.

4.6.4. Learning Domain Performance Comparison

Figure 16 compares the typical performance of different learning domain approaches based on reported F1-scores. In-domain approaches generally achieve the highest performance (average F1-score of 0.85) due to their ability to capture event-specific patterns. Cross-domain approaches demonstrate lower performance (with an average F1-score of 0.72) but offer greater generalizability. Hybrid approaches achieve a balance (average F1-score of 0.79), suggesting that combining domain-specific and transferable features can mitigate the performance gap while enhancing flexibility.

Overall, these findings highlight both technological progress and persistent limitations in AI-driven crisis learning systems. The dominance of deep learning reflects the field’s shift toward handling complex, multimodal inputs. However, its heavy data and computing demands restrict applicability in resource-constrained or low-data environments. The scarcity of pre-disaster research stems largely from the lack of annotated early-warning datasets and the unpredictability of future crises, suggesting the need for simulation-based and generative data augmentation strategies. The predominant focus on batch learning during the disaster phase highlights a reactive paradigm that needs to evolve toward adaptive learning frameworks to support real-time decision making. Similarly, the performance gap between in-domain and cross-domain models reveals a critical need for transfer learning, domain adaptation, and foundation models that generalize across diverse disaster types. Collectively, these findings suggest that future research should prioritize generalizable, data-efficient, and interpretable learning paradigms that can support proactive crisis management and robust real-time analytics.

5. Discussion, Limitations, and Future Directions

This section synthesizes our comprehensive review of AI-driven crisis management solutions, analyzing the interrelationships between data and learning aspects while identifying key trends, challenges, and future directions based on the 30 studies summarized in

Table 3. Summary of crisis management solutions based on their common aspects. The table highlights learning objectives, data types, learning phases, learning models, data labeling methods, learning types, data sources, and learning domains.

Ref.	Learning Objective	Data Type	Learning Phase	Learning Model	Data Labeling	Learning Type	Data Source	Learning Domain
[16]	Disaster classification	Textual	During-disaster	Supervised contrastive learning	Manual labeling	Batch learning	Twitter	In-domain
[17]	Damage assessment	Textual	During-disaster	SVR, SLR, and RF	Manual labeling	Batch learning	Twitter	Hybrid approach
[6]	Identifying informative data	Textual	During-disaster	SMM	Manual labeling	Batch learning	Twitter	In-domain
[18]	Identifying informative data	Textual	During-disaster	RF	Hybrid approach	Hybrid approach	Twitter	In-domain
[36]	Identifying humanitarian information	Multimodal	During-disaster	BERT, Deep-CNN	Manual labeling	Batch learning	Twitter	In-domain
[37]	Sentiment analysis	Multimodal	During-disaster	Transformer, LSTM, CLIP, ResNet-50	Hybrid approach	Batch learning	Twitter	In-domain
[33]	Damage assessment	Visual	Post-disaster	ANN, SVM, AlexNet, InceptionV3, ResNet-50	Manual labeling	Batch learning	Twitter, Facebook, Smart infrastructures	Cross-domain
[32]	Disaster classification	Textual	During-disaster	ESVM-ELM, TF-IDF	Hybrid approach	Hybrid approach	Twitter	In-domain
[30]	Event detection	Textual	During-disaster	Neural network, geocoding, NLP, DBSCAN	Manual labeling	Batch learning	Twitter	Cross-domain
[35]	Flooding phases tracking	Multimodal	During-disaster	NER, NeuroNER, ResNet, CNN	Manual labeling	Batch learning	Twitter	In-domain
[29]	Identifying informative data	Textual	Pre-disaster	NB, n-Grams, Log-likelihood similarity	Hybrid approach	Hybrid approach	Twitter	In-domain
[22]	Event detection and identifying humanitarian information	Textual	During-disaster	NER, BERT, graph-based clustering	Manual labeling	Batch learning	Twitter	In-domain
[31]	Event detection	Textual	Post-disaster	Text-CNN model	Hybrid approach	Batch learning	Twitter and Sina Weibo	In-domain
[23]	Identifying informative data	Textual	During-disaster	CNN	Hybrid approach	Hybrid approach	Twitter	In-domain
[40]	Disaster classification	Multimodal	Post-disaster	ResNet-50 and Bi-LSTM	Hybrid approach	Batch learning	Twitter	In-domain
[42]	Event detection	Multimodal	During-disaster	VGG-16, LSTM	Hybrid approach	Real-time learning	Twitter and Web sources	In-domain
[43]	Event summarization	Multimodal	Post-disaster	Bi-LSTM, LSTM, Transformer	Hybrid approach	Batch learning	Daily Mail’s service	In-domain
[44]	Event summarization	Multimodal	During-disaster	MOEA and dense captioning	Hybrid approach	Batch learning	Twitter	In-domain
[24]	Damage assessment	Textual	During-disaster	Word2Vec, CBOW, graph-based, LDA	Hybrid approach	Batch learning	Twitter, Sina Weibo, Damage, and Typhoon data	In-domain
[27]	Sentiment analysis	Textual	During-disaster	BERT	Manual labeling	Batch learning	Sina Weibo	In-domain
[19]	Damage assessment	Textual	Post-disaster	SVM, RCNNs	Manual labeling	Hybrid approach	Twitter	In-domain
[34]	Damage assessment	Visual	During-disaster	CD-NAS	Manual labeling	Real-time learning	Twitter	Cross-domain
[28]	Sentiment analysis	Textual	During-disaster	RF, DANN	Manual labeling	Hybrid approach	Twitter	Hybrid approach
[38]	Event detection	Multimodal	During-disaster	Laplacian Regularization	Manual labeling	Hybrid approach	Flickr and News media	Cross-domain
[20]	Event detection	Textual	During-disaster	JoSE	Hybrid approach	Batch learning	Twitter	In-domain
[21]	Rumor prediction	Textual	Post-disaster	R-CNN	Manual labeling	Batch learning	Sina Weibo	In-domain
[25]	Identifying humanitarian information	Textual	Post-disaster	BERT, GAT, Relation Network	Manual labeling	Batch learning	Twitter	In-domain
[39]	Identifying informative data	Multimodal	During-disaster	Bi-LSTM, VGG-16	Manual labeling	Batch learning	Twitter	In-domain
[41]	Damage assessment	Multimodal	During-disaster	BMLP, SDAE, HHNN	Hybrid approach	Batch learning	Twitter	In-domain
[26]	Event summarization	Textual	During-disaster	CNN	Hybrid approach	Hybrid approach	Twitter	Hybrid approach

