PROTEUS Dashboard | OneMap Integration

(i)

Indicative screening tool only. Results should not be used as the sole basis for community relocation or service closure decisions.

Land Area Inundated by Scenario

Population at Risk by Scenario

Top 10 Most Vulnerable Islands , Current Scenario

Flood Exposure Map

0-25%

25-50%

50-75%

75-100%

Show all islands (including uninhabited)

Island Vulnerability Rankings

Rank	Island	Atoll	Area (km )	Population	% Land	Pop. at Risk	VI Score

Historical Sea Level + ML Projections to 2100

Mal tide gauge · UHSLC Station 108 Historical: Oct 1989 - Feb 2026 (423 months) Projections: 2026 - 2100 Y-axis: anomaly relative to Oct 1989 baseline

Sea level at Mal has risen +27.98 cm since October 1989 , a rate of 4.4 mm/year. The Ensemble model (LSTM + Hybrid LSTM, weighted 0.6/0.4) achieves the best test performance (RMSE = 4.37 cm, R = 0.535). Dashed horizontal lines show IPCC AR6 thresholds: +50 cm (SSP1-2.6) and +100 cm (SSP5-8.5) by 2100. The shaded band shows Prophet's 80% confidence interval.

Model Accuracy on Held-Out Test Period

Test set: Nov 2018 - Feb 2026 (85 months) Not seen during training , true out-of-sample evaluation Includes 2019-2020 extreme positive IOD event

The test period includes the 2019-2020 extreme positive Indian Ocean Dipole (IOD) event (DMI peak = 0.96 degC), which pushed sea levels at Mal significantly above the historical trend. ARIMA and Prophet fail here (R < -1) because they can only model linear trends. The LSTM and Hybrid LSTM successfully track the IOD-driven anomaly through their recurrent memory.

ML Model Performance , All 8 Models

RMSE = Root Mean Square Error in cm (lower = better) R / NSE = Variance explained (1.0 = perfect, below 0 = worse than predicting the average) Skill Score = improvement over persistence baseline (above 0 = better than last month's value) F1 = threshold exceedance accuracy (sea level > +8 cm anomaly) Best model highlighted

F1 Score & Skill Score , Visual Comparison

F1 Score: ability to detect dangerous sea level months (threshold > +8 cm anomaly) Skill Score: improvement over a naive persistence forecast

Methodology & Data Sources

Phase 1: GIS Flood Inundation

Static bathtub model applied to SRTM 30m DEM data across all 181 inhabited Maldivian islands. Island elevation profiles are parameterised by mean elevation, maximum elevation, and fraction of land below 1m. Flood inundation is computed for 5 IPCC AR6 scenarios including a compound storm surge scenario.

Phase 2: Machine Learning Sea Level Prediction

Eight models trained and evaluated on 38 years of UHSLC tide gauge data (Station 108, Mal , 1988-2026). The Hybrid LSTM incorporates real NOAA monthly ONI and DMI climate indices as exogenous predictors. An 80/20 chronological train/test split was used; all models evaluated on identical held-out test set.

Model	RMSE (cm)	R	Framework
Ensemble (LSTM+Hybrid)	4.37	0.535	Pure JS Best
Hybrid LSTM (MSL+ONI+DMI)	4.38	0.534	Pure JS
LSTM (univariate)	4.58	0.490	Pure JS
Random Forest	5.61	0.235	Python/sklearn
Gradient Boosting	5.98	0.130	Python/sklearn
XGBoost	5.99	0.128	Python/XGBoost
ARIMA(1,1,0)	11.90	-2.44	Pure JS
Prophet (additive)	10.18	-1.52	Pure JS

Phase 3: Dashboard

Node.js/Express server with Chart.js and OneMap API frontend. All outputs pre-computed as JSON files in outputs/. Deploy with node server.js.

Data Sources

Dataset	Source	File
Monthly Sea Level	UHSLC Station 108	data/male_sealevel.csv
ONI Index	NOAA CPC	data/oni.csv
DMI Index	NOAA PSL	data/dmi.csv
Island registry	OneMap ArcGIS FeatureServer	data/islands.json
Elevation DEM	NASA SRTM	Parameterised

Run the Analysis

npm install && pip install -r requirements.txt
npm run analyse           # LSTM models
npm run analyse:advanced  # RF + GB + XGBoost
node server.js            # Dashboard at http://localhost:3000

Mahmood, M.Z. (2026) Predicting Sea Level Rise Impact on Maldivian Islands: A Machine Learning Approach for Climate Adaptation Planning. MSc Dissertation, Villa College / University of the West of England, Bristol. Module UFCF9Y-60-M.

