AI and Generative AI Transforming Disaster

Management: A Survey of Damage Assessment and

Response Techniques
Aman Raj Lakshit Arora Sanjay Surendranath Girija
amanraj@[Link] lakshit@[Link] sanjaysg@[Link]

Ankit Shetgaonkar Dipen Pradhan Shashank Kapoor

ankiit@[Link] dipenp@[Link] shashankkapoor@[Link]

Google

Abstract—Natural disasters, including earthquakes, wildfires methods like deep learning models offer a revolutionary
and cyclones, bear a huge risk on human lives as well as solution for disaster damage assessment. Models like
infrastructure assets. An effective response to disaster depends on Variational Autoencoders (VAEs) [2] and Generative
the ability to rapidly and efficiently assess the intensity of damage. Adversarial Networks (GANs) [3], can synthesize
Artificial Intelligence (AI) and Generative Artificial Intelligence representations of disaster-affected areas, enabling rapid
(GenAI) presents a breakthrough solution, capable of combining analysis and simulation. Generative AI models, particularly
knowledge from multiple types and sources of data, simulating task-finetuned ones, also are very well-suited to perform image
realistic scenarios of disaster, and identifying emerging trends at recognition and classification on multi-modal data from
a speed previously unimaginable. In this paper, we present a
affected regions, greatly improving the speed of assessment [4].
comprehensive review on the prospects of AI and GenAI in
damage assessment for various natural disasters, highlighting In this paper, we conduct a thorough evaluation of AI and
both its strengths and limitations. We talk about its application to GenAI’s capabilities in assessing damage from disasters like
multimodal data such as text, image, video, and audio, and also earthquakes, wildfires and cyclones, with a focus on their
cover major issues of data privacy, security, and ethical use of the strengths and limitations. We further discuss its capacity to
technology during crises. The paper also recognizes the threat of integrate disparate data across various modalities, including
Generative AI misuse, in the form of dissemination of satellite imagery, social media posts, sensor networks, and
misinformation and for adversarial attacks. Finally, we outline emergency reports, to generate real-time situational awareness.
avenues of future research, emphasizing the need for secure,
Towards the end we provide an overview, highlighting the most
reliable, and ethical Generative AI systems for disaster
management in general. We believe that this work represents the
important challenges and propose potential directions for future
first comprehensive survey of Gen-AI techniques being used in the research and development in this critical area.
field of Disaster Assessment and Response. The organization of this paper is as follows. In section II,
we discuss the various domain-specific literatures on the use of
Keywords— Generative Artificial Intelligence, Disaster GenAI in disaster assessment. In section III, we discuss the data
Response, Damage Assessment, Earthquake, Wildfire, Cyclone,
modalities across which GenAI techniques can be utilized. In
Deep Learning, Data Privacy, Adversarial Attacks, Misinformation.
section IV, we present how Gen-AI techniques can also be used
I. INTRODUCTION for combating misinformation and in an adversarial setting.
Section V brings the discussion on data privacy, security
The combination of growing urban areas and climate considerations and deployment challenges and section VI is
change is leading to more natural disasters and resulting in about our opinion on the future for Gen-AI techniques and
devastating human and economic losses. Precise and timely conclusions.
damage assessment is one of the most significant factors in an
effective disaster response. This calls for innovation in disaster II. DOMAIN-SPECIFIC LITERATURE REVIEW
management and assessment techniques. It is required for
This section provides a review of recent publications on
efficient resource allocation, rescue planning, and
damage assessment techniques for earthquakes, wildfires, and
reconstruction planning. Without clear and comprehensive
cyclones with a focus on the use of AI/GenAI.
damage assessment, the response is delayed and ineffective.
Traditionally, damage assessment has relied on manual A. Earthquakes
methods like ground survey, satellite data interpretation and Earthquakes present critical challenges to damage
physical inspection which are valuable, but nevertheless slow assessment due to the widespread and often unpredictable
and labor-intensive [1]. They are also hindered by extent of damage they inflict. Traditional earthquake damage
communication failures and other logistic issues. AI/GenAI assessment has relied heavily on visual inspection of structures

© 2025 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current
or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective
works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.
and infrastructure by professional engineers and emergency representation [14]. This approach helps to leverage its unique
teams [1]. These procedures, while vital in determining probabilistic nature for a more expressive latent representation,
structural soundness and identifying imminent safety hazards, aiming to better model the complex patterns in wildfire data
are time-consuming and often risky. Remote sensing compared to purely classical approaches. Using the pre-fire
technologies, satellite high-resolution images and information image as condition, the VQ-VAE decoder generates a
from unmanned aerial vehicles (UAV, or drones) are reconstruction of the post fire image. The difference between
increasingly being used along with machine learning to this generated image and the actual post-fire image is processed
automatically identify and distinguish damaged structures and to produce a segmentation map of the burned area and is then
infrastructure parts [5, 6]. evaluated using metrics like Intersection over Union (IoU) and
F1-Score.