and

Table 4. Comparative analysis of crisis management solutions across key design criteria. This table evaluates 30 studies based on six paired criteria categories: data type (unimodal/multimodal), learning domain (in-domain/cross-domain), learning model (traditional ML/DL), data labeling method (manual/automatic), learning type (real time/batch), and data source (single/multi-source).

Ref.	Data Type		Learning Domain		Learning Model		Data Labeling		Learning Type		Data Source
	Unimodal	Multimodal	In-Domain	Cross- Domain	Traditional ML	DL	Manual	Automatic	Real Time	Batch	Single- Source	Multi- Source
[16]	√	×	√	×	×	√	√	×	×	√	√	×
[17]	√	×	√	√	√	×	√	×	×	√	√	×
[6]	√	×	√	×	√	×	√	×	×	√	√	×
[18]	√	×	√	×	√	×	√	√	√	√	√	×
[36]	×	√	√	×	×	√	√	×	×	√	√	×
[37]	×	√	√	×	×	√	√	√	×	√	√	×
[33]	√	×	×	√	√	√	√	×	×	√	×	√
[32]	√	×	√	×	√	√	√	√	√	√	√	×
[30]	√	×	×	√	√	√	√	×	×	√	√	×
[35]	×	√	√	×	×	√	√	×	×	√	√	×
[29]	√	×	√	×	√	×	√	√	√	√	√	×
[22]	√	×	√	×	√	√	√	×	×	√	√	×
[31]	√	×	√	×	×	√	√	√	×	√	×	√
[23]	√	×	√	×	×	√	√	√	√	√	√	×
[40]	×	√	√	×	×	√	√	√	×	√	√	×
[42]	×	√	√	×	×	√	√	√	√	×	×	√
[43]	×	√	√	×	×	√	√	√	×	√	√	×
[44]	×	√	√	×	√	√	√	√	×	√	√	×
[24]	√	×	√	×	√	√	√	√	×	√	×	√
[27]	√	×	√	×	×	√	√	×	×	√	√	×
[19]	√	×	√	×	√	√	√	×	√	√	√	×
[34]	√	×	×	√	×	√	√	×	√	×	√	×
[28]	√	×	√	√	√	√	√	×	√	√	√	×
[38]	×	√	×	√	√	×	√	×	√	√	×	√
[20]	√	×	√	×	×	√	√	√	×	√	√	×
[21]	√	×	√	×	×	√	√	×	×	√	√	×
[25]	√	×	√	×	×	√	√	×	×	√	√	×
[39]	×	√	√	×	×	√	√	×	×	√	√	×
[41]	×	√	√	×	√	√	√	√	×	√	√	×
[26]	√	×	√	√	×	√	√	√	√	√	√	×
Total	20	10	26	7	14	25	30	14	10	28	25	5
Percentage	67%	33%	87%	23%	47%	83%	100%	47%	33%	93%	83%	17%

