AI-GRID: GeoRisk Intelligence for Disaster-Resilient Healthcare

Ananya Joshi1, Rishika Shinde2 and Kshamta Mathur3

1
Mukesh Patel School of Technology Management and Engineering, NMIMS, Mumbai, India
2
Mukesh Patel School of Technology Management and Engineering, NMIMS, Mumbai, India
3
Mukesh Patel School of Technology Management and Engineering, NMIMS, Mumbai, India

{ananya.joshi56, rishika.shinde24}@[Link], [Link]@[Link]

Abstract

In Assam, India’s flood-prone northeast, monsoon rains, snowmelt, and glacial runoff frequently swell
rivers, endangering lives and disrupting healthcare access. Existing disaster management still depends on
static GIS maps and historical flood records, which cannot capture rapidly shifting risks under climate
change. This study develops an AI–GIS framework for near real-time flood risk prediction and
healthcare accessibility assessment. The pipeline fuses Sentinel-1 radar imagery, NASA snow water
equivalent indices, SRTM elevation data, WorldPop population grids, and OpenStreetMap road
networks. A stacked ensemble of Histogram-based Gradient Boosting, Logistic Regression, and
CatBoost classifiers was trained on features such as elevation–rainfall ratios and rainfall–soil moisture
interactions, achieving 72% accuracy and an F1-score of 0.71. The results power a Streamlit dashboard
with Google Earth Engine maps, live weather feeds, and automated alerts via AWS SNS/SMTP,
enabling pre-positioning of medical supplies and risk-aware planning. The framework is scalable
globally and supports SDGs 3, 11, and 13.

Keywords -Flood Prediction, Machine Learning, Geographic Information Systems (GIS), Disaster Risk
Management, Healthcare Accessibility, Ensemble Learning

1. Introduction

Climate-driven disasters are no longer rare shocks, they arrive in relentless waves, upending towns,
stalling economies, and leaving deep scars on public health [5], [6]. In Assam, a lush northeastern state
of India, seasonal monsoons combine with Himalayan snowmelt and glacial runoff, swelling rivers until
they overflow into fields and villages [14], [17]. The result is recurring floods that wash out roads,
disrupt power lines, and leave communities cut off for days or even weeks [20], [22]. Among the
hardest-hit sectors is healthcare: low-lying hospitals are frequently submerged, clinics become
inaccessible, and medical supply chains grind to a halt, leaving patients without care at precisely the
moment they need it most [1], [2], [21].Traditional disaster response in Assam still relies on static GIS
maps and after-the-fact reporting useful for planning but too slow and blunt for real-time decisions [1],
[18]. This reactive approach means resources often arrive late, sometimes after preventable injuries and
deaths. Meanwhile, advances in artificial intelligence (AI) and geospatial technology are opening new
pathways for proactive disaster preparedness [6], [9], [11]. By learning from rich, multimodal
environmental data, machine-learning models can capture the complex interactions between rainfall,
elevation, soil saturation, and glacial melt — for example, how a sudden warm rain on packed snow can
rapidly swell rivers [8], [10], [13].

This study contributes to disaster informatics, geospatial AI, and climate-resilient healthcare planning by
introducing an AI–GIS fusion framework that blends machine-learning flood forecasts with geospatial
models of healthcare access. The framework is tailored for flood- and snowmelt-prone regions like
Assam, where river overflows can cut off entire villages [6], [17], [18]. A stacked ensemble classifier
was developed, integrating Histogram-based Gradient Boosting, Logistic Regression, and CatBoost,
enriched with hydrologically informed interaction features such as elevation-to-rainfall ratios, rainfall–
soil moisture products, and river–glacier proximity composites [7], [8], [10].

To operationalize the model, it is embedded in a Streamlit-powered dashboard that fuses live weather
data from the Open-Meteo API with high-resolution Google Earth Engine maps [26]. The platform offers
interactive visualizations, including rainfall and temperature scatterplots, feature distribution bar graphs,
and time-series weather curves, enabling decision-makers to explore key drivers of flood risk. A Multi-
Channel Early Warning System delivers automated alerts via AWS SNS and Gmail SMTP, often
accompanied by geocoded maps and timestamps, ensuring healthcare leaders and disaster-response

Beyond Assam, this framework is designed as a scalable blueprint for flood-prone regions worldwide,
from river valleys in Southeast Asia to low-lying deltas in West Africa [5], [15], [16]. By coupling
predictive analytics with healthcare accessibility modelling, it helps advance global sustainability efforts,
aligning with UN SDGs 3 (Good Health and Well-Being), 11 (Sustainable Cities and Communities), and
13 (Climate Action) [5], [6]. Together, these contributions shift disaster management from reactive
response to proactive, data-driven resilience planning.