Lately there has been significant research on using GenAI
techniques for the damage assessment. Mousavi, S. Mostafa, et Park et al. (2020) introduced a deep learning wildfire
al. (2024) explored the use of multimodal social media data to detection system using remote camera images with the
estimate earthquake shaking intensity [7]. Their work leveraged objective of solving data imbalance issues typically
Natural Language Processing (NLP) techniques, such as encountered in disaster datasets [15]. They used CycleGAN
sentiment analysis using transformer-based models, on text [16], a particular GAN for unpaired image-to-image translation,
data combined with image classification of user-submitted mostly to generate synthetic wildfire images from accessible
photos to assess damage levels in real-time. They addressed the non-wildfire forest photographs. Such data augmentation based
challenge of noisy social media data by employing filtering on CycleGAN increased training data numbers and variety
techniques based on source credibility and content verification. widely without employing paired examples to overcome the
Estêvão (2024) evaluated multimodal Generative AI models constraint of limited wildfire imagery. Then, a DenseNet-
(like OpenAI's GPT-4o and GPT-4o mini, and Google's Gemini driven classifier was trained on the data-augmented set and
1.5 Pro and Gemini 1.5 Flash) for classifying building damage performed better (98.27% accuracy and 98.16 F1-score) than
post-earthquake from images using the EMS-98 scale [59]. those with no GAN augmentation or by utilizing traditional pre-
Their study found varying accuracy (28.6%-75.0%) across trained models like VGG-16 [60] and ResNet-50 [61]. The
masonry and reinforced concrete structures, best accuracy research confirmed that GAN-generated synthetic data were
shown by GPT-4o. The work showed GenAI has the potential able to efficiently enhance wildfire detection accuracy.
for automated assessment but indicated that current
performance isn't optimal, suggesting significant improvements GenAI's capacity for synthetic training data generation
addresses one of the most important issues of wildfire
could be achieved via fine-tuning or Retrieval-Augmented
estimation: the unavailability of labeled data. Techniques like
Generation (RAG) [8].
CycleGANs can be used to transform pre-fire to post-fire
Bhadauria (2024) conducted a comprehensive review scenes, synthesizing realistic training examples for damage
exploring the blend of AI and Machine Learning (ML) tools segmentation models.
within earthquake engineering, focusing on damage assessment
C. Cyclones
and retrofitting strategies. Their work surveyed recent
applications of ML, Pattern Recognition (PR), and Deep Cyclone damage estimation demands prompt and correct
Learning (DL), emphasizing how these techniques surpass quantification of storm impacts along coastlines, structures, and
facilities to guide effective disaster management [17].
traditional model limitations in areas like risk prediction,
Traditional methods for this were based on the combination of
retrofitting design, and structural optimization. The review satellite imagery, aerial surveys, and ground observation data
identified challenges, limitations, and future research avenues [18]. Modern AI offers advanced alternatives for automating
for advancing seismic resilience engineering using AI/ML [9]. and streamlining such tasks.
B. Wildfires Arachchige & Pradhan (2025) developed a machine
Wildfire damage assessment is determining burn severity learning-based risk estimation of hurricane-induced flood
and comprehending the total impact on infrastructure, damage in Florida from granular insurance data enriched with
ecosystems, and communities. Classic approaches in the past remote sensing variables [19]. They compared several
employed remote sensing imagery from optical, thermal, and algorithms (Random Forest [62], XGBoost [63], GBM, Neural
radar sensors [10, 11, 12]. Generative AI is introducing new Networks) and selected a stacked ensemble model, which
paradigms for analyzing and interpreting such information. combines predictions from these base learners by a meta-
learner, because of its enhanced performance. This ensemble
Asanjan et al. (IGARSS 2024) proposed a technique for model predicted flood damage at the Zip Code Tabulation Area
wildfire segmentation from remotely sensed data using a
(ZCTA) level with a mean absolute error of 11.3%.
quantum-compatible, modified Vector-Quantized Variational- Explainability tools (Partial Dependence Plot [64]) were used
Autoencoder (VQ-VAE) [13]. Their methodology enhanced
for model interpretation, identifying key predictors like
probabilistic U-Net [57] models by incorporating a Restricted property value and showing geographic hotspots based on
Boltzmann Machine (RBM) [58], which is optimized with
factors like Gulf warmth and urbanization.
Parallel Tempering in PySA, into the prior encoder step for
improving latent space description and segmentation
 Figure 1: Taxonomy of Natural Disaster surveys and methods in this survey

Calton & Wei (2022) utilized transfer learning on advanced reactions in affected populations to be better identified.
Artificial Neural Networks (ResNet, MobileNet [65], Furthermore, aspect-based sentiment analysis is used to classify
EfficientNet [66]) to assess hurricane/cyclone building damage fine-grained sentiment on specific entities, like infrastructure or
from images [20]. Their work involved binary classification government response, further improving resource targeting
(flood/non-flood) and multi-class object detection (damaged accuracy. Tarasconi, Francesco, et al(2017) published research
roof, wall, flood, structural damage). MobileNet demonstrated suggesting that GenAI also supports relationship extraction and
superior performance, achieving 87% accuracy (0.88 F1-score) event correlation resolution beyond mere Named Entity
in classification and high confidence (up to 97.58%) in Recognition [23]. Models can identify complex relationships
detecting specific damage types, showing the efficiency of among damaged infrastructure, affected populations, and relief
transfer learning for rapid post-cyclone assessment. responses, providing richer situational context than traditional
rule-based systems.