√ = feature applied; × = not applied.

. In particular, this section addresses Research Question 5 (RQ5) by discussing the main challenges identified across the reviewed studies, such as data bias, generalizability, and real-time adaptability, and by outlining emerging research directions that aim to overcome these limitations.

5.1. Integrated Analysis of Crisis Management Approaches

Table 3 and Table 4 provide complementary perspectives on existing crisis management solutions. While Table 3 details the specific implementations across multiple dimensions, Table 4 offers a systematic binary comparison across six critical design criteria pairs. Our analysis reveals several patterns and relationships between data and learning aspects that significantly influence system performance and applicability.

5.1.1. Comparative Analysis of Design Criteria

Table 4 quantifies the distribution of design choices across the reviewed studies. The most prevalent approaches include the following:

• Data Type: Unimodal approaches (67%) dominate, but multimodal solutions (33%) show an increasing trend in recent years.
• Learning Domain: In-domain approaches (87%) are substantially more common than cross-domain approaches (23%), highlighting a limitation in generalizability.
• Learning Model: Deep learning techniques (83%) are preferred over traditional machine learning methods (47%) with some studies employing both approaches.
• Data Labeling: While all studies utilize manual labeling to some degree, only 47% incorporate automatic labeling techniques.
• Learning Type: Batch learning (93%) is overwhelmingly favored over real-time approaches (33%) with few studies implementing truly real-time solutions.
• Data Source: Single-source implementations (83%) predominate with multi-source approaches (17%) remaining underutilized despite their potential benefits.

5.1.2. Predominant Approaches and Their Effectiveness

As evident from Table 3, a clear trend emerges in the relationship between data types and learning models. Textual data analysis solutions (60% of studies in Table 3) predominantly employ traditional machine learning algorithms [6,17,18,19,22,24,28,29,30,32,33], while multimodal approaches (90%) almost exclusively rely on deep learning architectures [35,36,37,39,40,41,42,43,44]. This correlation highlights the inherent complexity of processing multiple data modalities, underscoring the need for the representational capabilities of deep neural networks.

Table 4 further confirms this relationship, showing that 90% of multimodal approaches utilize deep learning, compared to only 30% of unimodal approaches. This finding suggests that the complexity of integrating multiple data types typically necessitates more sophisticated model architectures.

Learning Phase Distribution: Table 3 reveals that the during-disaster phase dominates research focus (73% of reviewed studies) with solutions targeting real-time situation assessment [18,23,26,32,34,42]. The post-disaster phase (24%) emphasizes damage assessment and recovery planning [19,21,25,31,33,40,43],while pre-disaster applications (3%) remain significantly underexplored with only [29] focusing on early warning systems.

Learning Objectives and Data Types: Examining Table 3, it is evident that particular learning objectives demonstrate strong associations with specific data types. Damage assessment studies frequently employ visual or multimodal data [33,34,41], leveraging the rich information in images. Conversely, event detection and sentiment analysis predominantly rely on textual data [27,28,30], though multimodal approaches are emerging [37].

5.1.3. Temporal Evolution of Approaches

The chronological analysis of solutions in Table 3 reveals an evolutionary trajectory in crisis management approaches:

Earlier solutions primarily utilized unimodal data with traditional machine learning [6,17,18,19,22,24,28,29,30,32,33]. These foundational systems established basic frameworks for filtering relevant information but typically addressed single objectives.

A shift toward deep learning and hybrid approaches emerged [35,36,37,39,40,41,42,43,44], coinciding with an increased adoption of multimodal data. Solutions began addressing multiple simultaneous objectives and incorporated more sophisticated architectures.

Recent work shows significant advancement in real-time capabilities [18,19,23,26,28,29,32,34,38,42] and cross-domain applications [30,33,38]. The integration of foundation models and pretrained architectures has become prevalent, thereby improving generalizability across various disaster types.

