Source-linked AI summary
Natural Disasters Detection in Social Media and Satellite imagery: a survey
Naina Said, Kashif Ahmad, Michael Regular, Konstantin Pogorelov, Laiq Hassan, Nasir Ahmad, Nicola Conci
TL;DR
Timely disaster-damage information can be difficult to gather, despite its potential value for mitigating losses and supporting recovery. This survey reviews disaster detection and analysis in social-media text and images and satellite imagery, as well as benchmark datasets. It reports patterns in the reviewed methods, including deep-model use and the advantages of model fusion in visual analysis.
Problem
Timely information about disaster damage is difficult to gather, even though disasters can cause major property and human losses.
Source: §1 · p. 1Method
The paper surveys disaster detection and analysis approaches, benchmark datasets, trends, and challenges across social-media text and images and satellite imagery.
Source: §1 · p. 2, §7 · p. 14Results
The survey finds extensive Twitter analysis, growing use of deep architectures for social-media images, better performance from Places-pretrained than ImageNet-pretrained models for disaster events, and CNN improvements over handcrafted satellite features.
Source: §7 · p. 15Takeaways & Limitations
The review identifies benchmark datasets and cross-domain trends and challenges as reference points for future disaster-analysis research.
Source: §1 · p. 2, §7 · p. 16Large-scale benchmarks covering diverse disaster types and modalities remain scarce, and much satellite-image evaluation relies on non-public image sets.
Source: §6.3 · p. 13, §7 · p. 16
Abstract
from arXiv · showhide
The analysis of natural disaster-related multimedia content got great attention in recent years. Being one of the most important sources of information, social media have been crawled over the years to collect and analyze disaster-related multimedia content. Satellite imagery has also been widely explored for disasters analysis. In this paper, we survey the existing literature on disaster detection and analysis of the retrieved information from social media and satellites. Literature on disaster detection and analysis of related multimedia content on the basis of the nature of the content can be categorized into three groups, namely (i) disaster detection in text; (ii) analysis of disaster-related visual content from social media; and (iii) disaster detection in satellite imagery. We extensively review different approaches proposed in these three domains. Furthermore, we also review benchmarking datasets available for the evaluation of disaster detection frameworks. Moreover, we provide a detailed discussion on the insights obtained from the literature review, and identify future trends and challenges, which will provide an important starting point for the researchers in the field.
1. Introduction
Natural disasters can cause major losses, while timely damage information is difficult to gather and may be unavailable from news agencies. The survey reviews disaster analysis across social media and satellite imagery, including approaches, benchmark datasets, trends, and challenges.
1. Introduction
- Natural disasters can cause substantial property and human losses, and timely damage information may help mitigate losses and speed recovery.News agencies have sometimes been unable to report disasters promptly or at all because they lack reporters in the affected area.Source: §1 · p. 1
- The survey reviews disaster detection, retrieval, summarization, and analysis across social media and satellite imagery, alongside benchmark datasets, trends, and future directions.Its scope includes text, visual content from social media, and satellite imagery.Source: §1 · p. 2
2. Open/Key Research Challenges
The section identifies challenges in using social media and remote-sensed data for disaster analysis, including information quality, dataset availability, and satellite coverage and image quality.
2. Open/Key Research Challenges
- Social media content may be relevant but unauthentic, and existing work often lacks public, sufficiently large datasets covering varied disaster events.For Twitter, many studies use self-collected datasets; the survey calls for additional measures to check the authenticity of shared information.Source: §2 · p. 2
- Satellite imagery has low temporal frequency and provides a bird’s-eye view, while clouds, lighting, vegetation, and processing errors can obscure the ground.The New South Wales flood image makes the boundary concrete: flooding is visible, but its environmental and societal impact is difficult to determine; social-media information may enrich the imagery.Source: §2 · p. 2, Figure 1 · p. 2
- Large-scale annotated datasets remain scarce across Twitter text, social-media images, and satellite imagery, limiting training and evaluation.The section identifies content collection, handling, and analysis across platforms and modalities as related challenges.Source: §2 · p. 2
3. Disaster Detection in Social Media
The survey organizes social-media disaster analysis around Twitter text and user-shared visual content, then reviews benchmark initiatives for these modalities. Across the approaches, it describes varied text-processing methods and the use of visual features with metadata.