In addition, GenAI can assess infrastructure resilience by
generating hypothetical "damaged" copies of critical structures Models are able to leverage unstructured information and
[21]. For instance, it is possible to train a VAE to encode the contextual knowledge in order to deduce top damage indicators,
structural response of bridges subjected to various simulated estimate affected populations, and indicate emergent needs,
storm loads. The encoder maps the pre-cyclone bridge state into making reports tailored to the specific needs of the individual
a latent space, and the decoder generates potential post-cyclone stakeholder, either policymakers or first responders. Smart
states based on estimated damage parameters. This enables Conversation Agents (Chatbots) no longer stop at responding
predictive flagging of vulnerable infrastructure and informing to standard questions. GenAI makes adaptive dialog
targeted inspections following the storm [21]. management and contextual responses possible. Through a mix
of reinforcement learning and human feedback (RLHF),
III. DATA MODALITIES AND GENERATIVE AI chatbots can be optimized for more empathetic, contextually
APPLICATIONS knowledgeable interactions, creating greater user trust and
This section describes GenAI’s applications for damage encouraging active engagement with usable resources [24].
assessment and disaster response across various data Managing the inherent subtleties of text data requires
modalities. sophisticated methods. Text data from social media can be
A. Text noisy and unreliable, and requires filtering based on the source
for reliability [28]. Another possible issue is disinformation and
Text data offers dense, critical information sourced from
combating it requires GenAI to be able to recognize and mark-
various sources like social media posts, news services, off manipulated stories, integrate fact-checking modules, and
emergency notification services, call center reports, and actively recommend reliable sources of information [29]. When
government communication released following a disaster. Text dealing with multilingual sources of data, accurate machine
data analysis is of paramount importance due to its natural
translation and cross-lingual information searching are
language characteristics, thus enabling it to convey
required.
information, express emotions and reflect human
communication patterns. Generative AI techniques are B. Image Data
revolutionizing the use of this text data in disaster relief. Image data from satellites, aerial drones, roadside cameras,
Instead of relying on conventional sentiment analysis, there and social media sites is very important for visual and
are now newer approaches that employ transformer-based quantitative disaster damage assessments [30].
models such as BERT, RoBERTa, and XLM-RoBERTa [22]. Automated damage detection and segmentation now
Training or finetuning the models against multilingual, employ advanced Convolutional Neural Networks (CNNs) like
disaster-oriented corpora allows public opinion related to Mask R-CNN and U-Net, implementing semantic segmentation
developing events, resource constraints, and emotional
for pixel-level classification of specific damage types [31]. To GenAI is capable of aiding 3D reconstruction of damaged areas
improve robustness against variations in illumination, weather, from video. Techniques like Structure from Motion (SfM) and
and image quality, transfer learning is used by pretraining over Multi-View Stereo (MVS), complemented with deep learning
large datasets like ImageNet and then fine-tuning with disaster- for enhanced feature matching and depth estimation, can create
related datasets like xBD. Furthermore, techniques like high-resolution 3D models [37]. These models can provide
attention mechanisms and spatial pyramid pooling [50] extract structural strength estimation, rubble volume estimation, and
multi-scale contextual information [32], facilitating precise can be combined with virtual reality packages for remote
damage extraction even in partially obscured scenes. inspection.
When image data is incomplete, image inpainting Efficient use of video data has some open challenges. The
techniques are invaluable. Generative models, particularly massive volumes of data produced by video streams demand
those based on contextual autoencoders and Generative powerful computing and efficient processing techniques. The
Adversarial Networks (GANs), enable realistic restoration of removal of motion blurring effects, occlusion, illumination
corrupted or obscured regions [33]. The generator network variation, and view variation to maintain analysis quality is
learns feature representations from undamaged areas to crucial. Finally, the privacy of the people captured in the video
synthesize realistic in-paintings, while the discriminator data must be maintained by anonymization techniques and
enforces realism, guided by adversarial loss functions. Image adherence to ethical standards [38]. Research has to overcome
super-resolution uses architectures like Enhanced Deep Super- these challenges to utilize the full potential of video data in
Resolution Network (EDSR) [67] and Residual Channel disaster response.
Attention Networks (RCAN) [68] to enhance low-resolution
images. This uses residual blocks and channel attention to D. Audio Data
reconstruct high-frequency details and sharpen damage maps Audio data like emergency calls, social media audio and
[34]. recordings, and drone audio recordings hold crucial
information on disaster damage like location, affected
GANs and VAEs can produce new images simulating population size and impact severity. GenAI opens new
various levels of structural damages and weather conditions possibilities for tapping into this data to enhance disaster
addressing the scarcity of labeled training data. They response efforts.
significantly improve the strength and generalization
capabilities of AI algorithms to effectively recognize damage Distress call analysis employs GenAI coupled with signal
patterns in new scenarios. The key challenges in this area processing to identify and prioritize the most critical emergency
include accommodating variations in image resolution and situations [39]. Techniques like Mel-Frequency Cepstral
quality, and attaining high levels of precision for GenAI-driven Coefficients (MFCCs) [70] extract vocal features, but GenAI's
damage assessments. strength lies in using models like transformers to understand
call context beyond keywords. GenAI can be used to identify
C. Video Data panic or distress and generate a more nuanced "urgency score",
Video data creates a stream of information that is critical for factoring in emotional tone and acoustic environment.
