HeatFluxAI
Urban Heat Island Estimation System
Developed an end-to-end geospatial machine learning platform for estimating Urban Heat Island intensity using ERA5 weather data, MODIS satellite imagery, and Google Earth Engine across multiple global cities.
System Architecture
Cross-city geospatial regression architecture derived from the HeatFluxAI repository.
Climate + satellite data
ERA5 weather variables and MODIS land-surface observations are collected for the selected cities.
Geospatial feature pipeline
Raw observations are aligned, cleaned, and transformed into 26 environmental and spatial features.
Regression model comparison
Tree-based and boosting models are trained on approximately 4,500 observations.
City-aware evaluation
Cross-city splits test whether the learned heat relationships generalize beyond a single location.
Interactive heat analysis
The Streamlit interface exposes predictions, drivers, comparisons, and city-level exploration.
Urban heat estimates across 25 cities with traceable model inputs.
Project Overview
HeatFluxAI is an end-to-end geospatial machine learning system designed to estimate Urban Heat Island (UHI) intensity using real-world satellite imagery and climate data. The project combines ERA5 reanalysis weather data, MODIS vegetation and land surface temperature products, and Google Earth Engine to build a scalable pipeline capable of modeling urban heat patterns across 25 globally distributed cities.
The system automatically collects and processes approximately 4,500 observations spanning a 180-day period, generates a scientifically derived UHI target using MODIS Land Surface Temperature composites, and engineers 26 weather, geographic, temporal, satellite, and interaction-based features for predictive modeling.
To ensure robust geographic generalization, the project employs GroupKFold cross-validation and city-aware train-test splitting, preventing data leakage between cities. Up to eight regression algorithms are benchmarked when the optional XGBoost and LightGBM dependencies are available.
Hyperparameter optimization is performed with GridSearchCV, while held-out RMSE, MAE, R-squared, cross-validation error, and skill versus baseline are tracked for city-aware evaluation.
Beyond modeling, HeatFluxAI includes a comprehensive six-tab Streamlit dashboard featuring data exploration, preprocessing insights, model evaluation, interactive heatmaps, feature importance analysis, and live UHI prediction.
Problem
Urban heat patterns vary significantly by geography, vegetation, surface temperature, and weather conditions, making manual UHI estimation slow and difficult to generalize across cities.
Solution
Built a city-aware geospatial ML pipeline using ERA5, MODIS, and Google Earth Engine, then benchmarked multiple regression models with leakage-safe validation.