HeatWise AI

AI-powered geospatial platform that predicts urban heat hotspots.

Comment

HeatWise AI

Physics-guided geospatial intelligence for urban heat mitigation

HeatWise AI transforms urban heat observations into explainable, location-specific and monitorable cooling actions.

Inspiration

Urban heat is not distributed equally.

Two neighbourhoods in the same city can experience significantly different temperatures because of vegetation, impervious surfaces, building density, surface albedo, wind, humidity, pollution and population concentration.

Most thermal dashboards can show where high temperatures occur. However, municipal authorities must also answer:

  • Why is this location overheating?
  • Which people and infrastructure are exposed?
  • Which cooling strategy is appropriate for this location?
  • What improvement does the model predict?
  • How can we verify whether an implemented intervention worked?

We built HeatWise AI to connect hotspot detection, driver analysis, cooling-strategy comparison, spatial feasibility, implementation monitoring and municipal reporting in one platform.

Where is the heat? Why is it happening? What can be implemented? How will the result be verified?

What It Does

HeatWise AI provides an end-to-end decision-support workflow for urban heat management.

Stage HeatWise AI capability Decision supported
Observe City, season and hotspot-level thermal views Locate areas requiring attention
Predict Land-surface temperature regression Estimate local thermal intensity
Classify Hotspot probability and risk classification Prioritise high-risk locations
Diagnose Explainable environmental driver analysis Understand likely causes
Compare Baseline, no-action and intervention scenarios Evaluate cooling alternatives
Screen Buildings, roads, parks and water assets Assess spatial feasibility
Monitor Before-and-after and matched-control records Measure implementation outcomes
Report Location-specific municipal action briefs Communicate decisions and evidence

Core Features

National Observation Grid

The main dashboard provides a national overview of supported Indian cities, including:

  • City and model-area selection
  • Seasonal exploration
  • Peak land-surface temperature
  • Mean Urban Heat Island anomaly
  • Modelled hotspot sectors
  • Population exposure
  • Cooling opportunity indicators
  • Interactive thermal mapping

Heat Explorer

The Heat Explorer allows users to select a city and hotspot, inspect local thermal intensity, view risk classification, compare seasonal conditions and explore surrounding spatial context.

Driver Analysis

The Driver Analysis module explains which features most influenced the model for the selected location.

Potential heat drivers include:

  • Vegetation and canopy cover
  • Impervious surface fraction
  • Surface albedo
  • Building density and urban form
  • Wind and atmospheric conditions
  • Seasonal variables
  • Pollution and aerosol context
  • Population exposure

Feature importance is presented as a model association, not automatic proof of causation.

Scenario Lab

The Scenario Lab allows users to combine and compare cooling strategies such as:

  • Tree and green-cover expansion
  • Cool-roof coatings
  • Shade and green infrastructure
  • Permeable or reflective pavement

The system compares:

  • Baseline modelled temperature
  • Target-date no-action temperature
  • Intervention scenario temperature
  • Cooling relative to no action
  • Hotspot-risk transition
  • Physical feature assumptions

Scenario results are modelled counterfactuals, not measured post-implementation outcomes.

Actionability Engine

The Actionability Engine connects hotspot intelligence with nearby OpenStreetMap assets, including:

  • Building footprints
  • Named street corridors
  • Parks and open spaces
  • Waterways and water bodies
  • Other available mapped infrastructure

This provides local context before authorities conduct engineering and field feasibility surveys.

Implementation Monitoring

Municipal users can record:

  • Treated-location temperature before implementation
  • Treated-location temperature after implementation
  • Matched-control temperature before implementation
  • Matched-control temperature after implementation
  • Intervention type and coverage
  • Observation date and project status

The adjusted change can be estimated using a difference-in-differences comparison:

$$

\Delta_{\text{adjusted}}

(T_{\text{treated,after}}-T_{\text{treated,before}})

(T_{\text{control,after}}-T_{\text{control,before}}) $$

This controls for broader temperature changes affecting both locations. It is more defensible than a simple before-and-after comparison, although it does not prove causality by itself.