facilitating rapid damage assessment. Video maintains event
Sound event detection also benefits from GenAI. While
chronology, so it is possible to place incidents in a more
Hidden Markov Models (HMMs) were traditional, deep
nuanced context to estimate the intensity and progression of
damage. Drone videos, security camera footage, news learning models now learn complex acoustic patterns to detect
coverage, and live streaming offer various insights and GenAI events [40]. Another key advance through GenAI is the
techniques can help uncover actionable insights from this generation of synthetic audio data of disaster sounds (e.g.,
building collapse). This augments training and boosts detection
dynamic information.
accuracy, especially for rare events. Correlating these events
AI significantly enhances real-time damage assessment with other modalities can also increase efficacy [41].
capabilities through enabling classification and identification of
various forms of damage as and when they occur in video The key challenges with audio data are removal of
streams [35]. Spatiotemporal feature extraction with 3D background noise, accents, and maintaining privacy. While
Convolutional Neural Networks (3D-CNNs) and temporal conventional signal processing tools help, GenAI can
dependency with Recurrent Neural Networks (RNNs) like potentially generate sanitized audio samples or modify voices
LSTMs can track the general progression of a disaster event in to restrict accent bias and ensure data anonymization. Future
solutions could rely on GenAI to not just analyze audio data but
real time. For instance, object detection software based on
standalone frames can identify collapses of buildings or generate insights and improve data quality.
flooding water to issue evacuation warnings and guide rescue IV. MISINFORMATION AND ADVERSARIAL ATTACKS
forces [36]. The feedback loop in real-time increases situation
awareness so that resources are moved faster. The ability to generate very realistic but entirely fake
images and videos is a genuine threat to trustworthy GenAI-
GenAI also makes it possible to identify events like based assessment. Manipulated media ingested as authentic
collapses of buildings or explosions in video and trigger alerts. data can lead to flawed analysis and can disrupt the pipeline,
Object detection software like YOLO [69], enriched with leading to resource misallocation and even lethal consequences.
temporal reasoning, can categorize such incidents swiftly,
enabling quick damage evaluation and response. Moreover, One of the biggest issues is employing images produced by
GANs and diffusion models. GANs can produce manipulated
images by modifying textures, erasing or adding objects, or Explainable AI (XAI) approaches can provide insight into
changing environmental factors. Diffusion models that learn to model reasoning [49]. SHAP (SHapley Additive exPlanations)
reverse a gradual process of noise, introduce even greater [73] importance quantifies the contribution of every input
realism and control over image synthesis, thus producing bogus feature to a prediction, while LIME (Local Interpretable Model-
evidence of destruction. agnostic Explanations) [74] approximates local model behavior
with simpler, interpretable models that allow model output
CNNs can help with this by analyzing minute
understanding via local explanations to help establish trust in
inconsistencies that are oblivious to the naked eye [42]. They
stakeholders.
look at lighting, shading, texture, and even frequency-domain
artifacts that are introduced in GAN or diffusion model Possible biases of AI models should be addressed in order
processing. More recent architectures such as Vision to prevent unfair outcomes. Right training data choice and pre-
Transformers (ViTs) [71] and Swin-Transformers [72] also processing are necessary in order to present diverse populations
offer improved capabilities at capturing global context in and regulate possible biases [54].
images that assist in detecting subtle inconsistencies.
Techniques like inspection of noise patterns can detect the Deployment concerns add complexity in terms of real-time
computation, which is highly important within disaster zones,
characteristic "fingerprint" produced by different generative
models. In order to specifically counter deep fakes, techniques and calls for compression and optimization technology on
evaluating facial micro-expressions and lip-syncing models so they can be run effectively on all kinds of hardware.
inconsistency are essential [43, 44]. Development of "universal VI. FUTURE PROSPECTS AND CONCLUSION
detectors" which can identify manipulations in any generative
source is a key research goal. AI and GenAI ability to combine various data sources and
types in new ways can help transform damage assessment.
Effective mitigation should also involve prevention along Unleashing this potential will require focused research work in
with detection. Digital watermarking, for example, can embed priority areas.
invisible yet verifiable signatures into the original image for
authentication [45, 55]. Cryptography hash functions [56] along Advancing GenAI Model Capabilities: This requires focused
with perceptual image hashing provide an additional layer of research on improving GenAI models for precision,
content verification. Designing GenAI assessment systems with generalizability, and resilience. It needs exploring the
internal attention mechanisms to evaluate input reliability is application of powerful deep learning architectures, like
also essential [46]. Furthermore, strong data governance transformers with advanced attention and graph neural
imparts intrinsic trust upon GenAI models by providing networks (GNNs) [75] for spatial relationships modeling.
tangible assurances about the underlying training data. This Research also needs to be performed to develop methods that
involves technical procedures for data lineage tracking, can properly handle noisy and missing data, and to make the
automated quality validation, privacy enforcement minimizing models more robust against adversarial attacks and false
re-identification, secure data lifecycle management, and bias information using adversarial training and advanced data
auditing before training. augmentation. Unified datasets and benchmarks are also needed
to facilitate progress.