International Context

How does this study compare to similar work in other countries? Four comparison groups are presented below.

Pacific Small Island States

Storlazzi et al. (2018) modelled wave-driven flooding for Pacific atoll islands using a physics-based Boussinesq wave model. That study achieves higher physical accuracy but covers three islands. This framework covers all 181 inhabited Maldivian islands using a simpler bathtub model , a deliberate trade-off of precision for national coverage. The 38-year Male tide gauge record is also substantially longer than most Pacific records (typically 20 to 25 years), giving the ML component a meaningful training data advantage.

For Kiribati, Tuvalu and the Marshall Islands most national assessments use CMIP6 projections downscaled to individual islands rather than ML forecasting from local station records. The integrated ML-plus-GIS approach used here has not been applied at national scale to any Pacific SIDS.

Bangladesh FFWC , Closest Operational Equivalent

The Bangladesh Flood Forecasting and Warning Centre (FFWC) has deployed LSTM-based models for real-time tidal and river level forecasting (Kabir et al., 2020). Their published RMSE for tidal station predictions is 3 to 6 cm , directly comparable to the Ensemble RMSE of 4.37 cm achieved here. The key difference is that the FFWC system runs on live government servers with real sensor feeds and issues operational alerts. This dashboard uses pre-computed outputs. The implication is that the Maldives has the technical building blocks for an equivalent operational system.

Netherlands and Australia

The Netherlands uses Deltares Delft3D hydrodynamic models for coastal flood simulation, incorporating wave dynamics, storm surge and bathymetric data at fine resolution. This achieves higher physical accuracy but requires national LiDAR coverage, high-resolution bathymetric surveys and substantial computing infrastructure , none of which exist at national scale for the Maldives.

Australia's National Coastal Vulnerability Assessment (DCCEE, 2011) is the closest structural parallel: a national-scale vulnerability index producing ranked coastal segments. Australia uses a six-variable scoring system compared to the three-variable VI used here. This framework adds the ML forecasting component and the live OneMap API integration, neither of which appear in the Australian assessment.

Key Differentiator

Across all comparison groups, the unique contribution of this study is the combination of three things that have not previously been integrated: ML sea level forecasting from a long local tide gauge record, national-scale GIS vulnerability ranking for all inhabited islands, and live government data API integration via OneMap. Each component individually has precedents. Their combination in a SIDS context does not.

How This Study Compares

Study / System	Country	Approach	Islands / Scale	ML Component
This study	Maldives	Bathtub GIS + LSTM Ensemble	181 islands national	8 models, RMSE 4.37 cm
Storlazzi et al. (2018)	Pacific SIDS	Boussinesq wave model	3 islands	None
Kabir et al. / FFWC	Bangladesh	LSTM tidal forecasting	Operational stations	RMSE 3-6 cm
Deltares / Delft3D	Netherlands	Hydrodynamic model	National	None
DCCEE (2011)	Australia	6-variable VI assessment	National coastline	None
Amores et al. (2021)	Maldives	Wave-driven inundation	Selected islands	None

Glossary

Definitions for all technical terms used in this dashboard and the accompanying dissertation.

Model Performance Metrics

Term	Full Name	What it means
RMSE	Root Mean Square Error	Average prediction error in centimetres. Lower is better. The Ensemble achieves 4.37 cm on the 85-month test set.
MAE	Mean Absolute Error	Average absolute error in centimetres. Less sensitive to large outliers than RMSE.
R^2 / NSE	R-squared / Nash-Sutcliffe Efficiency	How much of the observed variance the model explains. 1.0 = perfect, 0 = no better than predicting the mean, below 0 = worse than the mean baseline.
MAPE	Mean Absolute Percentage Error	Error as a percentage of the mean absolute anomaly (5.11 cm). Lower is better.
Skill Score	Skill Score vs persistence	Improvement over a naive forecast that simply predicts last month's observed value. Above 0 means better than persistence. The Ensemble scores 0.07 (7% better than persistence baseline).
F1 Score	F1 Score	Accuracy of detecting months when sea level exceeds the +8 cm anomaly threshold. Balances precision (avoiding false alarms) and recall (catching real events). 1.0 = perfect, 0 = no skill.
Precision	Precision	Of all the months the model flagged as dangerous, what fraction actually were. High precision = few false alarms.
Recall	Recall	Of all the genuinely dangerous months, what fraction the model correctly identified. High recall = few missed events.

