An Adaptive AI‑Driven Multi‑Hazard Prediction and Early Alert

Framework for Real‑Time Emergency Response Using Sensor Fusion

and Deep Learning Models

Kavya Sree K.

, Ankit Kumar

, P. Mathiyalagan

, M. Vineesha

Kartheeswari M.

and Syed Hauider Abbas

Department of Artificial Intelligence and Machine Learning, Ballari Institute of Technology and Management,

Ballari‑583104, Karnataka, India

Department of Computer Application, Faculty of Science, Swami Vivekanand Subharti University, Meerut, Uttar Pradesh,

India

Department of Mechanical Engineering, J.J. College of Engineering and Technology, Tiruchirappalli, Tamil Nadu, India

Department of Computer Science and Engineering MLR Institute of Technology, Hyderabad‑500043, Telangana, India

Department of ECE, New Prince Shri Bhavani College of Engineering and Technology, Chennai, Tamil Nadu, India

Department of Computer Science & Engineering, Integral University, Lucknow, Uttar Pradesh, India

Keywords: Multi‑Hazard Prediction, Early Warning System, Sensor Fusion, Deep Learning, Emergency Response.

Abstract: Rapid, accurate and smart response is essential for any disaster management system to minimize life and

property loss. Research approaches that have addressed real-time simulation have been single-disaster focused

or have only provided non-real-time testing or little or no real-time interface. In order to circumvent these

limitations, this work suggests an adaptive AI-powered multi-hazard prediction and early warning system via

sensor ensemble and deep learning capabilities. Utilizing information from Internet of Things (IoT) sensors,

satellite imagery, weather stations and drones, the platform supports real-time detection and forecasting of a

range of disasters, including floods, earthquakes, forest fires and cyclones. The architecture uses an automated

deep learning pipeline combined with lifelong learning for updating environment dynamics. It also offers

emergency response and public alert dissemination accurately by scalable deployment at edge-cloud. The

model is tested with two real-world datasets of disaster situations and compared to approaches based on

traditional systems, showing higher precision in providing alerts to the affected public, lower response time,

and fault tolerance to the operation in different regions. It is this model that seeks to shift disaster management

from response to recovery.

1 INTRODUCTION

Natural and man-made disasters are confronted with

increasing intensity, putting global safety, urban

resilience, and humanitarian response under the test.

From flash floods inundating our towns and cities,

earthquakes effecting our infrastructure and wildfires

threatening lives, the very nature of these events

fluctuate on the spur of the moment and the scale

varies from one event to the next, so there is an urgent

need for not just quick action but for smart thinking

ahead of any such event. On the one hand, many

traditional disaster management systems are of great

value but are generally based on reactive scenarios,

static data models and isolated sensor schemes, which

significantly reduce the ability for predicting and

responding in real time.

Recent growth in Artificial Intelligence (AI),

along with the development of sensors technologies

as well as edge computing, have paved ways to

proactive responses toward disaster risk reduction.

However, existing AI systems so far tend to target

isolated risks and lack the cohesion necessary for

synchronous, multimode input handling. In addition,

most if not all of the works focus in controlled

simulations and fail under realistic emergency

conditions.

This paper fills this gap by developing an

adaptive AI based framework for multi-hazard

522

K., K. S., Kumar, A., Mathiyalagan, P., Vineesha, M., M., K. and Abbas, S. H.

An Adaptive AI-Driven Multi-Hazard Prediction and Early Alert Framework for Real-Time Emergency Response Using Sensor Fusion and Deep Learning Models.

DOI: 10.5220/0013868600004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 1, pages

522-529

ISBN: 978-989-758-777-1

disaster prediction and early warning. Leveraging

information from diverse sources, including IoT

sensors, satellite feeds, meteorological stations and

unmanned aerial vehicles, our proposed architecture

provides a strong foundation for real-time situational

awareness. The paradigm is based on hybrid deep

models that can learn on the fly and adapt themselves

to changing environmental dynamics acting in

dynamic scenarios.

Compared to the mainstream solutions, this one

not only focuses on detecting danger warnings in

advance but also can dynamically support the

emergency rescuing with real-time warning and

intelligent rescue coordination. The proposed model,

therefore, changes emergency management from a

manual and delayed to a synchronized and AI-

enabled life cycle, from early detection to immediate

data-driven response.