The assurance of integrity, security, and ethical handling of
core data provides the necessary verifiable foundation for trust Integration with existing technologies: Creating integrated
in the model's predictions during high-stakes events. disaster response platforms also need integration of GenAI with
other existing systems. This can be achieved by integration of
V. PRIVACY, SECURITY AND DEPLOYMENT CHALLENGES GenAI with i) Geographic Information Systems (GIS) [76] for
spatial analysis and visualization, ii) diverse sensor networks
It is important to have strong data security and privacy
safeguards when implementing GenAI for disaster response. for real-time data ingestion, iii) with communication systems
Overlooking them can weaken public trust, risk sensitive for disseminating actionable intelligence. Research should
information, and cause unintended harm. target efficient and scalable interfaces between these systems.

Protecting people's privacy by having strong anonymization Explainable AI (XAI): Developing systems that explains the
of data is critical. K-anonymity and l-diversity can be used to behavior and results produced by a model to improve
transparency and trust in AI-driven insights [52]. This includes
keep sensitive attributes from being easily traced back to
individuals in the data set [47]. Differential Privacy (DP) also visualizing model activations, identifying influential features
offers stronger, provable privacy guarantees [51]. The Laplace (e.g., via SHAP, LIME), and generating human-interpretable
Mechanism, a core component of DP, adds calibrated noise to explanations, potentially using standardized formats like model
data in a bid to hide individual contributions [48]. DP allows cards.
for fine-grain privacy budget tuning for privacy versus model Addressing Ethical and Societal Issues: Ethical
utility trade-offs in specific disaster assessment applications. considerations must guide GenAI design and deployment in
Secure data storage and transmission demand robust disaster response [53]. This requires ethical frameworks for
cryptographic controls. Penetration testing and regular security data handling, addressing privacy (e.g., via federated learning,
audits identify and correct GenAI system architecture DP) and security, mitigating biases through fairness-aware
learning and representative data, and ensuring equitable
weaknesses. Adopting zero trust architectures across all the
resource distribution. Interdisciplinary collaboration among
data is a main idea to keep in mind for all such safety types.
computer scientists, ethicists, and responders is essential to 11, no. 22, Jan. 2019, p. 2638. [Link],
[Link]
attain this goal. [12] Barmpoutis, Panagiotis, et al. “A Review on Early Forest Fire Detection
Systems Using Optical Remote Sensing.” Sensors, vol. 20, no. 22, Jan.
Developing Novel GenAI Applications: GenAI enables 2020, p. 6442. [Link], [Link]
advanced applications beyond basic assessment. Future work [13] Oord, Aaron van den, et al. Neural Discrete Representation Learning.
arXiv, 2017. [Link] (Datacite),
includes personalized disaster response plans using privacy- [Link]
preserving methods [51], predictive resource allocation models [14] Asanjan, Ata Akbari, et al. Wildfire Segmentation From Remotely Sensed
Data Using Quantum-Compatible Conditional Vector Quantized-
forecasting needs, realistic evacuation simulations to optimize Variational Autoencoders. 2024. [Link],
strategies, and automated damage report generation using [Link]
Natural Language Generation (NLG) [77] to synthesize [15] Park, Minsoo, et al. “Wildfire-Detection Method Using DenseNet and
CycleGAN Data Augmentation-Based Remote Camera Imagery.”
multimodal data into coherent summaries. Remote Sensing, vol. 12, no. 22, Nov. 2020, p. 3715. [Link] (Crossref),
[Link]
In conclusion, this survey shows that AI and GenAI provide [16] Zhu, Jun-Yan, et al. “Unpaired Image-to-Image Translation Using Cycle-
Consistent Adversarial Networks.” 2017 IEEE International Conference
a powerful instrument for transforming disaster damage on Computer Vision (ICCV), 2017, pp. 2242–51. IEEE Xplore,
assessment. Realizing its full potential depends on overcoming [Link]
[17] Radhika, Sudha, et al. “Cyclone Damage Detection on Building
significant challenges in model robustness, data quality, ethics, Structures from Pre- and Post-Satellite Images Using Wavelet Based
and integration. There are concerns around ethical deployment, Pattern Recognition.” Journal of Wind Engineering and Industrial
Aerodynamics, vol. 136, Jan. 2015, pp. 23–33. ScienceDirect,
measures against disinformation, anonymization and [Link]
adversarial attacks. The successful integration of strong ethical [18] Rahnemoonfar, Maryam, et al. “FloodNet: A High Resolution Aerial
frameworks and reliable technical safety measures will Imagery Dataset for Post Flood Scene Understanding.” IEEE Access, vol.
9, 2021, pp. 89644–54. IEEE Xplore,
ultimately determine the future effectiveness and societal [Link]
acceptance of GenAI in disaster assessment and response. [19] Arachchige, Sameera Maha, and Biswajeet Pradhan. “AI Meets the Eye
of the Storm: Machine Learning-Driven Insights for Hurricane Damage
Risk Assessment in Florida.” Earth Systems and Environment, Feb. 2025.