3.1. Disaster Detection in Twitter Text
- Twitter-text approaches span data cleaning and analysis, named-entity and cluster-based information extraction, Bayesian classification, and ontology-linked text and image analysis.The reviewed systems also include ensemble classifiers for heterogeneous emergencies and semi-supervised detection of unknown events.Source: §3.1 · pp. 3–4
- Other Twitter methods use spatio-temporal clustering, tweet-frequency signals, crisis mapping, and feature comparisons for eyewitness-report classification.Tweedr combines classification, clustering, and information extraction; its evaluation found Logistic Regression most reliable on several classification metrics, while the feature study compares five feature types across two data-split scenarios.Source: §3.1 · p. 4
- AIRD combines human intelligence and machine learning for rapid classification of disaster messages, with categories defined through crowdsourcing.The system continuously retrieves information and was tested on a 2013 Pakistan earthquake; a related top-down approach was designed around humanitarian agencies’ needs.Source: §3.1 · pp. 4–5
3.2. Disaster Detection in images from Social Media
- Social-media image analysis combines visual features with metadata such as tags, user information, time, and geolocation; reviewed work also uses human annotation and fine-tuned CNNs.The surveyed image methods include comparisons of global and deep features, classification algorithms, and late fusion.Source: §3.2 · p. 5
3.2.1 Benchmark competitions on disaster analysis in social media
- MediaEval-2017 evaluated flood-image retrieval using visual-only, metadata-only, and combined inputs, with submissions using deep features and fusion strategies.One described approach also translates users’ tags into English and combines handcrafted visual features with machine-learning classifiers.Source: §3.2.1 · pp. 5, 7
- MediaEval-2018 asked systems to identify images with road-passability evidence, then classify roads as passable or non-passable.Submissions commonly used deep models and metadata; one sequential approach first detects evidence and then classifies passability, while another account reports GAN experiments abandoned because of the small dataset.Source: §3.2.1 · p. 7
4. Disaster Detection in Satellite Imagery
The survey reviews satellite imagery for disaster monitoring and analysis, including CNN-based detection and damage assessment. It also covers flood-detection and road-passability benchmark tasks, while noting the limited availability of comparable satellite datasets elsewhere in the paper.
4. Disaster Detection in Satellite Imagery
- Satellite imagery supports disaster monitoring across management phases, and reviewed methods use CNNs, deep features, and UAV imagery for detection and damage assessment.Applications discussed include mapping hazards, assessing building damage, and identifying affected areas across several disaster types.Source: §4 · p. 9
4.0.1 Benchmark competitions on Flood detection in Satellite Imagery
- MediaEval-2017 framed satellite flood detection as flooded-versus-non-flooded region segmentation using Planet four-band imagery, while approaches included modified VGGNet, dilated convolutions, GANs, and spectral indices.The described GAN extension used RGB and RGB+IR inputs; its authors reported that the IR component may lead to false positives.Source: §4.0.1 · pp. 9–10
- MediaEval-2018 asked participants to identify passable roads in flood-affected satellite imagery, with three methods presented.The reviewed methods used RGB histograms with SVMs, fine-tuned ResNet-50, or transfer learning with Inception V-3; GAN experiments were abandoned because the dataset was small.Source: §4.0.1 · p. 10
5. Current Trends in Disasters Analysis
A growing trend is to combine social-media content with satellite imagery: satellite views provide a broad picture, while social media adds localized detail. Proposed systems collect, analyze, and link both sources for disaster exploration.