5.2. Critical Analysis of Current Approaches

As evident from Table 3 and Table 4, existing approaches exhibit several limitations across data and learning aspects, as detailed in the following.

5.2.1. Data Aspect Limitations

Data Source Dependence: Table 4 quantifies the overwhelming reliance on single-source data (83% of studies) with Twitter serving as the primary platform in most cases, as detailed in Table 3. This raises concerns about generalizability and representativeness, particularly in developing regions most vulnerable to natural disasters. Moreover, the heavy dependence on Twitter introduces potential geographic and demographic biases, as its users are not globally representative, which may further limit the inclusiveness and generalizability of the derived findings. Multi-source approaches [24,31,33,38,42] demonstrate superior coverage but remain significantly underutilized.

Data Labeling Challenges: As shown in Table 4, while all studies utilize manual labeling to some degree, only 47% incorporate automatic labeling techniques. This dependence on manual annotation creates a significant bottleneck for large-scale, real-time crisis management systems. The highest-scoring solutions in Table 4 typically employ hybrid labeling approaches [18,20,23,24,26,29,31,32,37,40,41,42,43,44], suggesting this as a best practice for balancing quality and scalability.

Multimodal Data Integration: While Table 4 shows that multimodal approaches represent only 33% of studies, they demonstrate superior potential for comprehensive analysis. However, they face significant challenges in effectively aligning and integrating heterogeneous data types. Current fusion techniques [35,38,39,40,42] often have limitations in balancing the complex relationships between modalities, particularly when text and image content are only partially correlated or entirely independent. As a result, these models often fail to leverage the complementary strengths of each modality fully.

5.2.2. Learning Aspect Limitations

Domain Adaptability Challenges: Table 4 quantifies the dominance of in-domain approaches (87% of studies), which achieve high performance for specific disaster types but struggle with generalization. The limited number of cross-domain solutions [30,33,34,38] highlights a critical gap in the research landscape, as practical deployment requires adaptation across diverse crisis scenarios.

Real-time Processing Constraints: Despite the critical need for immediate response during crises, Table 4 shows that only 33% of studies employ real-time learning approaches with most of these implementing hybrid rather than pure real-time processing. The predominance of batch learning (93% of studies) limits applicability in evolving disaster situations where streaming data processing is essential. Notably, the few studies implementing pure real-time learning [34,42] focus primarily on visual data analysis.

Model Complexity vs. Explainability: As models increase in complexity (with deep learning approaches comprising 83% of studies in Table 4), they offer improved performance but reduced explainability. This trade-off is particularly problematic in high-stakes crisis decision making, where transparency and interpretability are essential for establishing stakeholder trust.

5.2.3. Limitations of the Review Process

Despite adhering to a rigorous and transparent methodology, this review has certain limitations. The search was restricted to English-language journal articles published between 2020 and 2024, which may have excluded relevant studies published in other languages or outside this timeframe. Furthermore, the database coverage was limited to four major sources: Springer, ScienceDirect, Taylor & Francis, and IEEE Xplore, so potentially relevant studies indexed in other databases may not have been captured.

5.3. Emerging Trends and Future Directions

Based on our comprehensive analysis of Table 3 and Table 4, we identify several promising research directions that address current limitations and leverage emerging technologies.

5.3.1. Enhanced Data Strategies

Diverse Data Ecosystems: Future research should address the 83% single-source dependency identified in Table 4 by exploring more diverse data sources beyond Twitter, following examples like [24,38] that incorporate multiple platforms, news media, and specialized datasets. This diversity would improve coverage, reduce bias, and enhance situational awareness.

Efficient Labeling Paradigms: Given that only 47% of studies incorporate automatic labeling (Table 4), self-supervised and semi-supervised learning approaches offer promising avenues for reducing labeling requirements while maintaining performance. Active learning strategies, as demonstrated by [18,23], can optimize human involvement in the labeling process.

Cross-modal Learning: The 33% of studies utilizing multimodal data (Table 4) demonstrate the potential of this approach, but advanced fusion techniques that learn joint representations across modalities rather than processing them separately could significantly improve information extraction. Transformer-based architectures [37,43] show particular promise in this domain.

5.3.2. Advanced Learning Approaches

Pre-disaster Focus: With only 3% of studies addressing the pre-disaster phase and only one study [29] focusing on early warning systems, this limited attention represents a significant research opportunity. Early warning systems leveraging predictive analytics could substantially reduce disaster impacts through proactive response planning.