We believe that this work represents the first comprehensive Springer Link, [Link]
survey of GenAI techniques used for disaster assessment and [20] Calton, Landon, and Zhangping Wei. “Using Artificial Neural Network
Models to Assess Hurricane Damage through Transfer Learning.”
response. In particular, we provide a deep overview of how this Applied Sciences, vol. 12, no. 3, Jan. 2022, p. 1466. [Link] (Crossref),
emerging field of technology can be used for damage [Link]
assessment for earthquakes, wildfires and cyclones. The [21] Haggag, May, et al. “Infrastructure Performance Prediction under
Climate-Induced Disasters Using Data Analytics.” International Journal
research conducted in this paper demonstrates the potential for of Disaster Risk Reduction, vol. 56, Apr. 2021, p. 102121. [Link]
(Crossref), [Link]
the AI and GenAI in disaster response, and also the need for
[22] Qadir, Junaid, et al. “Crisis Analytics: Big Data-Driven Crisis Response.”
further research and investment in the domain to realize this Journal of International Humanitarian Action, vol. 1, no. 1, Aug. 2016,
potential. p. 12. Springer Link, [Link]
[23] Li, Tao, et al. “Data-Driven Techniques in Disaster Information
Management.” ACM Computing Surveys, vol. 50, no. 1, Jan. 2018, pp. 1–
REFERENCES 45. [Link] (Crossref), [Link]
[1] Ozisik, D., and N. Kerle. "Post-earthquake damage assessment using [24] Kaufmann, Timo, et al. A Survey of Reinforcement Learning from
satellite and airborne data in the case of the 1999 Kocaeli earthquake, Human Feedback. arXiv, 2023. [Link] (Datacite),
Turkey." Proc. of the XXth ISPRS congress: Geo-imagery bridging [Link]
continents. 2004. [25] Flah, Majdi, et al. “Machine Learning Algorithms in Civil Structural
[2] Kingma, Diederik P., and Max Welling. An Introduction to Variational Health Monitoring: A Systematic Review.” Archives of Computational
Autoencoders. 2019. [Link] (Datacite), Methods in Engineering, vol. 28, no. 4, June 2021, pp. 2621–43. Springer
[Link] Link, [Link]
[3] Goodfellow, Ian, et al. “Generative Adversarial Networks.” [26] Morton, Douglas C., et al. “Mapping Canopy Damage from Understory
Communications of the ACM, vol. 63, no. 11, Oct. 2020, pp. 139–44. Fires in Amazon Forests Using Annual Time Series of Landsat and
[Link] (Crossref), [Link] MODIS Data.” Remote Sensing of Environment, vol. 115, no. 7, July
[4] Ma, Zhengjing, et al. “Generative Deep Learning for Data Generation in 2011, pp. 1706–20. ScienceDirect,
Natural Hazard Analysis: Motivations, Advances, Challenges, and [Link]
Opportunities.” Artificial Intelligence Review, vol. 57, no. 6, May 2024, [27] Maeda, Eduardo Eiji, et al. “Predicting Forest Fire in the Brazilian
p. 160. Springer Link, [Link] Amazon Using MODIS Imagery and Artificial Neural Networks.”
International Journal of Applied Earth Observation and Geoinformation,
[5] Mangalathu, Sujith, et al. “Classifying Earthquake Damage to Buildings vol. 11, no. 4, Aug. 2009, pp. 265–72. ScienceDirect,
Using Machine Learning.” Earthquake Spectra, vol. 36, no. 1, Feb. 2020, [Link]
pp. 183–208. [Link] (Crossref),
[Link] [28] Havas, Clemens, et al. “E2mC: Improving Emergency Management
Service Practice through Social Media and Crowdsourcing Analysis in
[6] Kerle, Norman, et al. “UAV-Based Structural Damage Mapping: A Near Real Time.” Sensors, vol. 17, no. 12, Dec. 2017, p. 2766.
Review.” ISPRS International Journal of Geo-Information, vol. 9, no. 1, [Link], [Link]
Dec. 2019, p. 14. [Link] (Crossref),
[Link] [29] Tarasconi, Francesco, et al. “The Role of Unstructured Data in Real-Time
Disaster-Related Social Media Monitoring.” 2017 IEEE International
[7] Mousavi, S. Mostafa, et al. “Gemini and Physical World: Large Language Conference on Big Data (Big Data), 2017, pp. 3769–78. IEEE Xplore,
Models Can Estimate the Intensity of Earthquake Shaking from [Link]
Multimodal Social Media Posts.” Geophysical Journal International, vol.
240, no. 2, Dec. 2024, pp. 1281–94. [Link] (Crossref), [30] .Havas, Clemens, et al. “E2mC: Improving Emergency Management
[Link] Service Practice through Social Media and Crowdsourcing Analysis in
Near Real Time.” Sensors, vol. 17, no. 12, Dec. 2017, p. 2766.
[8] Estêvão, João M. C. “Effectiveness of Generative AI for Post-Earthquake [Link], [Link]
Damage Assessment.” Buildings, vol. 14, no. 10, Oct. 2024, p. 3255.
[Link] (Crossref), [Link] [31] Said, Naina, et al. “Natural Disasters Detection in Social Media and
[9] Bhadauria, P. K. S. “Comprehensive Review of AI and ML Tools for Satellite Imagery: A Survey.” Multimedia Tools and Applications, vol.