5. Current Trends in Disasters Analysis
- Joint-use tasks aim to enrich satellite imagery with localized social-media text, images, and videos alongside a broader view of the disaster.The trend is emerging through MediaEval’s Multimedia Satellite task and an ACM Multimedia 2016 challenge.Source: §5 · p. 10
- A contextual-enrichment system collects Landsat imagery and Twitter multimedia, filters irrelevant content, maps it using geolocation, and presents it through an exploratory interface.Geolocation comes from place names in tweets or available GPS annotations.Source: §5 · p. 10
- JORD links disaster information from EM-DAT, multilingual social-media content, and satellite imagery, and includes a GAN-based flood-detection test case.Its interface presents a joint view of disaster-related social-media multimedia and satellite imagery.Source: §5 · pp. 10, 12
6. Datasets
The section distinguishes general disaster databases from task-specific benchmarks for disaster detection. It surveys Twitter, social-image, and satellite-image datasets, noting that satellite evaluation remains constrained by limited shared data.
6. Datasets
- General disaster databases cover loss and damage, hazard catalogues, socio-economic indicators, and exposure, while this survey focuses on task-specific detection benchmarks.The general databases contain historical information, and their coverage and collection methods vary.Source: §6 · p. 12
6.1. Twitter Datasets
- Twitter benchmarks span crisis relevance, informativeness, humanitarian categories, and event metadata, including CrisisLexT26’s 250,000 tweets across 26 events.Other collections include Spanish-language Chile earthquake tweets, Italian disaster tweets, and 6,556,328 geo-tagged Sandy tweets without labels. ClimateCovE350 covers about 350 climate-related events with diverse tags, while a seven-disaster tweet-and-image dataset supports informativeness, humanitarian-category, and damage-severity tasks.Source: §6.1 · pp. 12–13
6.2. Images Datasets
- Social-image benchmarks cover flood detection and road passability, with DIRSM providing 5,280 development and 1,320 test images.The MediaEval 2018 dataset contains 11,070 images from Hurricanes Harvey, Irma, and Maria, supporting image-level passability evidence and passable-versus-non-passable road classification.Source: §6.2 · p. 13
6.3. Satellite Imagery Datasets
- Satellite-image benchmarks remain limited; most reviewed works use image sets unavailable for comparison, although flood and road-passability datasets are described.FDSI includes Planet four-band flood patches, with test subsets from development locations and unseen locations; the MediaEval 2018 road dataset contains 1,663 DigitalGlobe patches.Source: §6.3 · p. 13
7. Summary and future directions
The survey reviews disaster detection and analysis across social-media text and imagery and satellite imagery, alongside benchmark datasets and research challenges. It reports model and benchmark patterns, while identifying video analysis and richer information integration as future directions.
7. Summary and future directions
- The survey covers disaster detection, retrieval, summarization, and analysis across Twitter text, social-media visual content, and satellite imagery, as well as evaluation datasets.It also discusses trends and challenges across these subdomains.Source: §7 · pp. 13–15
- Twitter disaster detection has been extensively studied, while social-media image and satellite methods increasingly use CNNs and other deep architectures.Places-pretrained models tend to outperform ImageNet-pretrained models for disaster imagery, CNN fusion outperforms individual models, and metadata supports visual information.Source: §7 · p. 15
- 95% average accuracy on DIRSM and 65% on FCSM contrast with satellite mean F-scores of 65% on FDSI-2017 and 62% on FDSI-2018.The survey describes FCSM as significantly more complex than DIRSM and notes that self-collected Twitter datasets make quantitative method comparisons difficult.Source: §7 · pp. 15–16
- Most existing visual-content analysis relies on images; video analysis is a more challenging future direction that is expected to provide more detailed information.The survey says videos are expected to improve users’ experience.Source: §7 · p. 16
- OSM and other GIS data may improve AI performance alongside satellite imagery, while disaster statistics may inform impact, recovery costs, and future risk analysis.The survey proposes integrating and automatically analyzing these information sources as future research.Source: §7 · p. 16