2. Related Work and Research Gap

Early flood studies relied heavily on GIS-based mapping to assess hazard-prone zones and overlay them
with critical infrastructure such as hospitals and evacuation routes. For instance, the Surabaya study
demonstrated spatial variation in healthcare accessibility during floods, but its results remained largely
descriptive and static without predictive modeling or operational alerts [1], [20], [21], [22], [23].
Similarly, retrospective studies such as AGU (2022) quantified health impacts after flood events but
failed to enable proactive planning [2].

Satellite remote sensing approaches using Sentinel-1 SAR, Sentinel-2 optical data, and Landsat have
become central to flood monitoring, with large-scale automation often implemented on Google Earth
Engine pipelines [3], [4]. However, most of these efforts stop at delineating inundation extents and do
not integrate hazard drivers such as rainfall, glacier melt, or soil moisture, nor do they connect outputs to
healthcare facility data for hospital-specific planning. Global systems such as GloFAS have achieved
basin-scale flood forecasting accuracy but require significant downscaling and socio-spatial integration
to be operationally relevant [5].

Recent work has explored machine learning techniques such as Random Forest, CatBoost, XGBoost, and
MaxEnt to model non-linear interactions among rainfall, soil type, slope, and land cover, producing
accurate susceptibility maps [6], [7], [8]. Studies in Indonesia and Bangladesh have shown how
population rasters like WorldPop can be used to estimate affected populations and accessibility [9].
However, these approaches remain largely static and retrospective, unable to capture dynamic road
outages, bridge failures, or shifting weather patterns in real time. This highlights the need for solutions
that can integrate multi-hazard prediction with healthcare accessibility modelling and deliver actionable
insights through operational dashboards.

3. Methodology
The AI–GIS framework was developed in several stages, each aimed at tracing how environmental
shifts, river patterns, and local infrastructure combine to shape flood risk in Assam, and transforming
these forecasts into actionable insights for healthcare access. The work was structured into three main
phases: data collection and preprocessing, predictive model development with feature engineering, and
deployment on a decision-support dashboard with real-time visualizations and interactive charts.

The proposed AI–

GIS framework
(Figure 1) starts
with collecting data
from sources like
satellite imagery,
 meteorological records, and population maps, which are then cleaned and aligned spatially and
temporally.

Fig.1. Workflow of GeoRisk Intelligence Model

Key hydrological features such as elevation–rainfall ratios, soil moisture interactions, and river–glacier
proximity are used to train a stacked ensemble model and validated via stratified k-fold cross-validation.
The model is deployed on a Streamlit dashboard, integrating live weather data and Google Earth Engine
layers for real-time flood risk assessment providing geographic inputs (latitude, longitude, elevation, soil
moisture) generating flood predictions with confidence scores, visualized on an interactive map. High-
risk areas are highlighted, and alerts are sent via SMS and email, making the framework operational for
immediate disaster response.

3.1 Data Acquisition and Preprocessing

To capture Assam’s flood dynamics, remote-sensing, meteorological, and demographic data were
combined. Sentinel-1 SAR imagery was processed using NDWI thresholding to delineate flood extents,
while MODIS and NASA GES DISC datasets provided snow cover and snow water equivalent estimates.
Elevation information was extracted from SRTM DEM to represent terrain variability [3].

Meteorological variables — precipitation, temperature, pressure, wind speed, and visibility were obtained
from NASA archives and supplemented with real-time feeds from the Open-Meteo API. Infrastructure and
population data were sourced from OpenStreetMap (OSM) and WorldPop, including hospital locations,
road networks, and population density grids [24], [26].

All raster layers were standardized to a common spatial resolution and coordinate reference system
(EPSG:4326), and vector data were cleaned in QGIS to fix topology errors. Missing values were imputed
using median strategies, and continuous variables were scaled using z-scores. A stratified train–test split
was applied to preserve class balance despite the rarity of flood events.

Historical flood frequency

was overlaid with current
flood detections. Purple
points show past flood
locations, black points
indicate current flooding,
and overlaps mark
recurrently inundated zones.
These zones were converted
into a binary flood-prone
mask for model training
(Figure 2), confirming
persistent hotspots along river belts.
Fig. 2. Flood Risk Overlay

3.2 Feature Engineering and Predictive Modelling

To improve model expressiveness, meaningful features were created, including:

 Elevation–Rainfall Ratio (elev_rain_ratio): Captures orographic effects on rainfall.