78, no. 22, Nov. 2019, pp. 31267–302. Springer Link,
Earthquake Damage Assessment and Retrofitting Strategies.” Earth [Link]
Science Informatics, vol. 17, no. 5, Oct. 2024, pp. 3945–62. [Link]
(Crossref), [Link] [32] Yang, Lu. “Conditional Generative Adversarial Networks (CGAN) for
Abnormal Vibration of Aero Engine Analysis.” 2020 IEEE 2nd
[10] Lentile*, Leigh B., et al. “Remote Sensing Techniques to Assess Active International Conference on Civil Aviation Safety and Information
Fire Characteristics and Post-Fire Effects.” International Journal of Technology (ICCASIT, 2020, pp. 724–28. IEEE Xplore,
Wildland Fire, vol. 15, no. 3, Sept. 2006, pp. 319–45. [Link]
[Link], [Link]
[33] Xing, Ziyao, et al. “Flood Vulnerability Assessment of Urban Buildings
[11] Szpakowski, David M., and Jennifer L. R. Jensen. “A Review of the Based on Integrating High-Resolution Remote Sensing and Street View
Applications of Remote Sensing in Fire Ecology.” Remote Sensing, vol.
 Images.” Sustainable Cities and Society, vol. 92, May 2023, p. 104467. [54] Ferrara, Emilio. “Fairness and Bias in Artificial Intelligence: A Brief
ScienceDirect, [Link] Survey of Sources, Impacts, and Mitigation Strategies.” Sci, vol. 6, no. 1,
[34] Ghimire, Prashnna, et al. “Opportunities and Challenges of Generative AI Dec. 2023, p. 3. [Link] (Crossref), [Link]
in Construction Industry: Focusing on Adoption of Text-Based Models.” [55] Barni, M., et al. “Improved Wavelet-Based Watermarking through Pixel-
Buildings, vol. 14, no. 1, Jan. 2024, p. 220. [Link] (Crossref), Wise Masking.” IEEE Transactions on Image Processing, vol. 10, no. 5,
[Link] May 2001, pp. 783–91. [Link] (Crossref),
[35] Dong, Chao, et al. “Image Super-Resolution Using Deep Convolutional [Link]
Networks.” IEEE Transactions on Pattern Analysis and Machine [56] Stallings, William. Cryptography and network security: principles and
Intelligence, vol. 38, no. 2, Feb. 2016, pp. 295–307. IEEE Xplore, practice. Seventh edition., Pearson, 2022
[Link] [57] Ronneberger, Olaf, et al. U-Net: Convolutional Networks for Biomedical
[36] Alzantot, Moustafa, et al. “SenseGen: A Deep Learning Architecture for Image Segmentation. arXiv:1505.04597, arXiv, 18 May 2015. [Link],
Synthetic Sensor Data Generation.” 2017 IEEE International Conference [Link] Montufar, Guido.
on Pervasive Computing and Communications Workshops (PerCom [58] Restricted Boltzmann Machines: Introduction and Review.
Workshops), 2017, pp. 188–93. IEEE Xplore, arXiv:1806.07066, arXiv, 19 June 2018. [Link],
[Link] [Link]
[37] Viola, Paul & Jones, Michael. (2001). Robust Real-Time Object [59] Grünthal, G. (ed.), Musson, R., Schwarz, J., Stucchi, M. (1998): European
Detection. International Journal of Computer Vision - IJCV. 57. Macroseismic Scale 1998. Cahiers de Centre Européen de Géodynamique
[38] Khan, Talha Ahmed, et al. “Investigation of Flash Floods on Early Basis: et de Séismologie, Volume 15, Luxembourg
A Factual Comprehensive Review.” IEEE Access, vol. 8, 2020, pp. [60] Simonyan, Karen, and Andrew Zisserman. Very Deep Convolutional
19364–80. IEEE Xplore, Networks for Large-Scale Image Recognition. arXiv, 2014. [Link]
[Link] (Datacite), [Link]
[39] Galantucci, Rosella Alessia, and Fabio Fatiguso. “Advanced Damage [61] He, Kaiming, et al. Deep Residual Learning for Image Recognition.
Detection Techniques in Historical Buildings Using Digital arXiv, 2015. [Link] (Datacite),
Photogrammetry and 3D Surface Anlysis.” Journal of Cultural Heritage, [Link]
vol. 36, Mar. 2019, pp. 51–62. [Link] (Crossref),
[Link] [62] Breiman, Leo. “Random Forests.” Machine Learning, vol. 45, no. 1, Oct.
2001, pp. 5–32. Springer Link,
[40] Arbuckle, Luk, and Khaled El Emam. Building an Anonymization [Link]
Pipeline: Creating Safe Data. O’Reilly, 2020.