1.1 Problem Statement

Although disaster risk management technologies

have advanced, the systems in place today are

inadequate in providing precise and up-to-date

forecasts for a variety of hazards. The majority of

current approaches are either restrictively based on

particular hazards, exclusively historical data

dependant or infrastructure-dependent, e.g., working

with isolated sensing networks and centralized

hardware systems. There results in the scenarios

delayed notifications, lack of situational awareness,

and suboptimal coordination of emergency responses.

Beyond that, most AI models for disaster

prediction lack the ability to adapt to changing

environment, or to leverage various types of data

source such as IoT devices, satellite imagery, and

meteorological input. This leads to overfit models for

specific locales, or models that are not able to deal

with real-world uncertainty and infrastructure failure,

which is widespread in extreme events.

A solution that is smart, adaptive, and tough

enough to anticipate and withstand range of disaster

is required at speed. This system would need to

integrate heterogeneous sources, use deep learning

for accurate predictions, and provide alerts to

stakeholders such as first responders and vulnerable

populations in a timely manner.

This work aims to bridge the gap with an AI-

driven early warning system capable of anticipating

and detecting multiple hazards, dynamically learning

from ongoing hazardous events, and communicating

urgent information in real time with the ultimate goal

of transforming disaster response from a reactive

approach to one that prepares, informs and anticipates

- to establish smart resilience in advance of threats.

2 LITERATURE SURVEY

The utilization of Artificial Intelligence (AI) in

disaster prediction and emergency management has

received much attention in the past decade. Numerous

research studies have investigated the use of machine

learning, deep learning, and sensing technologies to

forecast disasters faster and better. However, these

studies tend to concentrate on a particular type of

disaster, such as using air pollution data rather than a

composite one, and they are not fully integrated,

thereby preventing practical deployment for real time

disasters. Alladi (2022) introduced an AI-based early

warning system in which environmental data are

used to predict disasters. Although useful in

simulation, this was not validated against real

disaster data. Likewise, Bhattarai (2021) used deep

learning and augmented reality for firefighting

awareness, but the system was not portable to outdoor

or urban-scale systems. Cani et al. (2025) have

developed the TRIFFID system to assist first

responders assisted by autonomous robots, but it is

still at the prototype stage and has not yet been

deployed.

Chamola et al. (2024) surveyed AI methods for

disaster prediction and identified the shortcomings of

real-time operations and field data inputs. Nevo et al.

(2021) applied a machine learning model in an

operational framework for flood forecasting, but

reported a difficulty of real-time adaptation for

sensors. Zhang et al. (2023) presented earthquake

early warning by wireless sensor networks whose

communication is degraded in destroyed or

disconnected situation. Melo et al. (2025) also

predicted floods using AI and process-based models,

and the system was limited to estuarine only. Wang et

al. (2022) adopted hybrid CNN-LSTM models for

predicting the seismic responses, which were still

effective but not universally feasible in other types of

disasters. Similarly, Lyu et al. (2021) focussed on

local predictions of landslides, with an emphasis on

model transferability.

On the integration front, Potter (2024)

highlighted IoT-AI partnership for disaster alerts

without addressing concerns such as latency or real-

time decision making. Early work the early warning

systems literature by Della Mura (2024) and Sahota

(2023) describe the concept and mechanism of early

warning in the context of war, but do not delve into

the algorithmic techniques. Deloitte (2024) and Step

An Adaptive AI-Driven Multi-Hazard Prediction and Early Alert Framework for Real-Time Emergency Response Using Sensor Fusion and

Deep Learning Models

523

of Web (2025) demonstrated AI’s policy and strategic

potential but not technical detail. In the recent years

also the multi-hazard situation is becoming

considered. Singh & Pal (2021) deployed ensemble

models for flood forecasting, and Thomas & Balan

(2022) proposed AI-supported edge device-based

forest fire detection. Lin et al. (2022) used AI on

clouds for tropical cyclone behavior but could not

generalize regionally. Ghosh & Ghosh (2023)

exploited satellite imagery to detect tsunamis,

though limited temporal resolution is another issue.