 Rainfall–Soil Moisture Product (rain_soil): Indicates potential runoff under saturated soil.

 Log Rainfall Transformation (log_rainfall): Reduces skew caused by heavy rainfall extremes.
 River–Glacier Proximity Composite: Represents risk from meltwater-fed rivers.

The task was framed as a binary classification problem (low-risk = 0, high-risk = 1) and addressed using
a stacked ensemble classifier combining:

 Histogram-Based Gradient Boosting (robust to missing values and mixed data),

 Logistic Regression (for well-calibrated probability estimates), and

 CatBoost as a Meta-learner (handling complex feature interactions and reducing overfitting) [8], [25].

Model robustness was validated using Stratified 5-Fold Cross-Validation to account for variability across
physiographic zones. Performance was evaluated using ROC–AUC, precision, recall, and F1-score, with
a focus on recall to minimize false negatives for high-risk cases.

3.3 Operational Deployment and Dashboard Integration

The modelling framework was operationalized as a Streamlit dashboard that integrates the trained model,
preprocessing pipeline, and live weather API feeds, enabling real-time forecast updates. The Real-Time
Prediction Interface allows users to input geographic details such as latitude, longitude, elevation, soil
moisture, and proximity to rivers or glaciers, while automatically incorporating live rainfall and weather
data from Open-Meteo. The processed features are passed through the stacked model, which predicts
flood risk and assigns each result a confidence score, for example, 92% for a high-risk zone.

The visualization layer (Figure 5) uses Folium maps to highlight high-risk regions in red and safer areas
in green. Additional panels display rainfall and temperature scatterplots, weather trends, and feature
distribution charts, allowing users to interpret what drives risk levels.

A dual-channel alert system was integrated, sending SMS notifications via AWS SNS and detailed email
alerts via SMTP with geocoded maps and timestamps, ensuring timely communication to hospital
administrators and disaster managers. Cloud integration with Google Earth Engine (GEE) enables
continuous updates with high-resolution satellite data, making the dashboard a dynamic, real-time
planning tool.

This modular deployment ensures that non-technical users including local authorities and healthcare staff
can interact with the system intuitively, supporting proactive decision-making even during ongoing flood
events.

4. Results

The ensemble model was evaluated on the test set using a variety of performance metrics. It achieved an
overall accuracy of 72.04% and a weighted F1-score of 0.71, demonstrating balanced performance across
both classes. Importantly, the model achieved a recall of 0.80 for flood-prone areas, which is crucial for
early warning applications where identifying as many true flood events as possible is the top priority.

Table 1. Classification report

Class Precision Recall F1-Score Support

0 0.63 0.50 0.56 382
1 0.72 0.86 0.78 610
Accuracy 0.72 992
Macro Avg 0.68 0.68 0.67 992
 Class Precision Recall F1-Score Support
Weighted Avg 0.72 0.72 0.71 992

The results show that the model performs especially well for the flood class, which is the primary focus
of this work. The high recall score means that most flood-prone cases were successfully detected.

To further validate the model, stratified train–test splits were applied to preserve class balance, and 5-
fold cross-validation was performed during training. The model consistently achieved comparable
accuracy and recall across folds, confirming stability and minimizing the risk of overfitting to a
particular data split. The dashboard includes multiple layers of contextual visualizations to support
decision-making.

Time-series plots of rainfall and

temperature (Figures 3 and 4) allow users
to spot weather trends that may signal
worsening conditions, such as sharp
rainfall spikes or persistent heat waves
that accelerate glacial melt. These
interactive graphs update automatically
with live API calls, ensuring that the
displayed patterns reflect the most current
environmental conditions.

Fig. 3. Rainfall intensity plotted over time,

helping users track storm development in near real time.

Fluctuations in temperature play a key role in

hydrological processes, as rising temperatures can
accelerate snowmelt and increase runoff, while
cooler periods may delay peak discharge.
Understanding these patterns is crucial for
assessing compound flood drivers and predicting
potential surges in river flow. The visual trend
highlights how sustained warming over
consecutive days can act as a precursor to
flooding events, complementing rainfall-
based indicators and improving early-
warning reliability.

Fig. 4. Temperature trends showing extreme

heat
periods that can contribute to snowmelt-driven
floods.

