youness-yach/uhi-business-challenge
Domaine:
environment and energygeospatialType de record:
project
Créateur:
you
Hôte:
Satellite-driven Urban Heat Island classification across Rio, Santiago, and Sierra Leone (Hult Business Challenge II, team project).
# Predicting Urban Heat Islands — A Machine Learning Approach Across Three Cities
Satellite-driven classification of Urban Heat Island (UHI) intensity in Rio de
Janeiro and Santiago, with a combined model transferred to predict UHI risk in
Freetown, Sierra Leone.
**Team project** — Hult International Business School, Business Challenge II.
Team: Carolina Trovisco, Filippo Beni, João Ponte, Mickias Ambaye,
**Youness Yachruti**, Yousra Sajjad. My focus on this team was model
research: testing and tuning multiple tree-based and gradient-boosting models
across all three locations to improve prediction accuracy. See a teammate's
complementary extraction-pipeline repo:
Mickias-Ambaye/uhi-pipe.
## Overview
Urban Heat Islands are urban areas that run warmer than their surroundings —
sometimes by 10°C+ locally — driven by dense building layouts, impervious
surface heat absorption, and waste heat from industry and transport. The
challenge: build a model that predicts UHI intensity from satellite data
alone, and test whether a model trained on two cities can transfer to predict
a third city it has never seen.
We extracted Sentinel-2 spectral indices and Landsat-8 thermal/elevation data
for Rio de Janeiro and Santiago (50,150 combined data points), engineered
interaction features (e.g. LST × NDVI, elevation × LST), trained per-city
classifiers, then tested whether a model combining both cities' signal could
predict UHI intensity in Freetown, Sierra Leone — a city with no training
labels of its own.
## Key results
All figures below are held-out classification performance (F1 / precision /
recall), taken directly from the team's final report — not backtested or
live figures, this is a classification task, not a trading strategy.
- **Rio de Janeiro:** F1 = 0.959 (Precision 0.960, Recall 0.959) — best model:
XGBoost. Rio's heat signal is dominated by raw thermal response and building
morphology; intense year-round solar exposure combined with high shares of
concrete/asphalt …