However, despite this progress, it is still a

challenge to develop a coherent, adaptive, and real-

time system using AI that processes all types of data,

i.e., satellite images, IoT sensors and meteorological

data, and reacts to a changing disaster scenario.

To this end, this study presents an adaptive multi-

hazard framework based on deep learning, real-time

sensor fusion, and edge-cloud computing, which will

generate accurate early warnings and facilitate quick

emergency responses in diverse environments.

3 METHODOLOGY

The envisaged approach is organized to establish an

adaptive AI framework in order to model, design, and

implement intuitively predicting various immediate

disaster types, remotely issuing early warning and

enhancing all agencies’ coordination in emergency

response. The overall architecture of the system

consists of five principal stages: data acquisition;

preprocessing and fusion; model training; real-time

alert triggering; and performance assessment (figure

1).

Figure 1: System Architecture Diagram.

The first phase is namely data collection, it gathers

multi-source data from different sensors or data

stores. The IoT devices installed in these exposed

areas present environmental measurements, including

temperature, humidity, seismic vibrations, and water

depth. At the same time, satellite image data is also

collected through open APIs like NASA Earthdata

and ESA Sentinel, which brings real-time spatial

data. Meteorological information is found on national

weather service’s databases, and historical records on

disaster events are sourced from open repositories

such as EM-DAT and Kaggle disaster datasets.

Drone footage is seamlessly incorporated where

available to improve aerial situational awareness in

disasters.

Table 1 gives the Sensor Data Sources and

Characteristics.

Table 1: Sensor Data Sources and Characteristics.

Sensor Type Data Collected Source Frequency Format

Weather Station Rainfall, Temp, Humidity IMD / NOAA Hourly CSV/JSON

IoT Water Level River/Drainage Level On-site Nodes Every 10 mins JSON

Seismic Senso

Vibrations, Tremors Geolo

ical De

tEver

5 secs CSV

Satellite Imagery Surface Images, Clouds NASA/ESA APIs 1–3 hours GeoTIFF

Drone Camera Aerial Imagery, Smoke Plumes Emergency Deployed On demand MP4/PNG

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

524

Emphasis is also laid on the synchronization of the

collected inputs in the data preprocessing and fusion

stage mainly using a common timestamp and

geotagging protocol. For the problem of missing data,

KNN algorithms are employed to fill in the gap, and

the noise is suppressed by using Kalman filters as

well as the method of low-pass smoothing. A

multimodal fusion approach is utilized afterwards,

combining the structured sensor data with the

unstructured satellite imagery through data alignment

based feature extraction. CNN is used for extracting

image data and perform a statistical normalization

pipeline with numerical sensor data.

A hybrid deep learning architecture is developed

for the model which integrates the CNN with Long

Short-Term Memory (LSTM) networks. The CNNs

are responsible for spatial pattern recognition over

satellite or drone imagery (e.g. flood extents, smoke,

cracks on surface). On the other hand, the temporal

dependencies in the sensor time series is modelled to

learn the evolution of disaster patterns using the

LSTM layers. This hybrid architecture is trained on

labeled datasets using cross-entropy loss and the

Adam optimizer. Reinforcement learning is

embedded in the model to increase model flexibility

where we implement a Q-learning style reward that

updates weights with real-time accuracy of

predictions.

Table 2: Deep Learning Model Configuration.

er T

e Details

Input Layer

Multimodal inputs (sensor +

ima

CNN Layers

3 Conv2D layers with ReLU

activation

LSTM Layers

2 stacked LSTM layers (64

units each)

Dense La

Full

connected, dro

out

(

0.3

)

Output Layer

Softmax activation (for hazard

classification

)

Optimize

Adam (learning rate: 0.001)

Loss Function Cate

orical Cross-Entro

Reinforcemen

t Lo

Q-Learning reward based on

alert accurac

The model, which has been trained, is deployed in

edge-cloud hybrid. In the real-time alert generation

block, the edge devices locally infer to identify on

going threats based on small and compressed versions

of the model. In parallel, the cloud performs high-

resolution analysis and prediction of disaster

propagation. Automatic alerts are activated when

prediction confidence exceeds a dynamically set

threshold and broadcasted to emergency response

groups and to public networks via SMS, email, public

broadcasting APIs, and mobile applications.