Feature distribution charts (Figure 5) further

enhance transparency by showing how input variables compare to historical data used in training. Users can
inspect how unusual current values are, such as whether rainfall levels are at the 90th percentile helping them
understand why a particular region was classified as high-risk.
 Fig. 5. Scaled rainfall distribution plots comparing current
environmental conditions against training data ranges, supporting interpretability.

The dashboard also provides a Weather

Metrics Overview (Figure 6), displaying
real-time trends for rainfall, temperature,
wind speed, visibility, and pressure on a
single timeline. This helps users see how
changing weather conditions relate to flood
risk predictions, offering quick situational
awareness. For example, rising rainfall and
falling pressure can signal worsening
conditions, allowing planners to take early action.

Fig. 6. Weather metrics time-series visualization

supporting real-time flood risk interpretation.

Figure 7 shows the spatial distribution of hospitals

across Chennai, India, mapped using latitude–
longitude data from OpenStreetMap. Each red point
represents a hospital location, allowing us to
visualize clusters of healthcare facilities relative to
flood-prone regions. Such geospatial visualization is
crucial because it enables researchers and
policymakers to quickly identify which hospitals are
likely to be affected by rising water levels and which
areas might suffer healthcare shortages during
disasters.

Fig. 7. Geospatial distribution of mapped hospitals in

Chennai used for healthcare accessibility analysis under flood scenarios

The integration of these visual elements ensures that predictions are not presented as “black-box” outputs
but rather as interpretable, data-rich insights that encourage informed action. This mapping approach
highlights the concentration of hospitals in the central and northern parts of the city, while certain
southern zones appear comparatively underserved. When overlaid with predicted flood zones, this
visualization can guide the pre-positioning of ambulances, allocation of mobile health units, and
prioritization of road-clearing operations in areas where healthcare access is most likely to be disrupted.

5. Discussion

The results highlight the strength of the proposed stacking ensemble model in predicting flood-prone
areas with both accuracy and reliability. Achieving an overall accuracy of 72.04% and a recall of 0.80 for
flood-prone zones, the model successfully meets the primary objective of maximizing early detection of
floods. This is a crucial outcome in disaster management, where the priority is to avoid missing potential
events and provide timely alerts for mitigation efforts.

The model’s performance is further supported

by the ROC curve, which achieved an AUC
score of 0.73. This indicates that the stacked
ensemble classifier is reasonably effective at
distinguishing between flood and non-flood
scenarios, performing well above random
chance (AUC = 0.5). The curve maintains a
strong separation from the diagonal baseline
across most thresholds, suggesting that the
model balances sensitivity (true positive rate)
and specificity (false positive rate).

Fig. 8. ROC-AUC curve

This result is particularly encouraging given the class imbalance in the dataset, where flood events are
comparatively rare. Despite this imbalance, the model demonstrates the ability to capture meaningful
patterns in the environmental and hydrological features, making it a reliable early-warning tool. An AUC
in this range also provides confidence to decision-makers that the predictions can be trusted to prioritize
resource allocation, such as pre-positioning medical supplies and evacuation planning.

A major contributor to this performance was the feature engineering process, which went beyond using
raw rainfall and elevation data. Features such as the elevation–rainfall ratio, rainfall–soil moisture
interaction, and log-transformed precipitation allowed the model to capture non-linear relationships
between topography, weather, and soil conditions. The use of polynomial interaction terms further
improved the model’s ability to detect subtle patterns that often precede flooding events.

Another important strength

is the interpretability of the
results. Feature importance
analysis confirmed that
elevation- and rainfall-
related variables were the
most influential in
determining flood risk,
aligning well with
hydrological
understanding. This
transparency makes the
solution easier to trust for
policymakers and disaster
management teams.

Fig. 9. Feature Importance based on model

The Assam case study highlights the model’s real-world impact by overlaying predicted flood zones with
hospital locations and road networks, revealing how floods disrupt medical access and delay emergency
response. These insights enable authorities to pre-position ambulances, stock essential medicines, and
plan temporary care centers before disaster strikes.
Unlike static risk maps, the system integrates live weather data, interactive geospatial layers, and dual-
channel alerts (SMS and email), making it a practical tool for healthcare preparedness. Hospital
administrators can input their facility coordinates and receive an instant flood risk score with actionable
guidance. Its modular design supports continuous retraining with new data, ensuring the model stays
relevant under shifting climate conditions and can scale to strengthen disaster response in other
vulnerable regions.

