Chennai Heat Island & Heat Equity Atlas

Hover over any ward to see LST, SC share, and heat burden index

---

Chennai’s urban heat is not uniform — it concentrates in specific corridors and neighbourhoods shaped by land use, surface materials, and decades of underinvestment in green infrastructure. This project builds a complete heat equity atlas for all 155 municipal wards, combining Landsat 8 satellite thermal data, K-means clustering of urban heat zones, and Census 2011 socioeconomic indicators to ask: who bears the greatest burden of extreme heat?

---

Key Findings

Land Surface Temperature across Chennai’s 155 wards ranges from 21°C in waterbodies and forested areas to over 55°C on exposed rooftops and industrial surfaces — a 34-degree spread within a single city boundary. The city-wide ward mean for 2023 is approximately 38°C, with the hottest wards concentrated in the northern industrial belt and dense inner-city fabric.

Wards with higher Scheduled Caste population shares are significantly overrepresented in the upper heat quintiles. OLS regression identifies SC share and NDVI as the two strongest predictors of ward-level LST — with SC share positively associated with heat exposure and NDVI negatively associated, after controlling for deprivation. The Heat Burden Index — a composite of thermal, social, and vegetation variables — reveals that the wards most exposed to heat are also those least equipped to cope with it.

Of the 155 wards analysed, 31 score above 0.6 on the Heat Burden Index, indicating co-occurring high LST, low canopy, high SC share, and high asset deprivation. Only 8 wards score below 0.3, clustered around the Adyar estuary, Guindy National Park, and the IIT Madras campus.

---

How It Works

Land Surface Temperature Retrieval

• Landsat 8 Collection 2 Level-2 imagery acquired via Google Earth Engine for four years: 2015, 2018, 2021, and 2023
• Images filtered to the March–July pre-monsoon season to capture peak heat stress conditions
• LST derived from Band 10 thermal infrared using the standard radiometric calibration: radiance scaling, conversion to brightness temperature, and atmospheric correction via the emissivity-based approach
• Median annual composite computed from all cloud-free scenes to suppress noise from individual overpasses

K-Means Clustering of Urban Heat Zones

• Valid LST pixels (non-masked, non-zero) extracted and clustered into 6 thermally distinct zones using K-Means (k=6, n_init=10, random_state=42)
• Clusters sorted by mean temperature and labelled: Water / Vegetation, Cool Urban, Moderate, Warm Urban, Hot Urban, and Extreme Hotspot
• Each cluster vectorized into a GeoJSON polygon layer for display in the interactive dashboard

Heat Equity Analysis: LST × Census 2011

• Ward-level zonal statistics (mean, max, 90th percentile LST; mean NDVI) computed using rasterstats against the 155-ward boundary layer
• Joined to Census 2011 data: Scheduled Caste share, Scheduled Tribe share, total households, and a deprivation index derived from the proportion of households reporting no listed assets (Census HH-14 table)
• OLS regression with lst_mean_2023 as dependent variable and deprivation_index + sc_share + ndvi_mean_2023 as predictors — SC share and NDVI emerge as significant at p < 0.05

Heat Burden Index

• Composite index constructed from four min-max normalised variables:

  HBI = 0.4 × LST_norm + 0.3 × Deprivation_norm + 0.2 × SC_share_norm + 0.1 × (1 − NDVI_norm)

• Weights reflect the relative contribution of thermal exposure, socioeconomic vulnerability, structural inequality, and green infrastructure deficit
• Index ranges from 0 (lowest burden) to 1 (highest burden) and is mapped as a ward-level choropleth

Tools and Stack

• Google Earth Engine — Landsat 8 image collection, LST computation, NDVI compositing
• Python (rasterio, scikit-learn, rasterstats, geopandas, statsmodels) — clustering, zonal statistics, regression
• Pandas — Census 2011 HH-14 table processing and ward-level aggregation
• Mapbox GL JS + Next.js — interactive dashboard with layer toggles and ward hover popups, deployed on Vercel

---

View the full code and methodology on GitHub