Machine Learning Models

Term	Full Name	Description
LSTM	Long Short-Term Memory	A recurrent neural network architecture that can retain information over long time periods. Used here in univariate form with only sea level as input.
Hybrid LSTM	Hybrid LSTM (MSL + ONI + DMI)	LSTM model using three input features per timestep: sea level, ONI, and DMI. The climate indices provide a 3-6 month lead time signal.
Ensemble	Ensemble (0.6 LSTM + 0.4 Hybrid)	Weighted average of LSTM and Hybrid LSTM predictions. Best-performing model in this study (RMSE = 4.37 cm, F1 = 0.59).
ARIMA	AutoRegressive Integrated Moving Average	Classical statistical time series model. Used as a baseline. Poor performance here (RMSE = 11.90 cm) because it cannot capture non-linear inter-annual variability from ENSO and IOD.
Prophet	Prophet (additive decomposition)	Open-source forecasting model by Meta. Decomposes series into trend, seasonality, and residuals. Better than ARIMA but still limited on inter-annual signals (RMSE = 10.18 cm).
RF	Random Forest	Ensemble of decision trees. Implemented with 500 trees and lag features. Intermediate performance (RMSE = 5.61 cm).
GB	Gradient Boosting	Sequential ensemble of shallow trees. Intermediate performance (RMSE = 5.98 cm).
XGB	XGBoost	Optimised gradient boosting implementation. Similar to GB (RMSE = 5.99 cm).

Climate and Oceanographic Terms

Term	Full Name	Description
ONI	Oceanic Nino Index	Monthly index measuring El Nino and La Nina strength in the equatorial Pacific. Affects Indian Ocean sea levels via atmospheric teleconnections.
DMI	Dipole Mode Index	Monthly index measuring the Indian Ocean Dipole. A positive IOD event in 2019-2020 (DMI peak = 0.964 deg C, second strongest on record) produced the largest sea level anomalies in the 85-month test period.
IOD	Indian Ocean Dipole	A coupled ocean-atmosphere pattern in the Indian Ocean. Positive IOD events drive sea level rise across the central and northern Indian Ocean via coastal Kelvin wave propagation.
ENSO	El Nino-Southern Oscillation	The dominant Pacific climate mode. Affects Indian Ocean sea levels and is captured by the ONI index used in the Hybrid LSTM.
MSL	Mean Sea Level	The average level of the ocean surface, used as the baseline for sea level rise measurements.
SLR	Sea Level Rise	The increase in mean sea level over time due to ocean thermal expansion and ice melt.
SSP	Shared Socioeconomic Pathway	IPCC AR6 climate scenarios. SSP1-2.6 assumes strong emissions cuts; SSP5-8.5 assumes continued high emissions and projects up to +1.0m by 2100.
IPCC AR6	Sixth Assessment Report	The authoritative international scientific assessment of climate change, published 2021 by the Intergovernmental Panel on Climate Change.

Data and Methods

Term	Full Name	Description
VI	Vulnerability Index	Composite score (0 to 1) for each island: VI = 0.5 x land flooded + 0.3 x normalised population + 0.2 x (1/island area). Higher score = more vulnerable. All sub-indicators normalised to [0,1] before weighting.
GIS	Geographic Information System	Tools and methods for spatial data analysis. Used here for flood inundation mapping of all 181 inhabited islands.
Bathtub model	Bathtub inundation model	Simple flood model that classifies all land below a sea level threshold as permanently inundated. Standard approach for national-scale screening where LiDAR data is unavailable.
SRTM	Shuttle Radar Topography Mission (+/-5--10m vertical accuracy)	NASA satellite elevation dataset at 30m resolution. Used to estimate island elevation in the absence of national LiDAR coverage for the Maldives.
OneMap	OneMap Maldives	The official national mapping platform of the Maldives government. Provides island boundary polygons, names, and atoll classification via an ArcGIS FeatureServer API.
UHSLC	University of Hawaii Sea Level Center	Maintains the 38-year tide gauge record at Male, Maldives (Station 108) used to train and test all ML models in this study.
ArcGIS	ArcGIS JavaScript API	Esri mapping platform used to display real OneMap island boundaries, flood bubble overlays, and reef/lagoon layers in the dashboard map.