Last, the performance phase involves offline

validation as well as real-time system monitoring.

Confusion matrix and time delay analysis are used for

computing accuracy, precision, recall, F1-score and

latency on the test datasets. Performance is checked

on the fly by comparing timestamps of alerts with

incident incident reports. Furthermore, user feedback

from emergency responders is obtained to evaluate

usability and actionability. This approach guarantees

that the proposed system is not only technically

sound and smart but also feasible and deployable in

various geographical and infrastructural conditions.

4 RESULT AND DISCUSSION

The adaptive AI-based framework was tested under

historical and up-to-the-minute data in multiple

hazard domains (floods, fires, earthquakes, and

cyclone). Figure 3 shows Confusion Matrix Showing

Hazard Classification Accuracy Across Disaster

Types The hybrid CNN-LSTM model was trained on

a combined dataset sourced from government

meteorological data, IOT sensor logs and satellite

logs. Figure 2 illustrates the training and validation

loss convergence during model training Overall

system performance was evaluated according to the

confidence placed in models based on standard

classification metrics, as well as latency measures to

evaluate the potential for real-time deployment of the

emergency response system.

Table 3 tabulates the

performance metrics for hazard detection.

Figure 2: Training Vs. Validation Loss Over Epochs.

An Adaptive AI-Driven Multi-Hazard Prediction and Early Alert Framework for Real-Time Emergency Response Using Sensor Fusion and

Deep Learning Models

525

Table 3: Performance Metrics for Hazard Detection.

Disaster

Type

Precision Recall

F1-

Scor

Accurac

Floo

0.97 0.95 0.96 0.96

Earthquake 0.93 0.91 0.92 0.94

Wildfire 0.91 0.88 0.89 0.92

Cyclone 0.95 0.93 0.94 0.95

Overall

Average

0.94 0.92 0.93 0.94

Overall, the model achieved a 94.3% accuracy for

predicting all hazard types, with the highest accuracy

occurring for floods (97.2%), and somewhat lower

performance in wildfires (91.8%), primarily because

plumes of smoke from fires and cloud patterns in

satellite imagery are visually similar. The average F1-

score was 0.93, showing a good balance of precision

and recall across the disaster classes. The rate of false

positives stayed just under 6%, an important standard

for keeping panic in check and ensuring only alarms

worth acting on go to first responders and the public.

Figure 3 gives the confusion matrix for multi-hazard

classification.

Figure 3: Confusion Matrix for Multi-Hazard

Classification.

In terms of responsiveness, the edge-deployed variant

of the model achieved a median time of 3.5 seconds

during the data ingestion when the alerts were

generated. Figure 4 illustrates the average latency of

adversarial alert generation among three deployment

modes (edge, cloud, and hybrid).This low latency

was achieved by deploying compressed, quantized

models on the leaf nodes, but keeping the full-

resolution model in the cloud for deep pattern

analysis. The cloud based model enabled ramp-up

forecasting of disasters with forecasting horizons of

up to 6 hours, obtaining a Root Mean Square Error

(RMSE) of 0.07 when used to forecast disaster

severity evolution over time.

The proposed model was compared with state-of-

the-art single-source models, e.g., isolated CNNs and

rule-based expert systems. The purely data-driven

models performed between 17–21% worse overall,

as well as being up to 42% slower with respect to alert

time. Furthermore, it showed increased robustness in

data-limited scenarios because of the multimodal

inputs and the reinforcement learning module, which

provided the possibility of real-time fine-tuning.

Figure 4 and table 4 shows the alert generation

latency by deployment mode.

Figure 4: Alert Generation Latency by Deployment Mode.

A qualitative data was collected from

experimental emergency scenarios in three local

disaster relief working groups. Respondents praised

the system’s visual dashboard for being easy to

understand, its real-time heat maps, and the clarity of

the alert messages. The integration with the

communication API allowed the sending of

automated alerts to responders using SMS, to the

public using a prototype mobile app, which increased

information-flow rate and area coverage.

Table 4: Alert Generation Latency Analysis.

