Urban Resilience Prototype (SDG-13)
The AI Flood Intelligence System is a near real-time urban flood and drain overflow alert prototype designed for low-lying areas.
The system integrates drainage health modeling, zone-level aggregation, and machine learning-based probabilistic prediction to generate localized flood risk alerts.
Unlike rainfall-only alert systems, this project models how the drainage network behaves under stress and predicts the probability of flooding within the next one hour.
The system is modular, interpretable, and designed to support future municipal deployment.
Drain-level sensor data
Zone-level aggregation
Physics-informed modeling
Machine Learning-based risk prediction
Interpretable alert generation
The goal is to enable early, explainable, and localized flood warnings that can support:
Municipal authorities
Urban planners
Citizen reporting systems
Climate resilience initiatives
Urban flooding in low-lying regions is often caused by:
Drain blockages
Reduced drainage efficiency
Cascading drain failures
Persistent overload conditions
Rainfall alone does not reliably indicate flooding risk. The drainage system’s health and degradation dynamics must also be considered.
This project addresses that gap by combining physics-informed features with a probabilistic AI model.
Sensors / Rainfall / Reports
→ Drain-Level Health (DCI)
→ Zone-Level Aggregation
→ Degradation and Flood-Time Dynamics
→ AI Probability Model
→ Risk Interpretation Layer
→ Localized Alerts and Dashboard
Raw drain sensor readings (water level, flow efficiency, blockage indicators) are normalized into a Drainage Condition Index:
DCI ∈ [0, 1]
Where:
1 represents a fully healthy drain
0 represents a failed or blocked drain
The index is smoothed over time to obtain:
DCI_eff
This reduces noise and captures effective drain health.
Drains are grouped into zones. From drain-level DCI values, the following zone-level features are computed:
mean_DCI_eff
min_DCI_eff
pct_low_DCI
mean_degradation_rate
min_degradation_rate
degradation_volatility
zone_flood_time_now
zone_flood_slope
zone_flood_volatility
time_under_2hr_min
drains_TTF_lt_1hr
drains_TTF_lt_30min
These features describe:
Current system health
Rate of degradation
Instability and volatility
Persistence under risk conditions
Cascading failure clustering
Model Type: Logistic Regression
The model predicts:
Probability of flooding within the next 1 hour
Output:
AI_flood_risk_score ∈ [0, 1]
The choice of logistic regression ensures:
Interpretability
Stable probabilistic output
Real-time compatibility
Low computational overhead
Model evaluation:
ROC-AUC ≈ 0.78
Strong early-warning separation between flood and non-flood states
Predicted probability is converted into actionable risk levels:
Probability < 0.30 → LOW
0.30 – 0.50 → WATCH
0.50 – 0.70 → WARNING
≥ 0.70 → CRITICAL
This allows gradual escalation instead of binary alerts.
The Streamlit dashboard provides:
Dataset snapshot
AI flood probability viewer
Risk level visualization
ROC curve and performance summary
Probability distribution plots
Feature correlation heatmap
System interpretation text
The dashboard is intended for demonstration, evaluation, and monitoring.
Drain_Intelligence/
├── processing/
│ ├── compute_dci.py
│ ├── compute_degradation.py
│ ├── compute_time_to_failure.py
│ └── compute_zone_basin_flood_time.py
│
├── output/
│ ├── ai_zone_training_dataset.csv
│ ├── ai_zone_risk_score.csv
│ └── zone_flood_prediction.csv
│
├── AI_Flood/
│ ├── training/
│ │ └── train_zone_flood_model.py
│ │
│ ├── inference/
│ │ ├── predict_zone_flood_risk.py
│ │ └── risk_interpretation.py
│ │
│ ├── models/
│ │ └── zone_flood_logistic.pkl
│ │
│ └── analysis/
│ └── flood_model_analysis.ipynb
│
├── dashboard/
│ └── app.py
│
└── README.md
README.md
Train the model:
python AI_Flood/training/train_zone_flood_model.py
Run the dashboard:
streamlit run dashboard/app.py
Run inference programmatically:
from AI_Flood.inference.predict_zone_flood_risk import predict_zone_flood_probability
This system contributes to climate resilience by:
Enabling early flood warnings
Modeling drainage infrastructure vulnerability
Supporting localized disaster mitigation
Providing interpretable decision support
The architecture is modular and can be extended for ward-level or city-level deployment.
Physics-informed feature engineering
Zone-level abstraction
Interpretable probabilistic AI
Real-time safe computation
Modular separation of training and inference
Explainable risk thresholds
Currently operates on structured or simulated sensor datasets
Rainfall API integration is not yet live
Citizen reporting integration is conceptual
Geospatial visualization layer is limited
Real-Time Sensor Integration
MQTT / IoT streaming pipeline
Live DCI updates
Edge-based preprocessing
Rainfall API Integration
OpenWeather or IMD integration
Rainfall-trigger adjustments in zone features
Citizen Reporting Module
WhatsApp or SMS integration
Report-density feature integration
Crowd-sourced validation
Geospatial Mapping
Zone heatmaps
GIS-based overlays
Vulnerability visualization
Advanced Modeling
Random Forest and XGBoost comparison
Time-series models
Adaptive threshold tuning
Alert Automation
SMS push alerts
Escalation pipeline for municipal authorities
Municipal Deployment Path
Ward-level scaling
API service layer
Cloud deployment architecture
Urban flood prediction should not rely solely on rainfall intensity. By modeling drainage health, degradation dynamics, and cascading failures, the AI Flood Intelligence System provides early, explainable, and localized flood risk alerts suitable for resilient urban infrastructure planning.