Municipal Reports

The platform creates a location-specific action brief containing:

  • Temperature and hotspot-risk summary
  • Population exposure
  • Key heat drivers
  • Environmental parameters
  • Proposed cooling portfolio
  • Spatial asset context
  • Monitoring requirements
  • Model assumptions and limitations

Contextual AI Copilot

The AI assistant receives the currently selected city, hotspot, season and model context.

It is designed to:

  • Answer only the requested question
  • Provide concise, location-aware responses
  • Explain model outputs in understandable language
  • Avoid inventing unavailable observations
  • Separate predictions from verified facts
  • Format technical content clearly

How We Built It

HeatWise AI combines geospatial data, machine learning, physics-guided reasoning and an interactive web application.

Model 1: XGBoost LST Regressor

The regression model estimates land-surface temperature using environmental, spatial and seasonal inputs.

Example inputs include:

  • Vegetation indices
  • Canopy-cover fraction
  • Impervious-surface fraction
  • Surface albedo
  • Land-use and land-cover information
  • Building and urban-form indicators
  • Air temperature
  • Relative humidity
  • Wind speed
  • Pollution or aerosol context
  • Seasonal encodings
  • Location-specific features

The output is a predicted land-surface temperature in degrees Celsius.

XGBoost was selected because it performs strongly on structured tabular data, captures nonlinear relationships, models interactions between environmental variables and provides efficient inference.

Model 2: XGBoost Hotspot Classifier

The classification model uses environmental features and modelled temperature to produce:

  • Hotspot probability
  • Binary hotspot decision
  • Operational risk category

The decision threshold is selected according to the required trade-off between precision and recall instead of automatically using a threshold of 0.5.

Model Performance

Temperature regression

Metric Result
Mean Absolute Error 0.98°C
Root Mean Squared Error 1.23°C
Coefficient of Determination 0.977

Hotspot classification

Metric Result
Accuracy 93.07%
Precision 79.95%
Recall 41.06%
F1 score 54.25%
ROC-AUC 95.16%
PR-AUC 73.07%
Decision threshold 95.25%

The high decision threshold prioritises precision and reduces false hotspot alerts. This produces a recall trade-off that must be considered when using the model for screening.

These metrics apply to the held-out evaluation distribution and do not guarantee identical performance across every city, season or sensor.

Physics-Guided Reasoning

The model design follows the urban surface-energy balance:

$$ R_n = H + LE + G $$

Where:

  • $R_n$ is net radiation
  • $H$ is sensible heat transferred to the atmosphere
  • $LE$ is latent heat associated with evapotranspiration
  • $G$ is ground and surface heat storage

This provides the physical interpretation behind the model features:

  • Vegetation can increase latent cooling through evapotranspiration.
  • Reflective surfaces can reduce absorbed solar radiation.
  • Concrete and asphalt can increase heat storage.
  • Wind and urban geometry can affect ventilation.
  • Moisture availability affects the balance between sensible and latent heat.

Physics guides feature selection, scenario constraints and interpretation. It does not replace real physical measurements or independent validation.

Experimental PINN Cross-Check

We also developed an experimental Physics-Informed Neural Network as a secondary temperature cross-check.

The PINN includes a physics loss that penalises predictions inconsistent with a partial surface-energy balance.

The PINN is not used as the operational hotspot classifier. XGBoost remains the primary model because it performs better on the available structured dataset and supports efficient, explainable inference.

Data Strategy

Data category Candidate sources Purpose
Thermal observations Landsat 8/9, ECOSTRESS and MODIS LST targets and temporal validation
Vegetation and land cover Sentinel-2 and Landsat NDVI, canopy and surface-cover features
Meteorological conditions ERA5 and local stations Air temperature, humidity and wind
Air quality CPCB observations Pollution and aerosol context
Urban morphology OpenStreetMap and GHSL Buildings, roads and built-up structure
Population Census or validated gridded datasets Exposure and prioritisation

For operational use, observations must be aligned by geographic location, acquisition time, coordinate system, spatial resolution, quality flags, sensor characteristics and seasonal context.