Model

Deployment

Type

Average

Latency

(sec)

Standard

Deviation

Connectivit

y Required

e Device 3.5 0.8 Low

Cloud Serve

7.9 1.3 High

Hybrid

Mode

4.2 0.6 Medium

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

526

There were some restrictions, though. The system

performance degraded slightly (by approximately 4–

5%) in the case of extreme low-connectivity

environments, when edge devices could barely access

any real-time satellite data. the Figure 5 and table 5

gives the accuracy and alert delay comparison

between proposed model and traditional systems. It

indicates that the model effeciency can be improved

in disconnected setting or partial connected setting in

the future, by felderrated learning, or good use of

local synthetic data.

Figure 5: Performance Comparison With Existing Models.

Table 5: Comparison With Existing Methods.

Model Accuracy

Alert Delay

(sec)

Multi-Hazard Support Adaptability

Rule-Based Expert

System

78.2% 12.4

✗

Low

CNN Onl

85.6% 8.1 Limite

Mediu

LSTM Only 87.9% 7.5

✗

Medium

Proposed CNN-LSTM +

94.3% 3.5

✓

High

The overall results technically validate the proposed

multi-hazard prediction and alert system as

technically sound, faster and operationally feasible.

Its general applicability across different types of

disasters, real-time features, and user-orientation,

make it a good contender to be incorporated into

present disaster preparedness infrastructure.

5 CONCLUSIONS

In the context of rising number and severity of

natural catastrophe incidents, a cornerstone of

resilient disaster management includes predicting

hazardous events and reacting in real-time. This

study proposed an adaptable AI-integrated system to

overcome major drawbacks of prior systems, which

are lack of scalability, single-hazard orientation, and

lag-induced alert generation. Through a combination

of deep learning models and sensor fusion

methodologies, employed on a hybrid edge-cloud

infrastructure, the presented system achieved high

precision, low latency, and multi-danger versatility.

The addition of a CNN-LSTM hybrid structure

effectively facilitated the learning of spatial-temporal

disaster patterns, and the introduction of

reinforcement learning made the system more

capable of adjusting to new data and situations. Real-

world simulations and empirical evaluation on

An Adaptive AI-Driven Multi-Hazard Prediction and Early Alert Framework for Real-Time Emergency Response Using Sensor Fusion and

Deep Learning Models

527

various real-world datasets proved the robustness of

the system, as it achieves strong performance for

floods, earthquakes, wildfires, and cyclones.

Furthermore, its wide-spread alert in short time and

the potential use for emergency teams by providing

actionable information are promising on the practical

level.

More than just an app the system is also designed to

be accessible, scalable and work in conjunction with

any existing emergency protocols, and thus it’s a

highly useful asset for governments, humanitarian

organizations, and local authorities. Although the

proposed framework achieves strong baseline

performance, future work could improve the solution

from the offline aspect by using federated learning,

extend the disaster types, and make real-time drone-

based anomaly detection better.

Then there’s the fact that this research represents

an enormous step toward proactive AI-enabled

resiliency: changing the game in how communities

can forecast, plan for and respond to emergencies

with intelligence, velocity and precision.

REFERENCES

Alladi, D. (2022). AI-driven early warning systems for

natural disaster prediction. International Journal of

Smart Distributed Computing Systems, 4(4).

https://ijsdcs.com/index.php/ijsdcs/article/view/628

Bhattarai, M. (2021). Integrating deep learning and

augmented reality to enhance situational awareness in

firefighting environments arXiv.

https://arxiv.org/abs/2107.11043

Cani, J., Koletsis, P., Foteinos, K., Kefaloukos, I.,

Argyriou, L., & Papadopoulos, G. T. (2025). TRIFFID:

Autonomous robotic aid for increasing first responders'

efficiency. arXiv. https://arxiv.org/abs/2502.09379

Chamola, V., Linardos, A., & others. (2024). Artificial

intelligence and machine learning for disaster

prediction: A review. Natural Hazards, 110(3), 1234–

1256. https://doi.org/10.1007/s11069-024-06616-y

Dattamajumdar, A. (2021). An early warning AI-powered

portable system to reduce workload and inspect

environmental damage after natural disasters. arXiv.

https://arxiv.org/abs/2104.00876

Della Mura, M. T. (2024). Early warning systems:

Harnessing AI to mitigate risks. Tech4Future.

https://tech4future.info/en/early-warning-systems-ai/

Deloitte Insights. (2024). Leveraging AI in emergency

management and crisis response.

https://www2.deloitte.com/us/en/insights/industry/pub

lic- sector/automation-and-generative-ai-in-

government/le veraging-ai-in-emergency-

management-and-crisis-response.html

Ghosh, S., & Ghosh, D. (2023). Tsunami detection using

satellite images and deep learning. Remote Sensing

Letters, 14(2), 177–

187. https://doi.org/10.1080/215070 4X.2023.2164729

Jaiswal, A., & Joshi, R. (2023). Drone-assisted real-time

disaster response system using AI and computer vision.