5.1 Comparison with Traditional Approaches

Unlike static GIS maps or rule-based hydrological models that require manual calibration and are
updated infrequently, the proposed ensemble model dynamically adapts to new data[5], [18]. By
combining elevation, rainfall, soil moisture, and derived interaction features, the model can capture real-
time variability in flood risk[6], [7], [10], [11]. This data-driven approach improves both coverage and
accuracy compared to conventional methods, which often rely on a limited number of rain gauges or
historical averages and may fail to reflect changing climate patterns.

A major strength of this work lies in its ability to provide transparent and explainable predictions.
CatBoost’s feature importance analysis revealed that elevation-related and rainfall-driven features were
the most significant drivers of flood prediction, aligning closely with established hydrological principles.
This interpretability fosters trust among stakeholders, as decision-makers can understand why a
particular area was classified as high-risk rather than relying on a “black box” output [19]. Such
transparency is critical for government agencies and local authorities when justifying early interventions
to communities.

5.2 Implications for Disaster Preparedness and Healthcare Resilience

The model’s predictive power goes beyond simply flagging flood-prone zones. By integrating flood
forecasts with hospital locations and road networks, it becomes a decision-support tool for disaster
preparedness. Authorities can use this information to pre-position medical supplies, deploy ambulances
strategically, and plan alternate evacuation routes before a disaster occurs[14], [15], [17]. This kind of
proactive planning can significantly reduce disruption to healthcare services, ensuring that even during
severe floods, critical patients receive timely care.

6. Limitations of the Current Framework

Although the results are encouraging, there are a few important limitations to note. The model was
trained mainly on data from Assam, so its performance might not directly translate to regions with very
different terrain, rainfall patterns, or flood behaviour. For example, areas with snowmelt-driven floods,
coastal storm surges, or desert flash floods would likely require retraining or fine-tuning before
deployment. The system also relies heavily on remote-sensing and weather data, which, while reliable
overall, can sometimes face issues such as delays in availability, missing readings, or lower spatial
resolution during extreme weather. This means that very local, street-level predictions may not always be
accurate.

Another limitation comes from the data balance. Since the dataset contained more flood-labeled cases
than non-flood ones, the model learned to lean slightly toward predicting floods. While this is useful for
early warning systems where missing a flood is more costly than issuing a false alert, it can lead to a few
overpredictions, which, if repeated too often, could cause people to ignore alerts.

Finally, while the model does a good job of capturing general flood patterns, its precision at a micro level
(e.g., small villages or neighborhood clusters) could be improved. This level of granularity is critical for
disaster managers who need to plan logistics and allocate resources to very specific locations. Future
updates could use finer spatial data and more localized ground reports to address this gap
7. Future Work

This work opens several promising directions for future research and development. A key step forward
would be integrating IoT-based water-level sensors and river gauges, providing high-frequency ground
data to complement satellite imagery and capture sudden changes that precede flash floods. Incorporating
crowdsourced information including geotagged images, local rainfall measurements, and road condition
updates could significantly refine local predictions while building community trust. The framework can
be expanded into a multi-hazard platform, modelling related risks such as landslides, road washouts, and
post-flood disease outbreaks, enabling health authorities to anticipate secondary impacts.

Future efforts should also enhance last-mile alert delivery through SMS, IVR calls, and WhatsApp
notifications. On the modelling side, addressing class imbalance using techniques like SMOTE or cost-
sensitive learning, integrating higher-resolution spatial data, and including temporal features such as
rainfall lag would further improve performance. Finally, retraining the model on datasets from diverse
geographies would increase its robustness and scalability, turning it into a globally deployable decision-
support system for disaster preparedness.

8. Conclusion

This work presented an AI–GIS based flood risk prediction framework designed to support disaster
preparedness and healthcare accessibility planning in Assam. By combining feature-engineered
hydrological variables with a stacking ensemble model (Histogram-based Gradient Boosting, Logistic
Regression, and CatBoost meta-learner), the system achieved a balanced predictive performance with a
recall of 0.80 for flood-prone areas. The integration of geospatial hospital and road network data further
demonstrated how predictive outputs can be translated into actionable insights for healthcare planning
during flood events. A cloud-based dashboard was developed to visualize risk zones and make the results
accessible in real time.

8.1 Policy and Practical Implications

The model’s ability to identify high-risk areas before a flood occurs enables proactive disaster
management, including pre-positioning of medical supplies, planning evacuation routes, and allocating
resources efficiently. Such a data-driven approach supports local governments and disaster management
authorities in shifting from reactive to preventive measures, reducing the overall impact of flooding on
public health systems.