[63] Chen, Tianqi, and Carlos Guestrin. XGBoost: A Scalable Tree Boosting
[41] Smadi, Takialddin Al, et al. “Artificial Intelligence for Speech System. 2016. [Link] (Datacite),
Recognition Based on Neural Networks.” Journal of Signal and [Link]
Information Processing, vol. 06, no. 02, 2015, pp. 66–72. [Link]
(Crossref), [Link] [64] Friedman, Jerome H. 2001. “Greedy Function
Approximation: A Gradient Boosting Machine.” The Annals of
[42] Gregory, Sam. “Deepfakes, Misinformation and Disinformation and Statistics 29 (5): 1189–1232. [Link]
Authenticity Infrastructure Responses: Impacts on Frontline Witnessing,
Distant Witnessing, and Civic Journalism.” Journalism, vol. 23, no. 3, [65] Howard, Andrew G., et al. MobileNets: Efficient Convolutional Neural
Mar. 2022, pp. 708–29. [Link] (Crossref), Networks for Mobile Vision Applications. arXiv, 2017. [Link]
[Link] (Datacite), [Link]
[43] Farid, H. “Image Forgery Detection.” IEEE Signal Processing Magazine, [66] Tan, Mingxing, and Quoc V. Le. EfficientNet: Rethinking Model Scaling
vol. 26, no. 2, Mar. 2009, pp. 16–25. [Link] (Crossref), for Convolutional Neural Networks. 2019. [Link] (Datacite),
[Link] [Link]
[44] Sharma, Dilip Kumar, et al. “A Survey of Detection and Mitigation for [67] Lim, Bee, et al. Enhanced Deep Residual Networks for Single Image
Fake Images on Social Media Platforms.” Applied Sciences, vol. 13, no. Super-Resolution. arXiv, 2017. [Link] (Datacite),
19, Oct. 2023, p. 10980. [Link] (Crossref), [Link]
[Link] [68] Zhang, Yulun, et al. Image Super-Resolution Using Very Deep Residual
[45] Caled, Danielle, and Mário J. Silva. “Digital Media and Misinformation: Channel Attention Networks. arXiv, 2018. [Link] (Datacite),
An Outlook on Multidisciplinary Strategies against Manipulation.” [Link]
Journal of Computational Social Science, vol. 5, no. 1, May 2022, pp. [69] Redmon, Joseph, et al. You Only Look Once: Unified, Real-Time Object
123–59. Springer Link, [Link] Detection. arXiv, 2015. [Link] (Datacite),
[46] Wang, Xiaofeng, et al. “A Visual Model-Based Perceptual Image Hash [Link]
for Content Authentication.” IEEE Transactions on Information Forensics [70] Abdul, Zrar Kh., and Abdulbasit K. Al-Talabani. “Mel Frequency
and Security, vol. 10, no. 7, July 2015, pp. 1336–49. IEEE Xplore, Cepstral Coefficient and Its Applications: A Review.” IEEE Access, vol.
[Link] 10, 2022, pp. 122136–58. IEEE Xplore,
[47] Zanardelli, Marcello, et al. “Image Forgery Detection: A Survey of [Link]
Recent Deep-Learning Approaches.” Multimedia Tools and Applications, [71] Dosovitskiy, Alexey, et al. An Image Is Worth 16x16 Words:
vol. 82, no. 12, May 2023, pp. 17521–66. Springer Link, Transformers for Image Recognition at Scale. arXiv, 2020. [Link]
[Link] (Datacite), [Link]
[48] Szegedy, Christian, et al. Intriguing Properties of Neural Networks. [72] Liu, Ze, et al. Swin Transformer: Hierarchical Vision Transformer Using
arXiv:1312.6199, arXiv, 19 Feb. 2014. [Link], Shifted Windows. arXiv, 2021. [Link] (Datacite),
[Link] [Link]
[49] Ali, Sajid, et al. “Explainable Artificial Intelligence (XAI): What We [73] Lundberg, Scott, and Su-In Lee. A Unified Approach to Interpreting
Know and What Is Left to Attain Trustworthy Artificial Intelligence.” Model Predictions. arXiv, 2017. [Link] (Datacite),
Information Fusion, vol. 99, Nov. 2023, p. 101805. [Link] (Crossref), [Link]
[Link] [74] Ribeiro, Marco Tulio, et al. “Why Should I Trust You?”: Explaining the
[50] Yu, Yang, et al. “Multi-Scale Spatial Pyramid Attention Mechanism for Predictions of Any Classifier. arXiv, 2016. [Link] (Datacite),
Image Recognition: An Effective Approach.” Engineering Applications [Link]
of Artificial Intelligence, vol. 133, July 2024, p. 108261. ScienceDirect, [75] Zhou, Jie, et al. “Graph Neural Networks: A Review of Methods and
[Link] Applications.” AI Open, vol. 1, 2020, pp. 57–81. [Link] (Crossref),
[51] Dwork, Cynthia. “Differential Privacy.” Automata, Languages and [Link]
Programming, edited by Michele Bugliesi et al., Springer, 2006, pp. 1– [76] Mennecke, B. E., and M. D. Crossland. “Geographic Information
12. Springer Link, [Link] Systems: Applications and Research Opportunities for Information
[52] Hassija, Vikas, et al. “Interpreting Black-Box Models: A Review on Systems Researchers.” Proceedings of HICSS-29: 29th Hawaii
Explainable Artificial Intelligence.” Cognitive Computation, vol. 16, no. International Conference on System Sciences, vol. 3, 1996, pp. 537–46
1, Jan. 2024, pp. 45–74. Springer Link, [Link] vol.3. IEEE Xplore, [Link]
023-10179-8. [77] Dong, Chenhe, et al. A Survey of Natural Language Generation. 2021.
[53] M. Anderson and S. L. Anderson, “Machine Ethics: Creating an Ethical [Link] (Datacite), [Link]
Intelligent Agent”, AIMag, vol. 28, no. 4, p. 15, Dec. 2007.