Technology Stack

Layer Technologies
Frontend Next.js, React and TypeScript
Styling Tailwind CSS
Mapping MapTiler, MapLibre and OpenStreetMap
Machine learning XGBoost regression and classification
Physics cross-check Experimental PINN
Explainability Feature importance and driver analysis
AI assistant Groq-powered contextual assistant
Automation n8n webhook integration
Reporting Browser-generated municipal reports and PDFs
Deployment Vercel

Challenges We Faced

Turning Predictions into Decisions

Our initial implementation focused heavily on temperature and risk predictions. We realised that a temperature value alone does not tell a municipality what action to take.

We redesigned the platform around the complete workflow of detection, diagnosis, comparison, implementation and monitoring.

Separating Simulation from Evidence

A predicted temperature reduction is not the same as an observed real-world improvement.

We explicitly separated:

  • Available observations
  • Model predictions
  • Counterfactual simulations
  • Post-implementation measurements

This prevents simulated values from being presented as verified outcomes.

Handling Incomplete Data

Not every environmental measurement is available for every location.

Instead of silently presenting missing values as ground truth, the system distinguishes available observations from model-filled features and communicates these limitations to the user.

Making AI Understandable

Raw feature names and model outputs are difficult for non-technical users.

We translated model features into understandable urban-heat drivers while retaining access to the underlying technical parameters.

Evaluating Real Interventions

A simple before-and-after comparison can be misleading because both dates may have different weather conditions.

We added matched-control monitoring to help separate intervention-related change from city-wide temperature variation.

Building a Deployable System

We integrated model inference, interactive mapping, intervention comparison, reporting and the AI assistant into a browser-accessible application that can be deployed on Vercel.

What We Learned

Building HeatWise AI taught us that an effective AI product requires more than a high-performing model.

A reliable AI product also needs:

  • Transparent assumptions
  • Understandable explanations
  • Confidence-aware decisions
  • Safe fallbacks
  • Location-specific context
  • Post-implementation monitoring
  • Clear separation between observed, predicted and simulated data

Most importantly, we learned that responsible AI sometimes means saying, “This requires local verification,” instead of presenting an uncertain prediction as fact.

What Makes HeatWise AI Different

  • Goes beyond heat-map visualisation
  • Combines temperature regression with hotspot classification
  • Explains why a location is predicted to be hot
  • Connects model outputs with surface-energy physics
  • Supports multi-strategy cooling portfolios
  • Compares interventions against a no-action baseline
  • Uses local spatial assets for feasibility context
  • Separates simulated impact from measured impact
  • Supports matched-control monitoring
  • Generates location-specific municipal reports
  • Includes a contextual AI assistant
  • Runs as a deployable web application

Current Limitations

HeatWise AI is a decision-support prototype.

Operational municipal deployment would require:

  • Aligned and quality-controlled satellite scenes
  • Verified meteorological and ground observations
  • Local municipal GIS data
  • Engineering and field feasibility surveys
  • Independent validation across cities and seasons
  • Prediction uncertainty estimates
  • Security and access-control review
  • Approval from relevant authorities

Model predictions should not be interpreted as guaranteed real-world outcomes.

What’s Next

Our future roadmap includes:

  • Reproducible ingestion of satellite and ERA5 data
  • Automated cloud masking and temporal alignment
  • Validation across additional cities and seasons
  • Prediction uncertainty intervals
  • Authenticated municipal project portfolios
  • Intervention implementation tracking
  • Satellite-based post-implementation verification
  • Causal modelling after sufficient audited observations
  • Model cards and data-lineage documentation
  • Municipal GIS integration
  • Continuous model-drift monitoring

Try It

Built With

Share this project:

Updates

Submission history