Computers, Materials & Continua, 75(2), 2345–2360.

https://doi.org/10.32604/cmc.2023.030261

Jamison, T. (2024). Smart dispatching: How artificial

intelligence is reshaping emergency response. Police1.

Kim, H., & Lee, S. (2021). Urban flood risk mapping using

machine learning and geographic information systems.

Environmental Modelling & Software, 144, 105160.

https://doi.org/10.1016/j.envsoft.2021.105160

Kuglitsch, M. (2024). AI to the rescue: How to enhance

disaster early warnings with artificial intelligence.

Nature. https://www.nature.com/articles/d41586-024-

03149-z

Lin, K., Chen, Z., & Fang, Y. (2022). Tropical cyclone

prediction using AI-driven cloud platforms. Weather

and

Climate Extremes, 38, 100469. https://doi.org/10.101

6/j.wace.2022.100469

Lyu, Y., Hu, Y., & Cao, Y. (2021). Real-time landslide

prediction using deep learning and environmental

sensor

networks. Sensors, 21(8), 2675. https://doi.org/10.339

0/s21082675

Melo, W. W. de, Iglesias, I., & Pinho, J. (2025). Early

warning system for floods at estuarine areas:

Combining artificial intelligence with process-based

models. Natural Hazards, 121, 4615–

4638. https://doi.org/10.1007/s1106 9-024-06957-8

Nevo, S., Morin, E., Gerzi Rosenthal, A., Metzger, A.,

Barshai, C., & Matias, Y. (2021). Flood forecasting

with machine learning models in an operational

framework. arXiv. https://arxiv.org/abs/2111.02780

Potter, K. (2024). AI and IoT integration for disaster

management and early warning systems. ResearchGate.

https://www.researchgate.net/publication/384662288

RoX818. (2025). AI predicts & prevents humanitarian

disasters early. AI Competence. https://aicompetence.

org/ai-predicts-prevents-humanit arian-disasters-early/

Sahota, N. (2023). AI in disaster management: AI’s role in

disaster risk reduction. https://www.neilsahota.com/ai-

in-disaster-management-ais-role-in-disaster-risk-

reduction/

Scientia Educare. (2025). The future of AI in disaster

response and emergency management. https://scientia

educare.com/the-future-of-ai-in-disaster-response-and-

emergency-management/

Singh, A., & Pal, R. (2021). A hybrid ensemble model for

flood prediction and disaster management. Journal of

Hydrology, 603, 127017. https://doi.org/10.1016/j.jhy

drol.2021.127017

Step of Web. (2025). The impact of AI on emergency

response systems: Transforming the way we respond to

crises. https://stepofweb.com/ai-enhanced-emergency-

response-systems/

Thomas, M., & Balan, R. (2022). Forest fire detection and

alert system using AI-enabled edge devices. IEEE

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

528

Sensors Journal, 22(15), 14860–

14867. https://doi.org/1 0.1109/JSEN.2022.3173915

Wang, S., Huang, Y., & Zhang, H. (2022). Deep learning-

based seismic hazard prediction using CNN-LSTM

models. IEEE Access, 10, 21734–21743. https://doi.or

g/10.1109/ACCESS.2022.3156552

Zhang, X., Li, J., & Liu, W. (2023). Earthquake early

warning using hybrid AI models and wireless sensor

networks. Engineering Applications of Artificial

Intelligence, 118, 105614. https://doi.org/10.1016/j.en

gappai.2023.105614

An Adaptive AI-Driven Multi-Hazard Prediction and Early Alert Framework for Real-Time Emergency Response Using Sensor Fusion and

Deep Learning Models

529