8.2 Global Relevance and Scalability

Although this study focuses on Assam, the methodology is generalizable to other flood-prone regions.
By retraining the model on local data, it could be deployed in delta regions, coastal flood zones, and
snowmelt basins globally. The modular design of the pipeline allows for integration of new data sources,
including IoT-based sensors and live weather APIs, making it a scalable solution for climate resilience
planning. This approach contributes to global efforts to strengthen disaster preparedness, healthcare
resilience, and sustainable development, aligning with SDGs 3 (Good Health), 11 (Sustainable Cities),
and 13 (Climate Action).

References

1. Author et al., “Using GIS to Understand Healthcare Access Variations in Flood Situations,” ResearchGate, 2022.
2. Author et al., “Health Impacts of Flooding: A Retrospective Study,” AGU Advances, 2022.

3. NASA, “Shuttle Radar Topography Mission (SRTM),” 2023.

4. GLIMS, “Global Land Ice Measurements from Space,” 2023.

5. “A Critical Review of Flood Risk Assessment in Kerala Post-2018: Methodological Approaches, Gaps, and Future
Directions,” Journal of Hydrology, 2025.

6. “Improving Flood Hazard Susceptibility Assessment by Integrating Hydrodynamic Modeling Remote-Sensing and
Ensemble Machine Learning,” Natural Hazards, 2025.

7. “Flood Hazard and Risk Mapping by Applying an Explainable Machine Learning Framework Using Satellite
Imagery and GIS Data,” Sustainability, 2022.

8. M. Sabera et al., “Examining LightGBM and CatBoost Models for Wadi Flash FloodSusceptibility Prediction,”
Geocarto International, vol. 37, no. 25, 2021.

9. “Comparison of Machine Learning Algorithms for Flood Susceptibility Mapping,” Remote sensing, 2023.

10. “Flood Susceptibility Assessment Using Multi-Tier Feature Selection and Ensemble Boosting Machine Learning
Models,” Water, vol. 17, no. 14, 2025.

11. C. Zeng and D. Bertsimas, “Global Flood Prediction: A Multimodal Machine Learning Approach,” arXiv
preprint, 2023.

12. W. J. et al., “Risk Assessment of Flood Disasters with Multi-Source Data and Its Spatial Differentiation
Characteristics,” Sustainability, vol. 17, 2023.

13. L. M. Sidek et al., “Developing the Flood Risk Matrix for Impact-Based Forecasting in Kelantan River Basin,
Malaysia,” Lecture Notes in Civil Engineering (LNCE), Springer, 2022.

14. Q. Xu et al., “Large-Scale Flood Modeling and Forecasting with FloodCast,” arXiv preprint,2024.

15. M. L. Salvana, “Multi-Hazard Bayesian Hierarchical Model for Damage ˜Prediction,” arXiv preprint, 2025.

16. NIT Warangal and CBIT Hyderabad, “Flood Hazard Assessment of Osmansagar Basin Using GIS, Hydraulic
Modeling and ML,” Scientific Reports, 2025.

17. S. Malik et al., “GIS-Based Statistical Model for the Prediction of Flood Hazard Susceptibility,” Environment,
Development and Sustainability, 2021.

18. Y. S. et al., “Examining the Role of Key Conditioning Factors in Flood Susceptibility Mapping Through
Machine Learning Approaches,” Earth Systems and Environment, 2023.

19. H. Desalegn and A. Mulu, “Flood Vulnerability Assessment Using GIS at Fetam Watershed, Ethiopia,” Heliyon,
2021.

20. J. S. Cabrera and H. S. Lee, “Flood-Prone Area Assessment Using GISBased Multi-Criteria Analysis: Davao
Oriental,” Water, 2019.

21. K. Dandapat and G. K. Panda, “Flood Vulnerability Analysis and Risk Assessment Using AHP & GIS:
Coochbehar, West Bengal,” Natural Hazards, 2019.

22. P. Ramkar and S. M. Yadav, “Flood Risk Index in Data-Scarce River Basins Using the AHP and GIS
Approach,” Natural Hazards, 2021.

23. M. Cerri et al., “Are OpenStreetMap Building Data Useful for Flood Vulnerability Modelling?,” Natural Hazards
and Earth System Sciences, 2021.

24. “CatBoost: Gradient Boosting with Categorical Features,” [Online]. vailable: [Link] Accessed: Feb.
2024.

25. N. Gorelick et al., “Google Earth Engine: Planetary-scale Geospatial Analysis for Ev