Federated Learning: Distributed learning approaches could enable collaborative model training across organizations without centralizing sensitive data. This approach addresses privacy concerns while improving model robustness through diverse training experiences.

Adaptive Learning Frameworks: With only 33% of studies incorporating real-time learning capabilities (Table 4), solutions that can adapt to evolving crisis situations through continuous model updating, as partially implemented by [34,42], represent a critical advancement for practical deployment.

Foundation Models for Crisis Management: Recent advances in large language and vision models could be leveraged through specialized fine tuning for crisis contexts, potentially addressing the cross-domain adaptation challenge (only 23% of studies in Table 4) while reducing training data requirements.

5.3.3. Integration with Decision Support Systems

Human–AI Collaborative Frameworks: Future systems should prioritize human–AI collaboration over full automation, integrating AI’s processing capabilities with human contextual understanding and ethical judgment.

Multimodal Interactive Visualization: Advanced visualization techniques that effectively communicate complex multimodal information to decision-makers represent a vital area for future work.

Scenario Simulation: The integration of crisis management models with simulation environments could enable the rapid testing and improvement of response strategies before real-world deployment. Recent reviews on Digital Twins and Metaverse-enabled simulation ecosystems report operational gains from virtualized and predictive testing [51], offering valuable insights for the development of crisis-oriented simulation frameworks.

Optimization-Oriented Frameworks: Complementing simulation-driven approaches, future research can benefit from incorporating lean optimization philosophies that emphasize efficiency, waste reduction, and continuous improvement. Mirali et al. [52] highlight how lean strategies, supported by AI and digitalization, enhance process performance through adaptive learning and real-time monitoring. Applying similar principles to crisis management could guide the development of AI-based frameworks that optimize resource allocation, reduce redundancy, and streamline emergency workflows, aligning with the broader goals of sustainable and efficient response systems.

5.3.4. Ethical and Privacy Considerations

As AI and social media become increasingly integrated into crisis management, ensuring the ethical use of personal data is critical. Future research should prioritize privacy-preserving approaches, such as anonymization, differential privacy, and federated learning, to protect user identities. Additionally, transparent data collection policies, informed consent, and clear ethical guidelines are also necessary to maintain public trust and ensure the responsible use of social media information during emergencies.

6. Conclusions

This research systematically examined the integration of social media platforms and AI techniques in crisis management systems. Through our comprehensive review, we selected representative and credible studies to provide a detailed overview of state-of-the-art solutions while highlighting prevalent trends and commonalities across different approaches. Our analysis revealed that crisis management solutions incorporating social media and AI techniques share common aspects, which we categorized into data and learning dimensions. The data aspect encompassed data types, labeling methods, and sources, while the learning aspect included models, domains, phases, types, and objectives.

Our review thoroughly analyzed these essential aspects and their applications in crisis management. We summarized the identified components in a structured taxonomy and discussed proposed solutions accordingly. This dual-aspect analytical taxonomy provides a theoretical foundation for interpreting the relationships between data characteristics and learning strategies and serves as a framework for guiding future research in AI-driven crisis management solutions. Furthermore, we presented key findings, challenges, and directions for future research.

The findings demonstrated the increasing adoption of multimodal data to enhance the comprehensiveness and accuracy of crisis analysis. We observed that hybrid strategies, which combine labeling methods, learning domains, and learning types, significantly improve both the efficiency and effectiveness of the proposed solutions. Our analysis confirmed that leveraging multiple data sources is critical for gaining comprehensive insights into crisis situations.

Based on these findings, future research should prioritize the development of ethical and privacy-preserving frameworks to ensure the responsible use of social media data during crises. Advancing adaptive and federated learning paradigms will further enable real-time, secure, and privacy-aware decision support. Additionally, the adoption of multimodal transformer-based and foundation models presents promising opportunities for enhancing cross-domain generalization and contextual reasoning. Future research should also explore the development of human–AI collaborative visualization and simulation systems to bridge analytics with real-world decision making. Expanding data sources beyond Twitter and linking analytical models with decision-support systems will enhance transparency, trust, and the operational deployment of AI-driven crisis management solutions.

Overall, this review provides a comprehensive conceptual basis and actionable roadmap for advancing next-generation crisis management frameworks that are adaptive, reliable, and ethically grounded, thereby strengthening societal resilience in the face of future disasters.