Kporha et al. (2026) Comparing daily and 8-day MODIS land surface temperature data for urban heat island assessment using random forest modeling in data-limited regions

Identification

• Journal: Remote Sensing Applications Society and Environment
• Year: 2026
• Date: 2026-01-01
• Authors: Vannesah K. Kporha, Dennis M. Fox, Mostafa Banitalebi, Yacine Bouroubi, Richard Fournier
• DOI: 10.1016/j.rsase.2026.101904

Research Groups

• Université Côte d’Azur, CNRS, AMU, Avignon Université, UMR ESPACE, France
• Université de Sherbrooke, Department of Applied Geomatics, Sherbrooke, Canada

Short Summary

This study evaluates the utility of MODIS daily and 8-day composite data for modeling summer Land Surface Temperature (LST) and Urban Heat Island (UHI) effects across 137 cities in continental France over a 10-year period. It finds that while daily MODIS data offers higher accuracy with meteorological inputs, 8-day composites provide a robust alternative for daytime LST and UHI prediction in regions with limited meteorological data, significantly reducing cloud-related data gaps.

Objective

• To assess the viability of MODIS 8-day composite data as an alternative to daily data in LST and UHI modeling, particularly in settings with limited ground-based weather data.
• To evaluate how the inclusion or exclusion of meteorological data affects predictive performance of Random Forest (RF)-based LST and UHI models.
• To characterize and contrast daytime and nighttime LST and UHI dynamics, offering insights into diurnal heat variation critical for time-sensitive urban heat mitigation strategies.

Study Configuration

• Spatial Scale: 137 cities across continental France, with urban extents ranging from 30 km² to 2193 km². Peri-urban zones were defined by a 15 km buffer around each city, incorporating urban, agricultural, and forest land cover classes.
• Temporal Scale: A 10-year period (2014–2023) focusing on summer months (June to September). Both daily and 8-day composite MODIS LST data were used.

Methodology and Data

• Models used: Random Forest (RF) machine learning model for LST and UHI prediction, supplemented by SHapley Additive exPlanations (SHAP) analysis for model interpretation.
• Data sources:
  • Land Surface Temperature (LST): MODIS Terra Satellite (MOD11A1 and MOD11A2.061 products), 1 km spatial resolution.
  • Land Cover: ESA WorldCover v1.0 (10 m) and Copernicus CORINE Land Cover 2018 (100 m).
  • Weather Data: Météo-France (interpolated to 8 km resolution) for air temperature, relative humidity, and wind speed.
  • Topographic Data: Shuttle Radar Topography Mission (SRTM) Digital Elevation Model (DEM) (30 m spatial resolution) from USGS EarthExplorer.
  • Vegetation Index (NDVI, Proportion of Vegetation): Landsat-8/9 OLI/TIRS Level 2 images (100 m resolution).
  • Proportion of Imperviousness: Derived from the IGN BD-TOPO database (building footprint and road network data) using an empirical framework.

Main Results

• LST Prediction with Meteorological Variables: MODIS daily data achieved higher accuracy (Daytime R² = 0.85, RMSE = 2.39 °C, MAE = 1.80 °C; Nighttime R² = 0.85, RMSE = 1.65 °C, MAE = 1.15 °C) compared to 8-day composites (Daytime R² = 0.75, RMSE = 2.61 °C, MAE = 2.02 °C; Nighttime R² = 0.70, RMSE = 2.10 °C, MAE = 1.56 °C). Air temperature and humidity were the dominant predictors.
• LST Prediction without Meteorological Variables: MODIS 8-day composite data outperformed daily data for daytime LST (R² = 0.57 vs. R² = 0.48) and nighttime LST (R² = 0.52 vs. R² = 0.45). Month and proportion of vegetation were key drivers for daytime, while month, imperviousness, and distance to ocean/sea were important for nighttime.
• UHI Prediction with Meteorological Variables: The 8-day product outperformed daily data for daytime UHI (R² = 0.65 vs. R² = 0.55). Nighttime UHI prediction accuracy was lower for both, with daily data slightly better (R² = 0.35 vs. R² = 0.25). Month, elevation, distance from ocean/sea, and city size were influential for daytime, while city size and imperviousness dominated nighttime.
• UHI Prediction without Meteorological Variables: Model performance remained consistent with the inclusion of meteorological variables. The 8-day product maintained higher accuracy for daytime UHI (R² = 0.65 vs. R² = 0.54 for daily). Nighttime performance declined, with daily data yielding R² = 0.32 compared to R² = 0.25 for 8-day.
• Diurnal Dynamics: Daytime UHI intensity (mean 5.5–6.5 °C) was consistently higher than nighttime UHI (mean around 1.5 °C). Daytime UHI peaked in June, and nighttime UHI peaked in July.
• Data Continuity: The 8-day composite significantly reduced cloud-related data loss from 28.78% (daily) to 0.15%.
• Key Predictor: "Month" consistently emerged as the most influential predictor across nearly all scenarios, serving as a proxy for seasonal climatic and biophysical variations.

Contributions

• Provides a comparative application of Random Forest modeling with dual-temporal satellite-derived LST data to evaluate performance trade-offs, addressing an underexplored area in UHI research.
• Implemented at a broad spatial and temporal scale across 137 cities in continental France, enhancing the generalizability of its findings.
• Offers practical insights into UHI model performance under varying data constraints, supporting more inclusive and resilient urban climate planning, particularly in data-limited regions.
• Demonstrates the strong predictive value of "Month" in both LST and UHI models, capturing seasonal and lagged thermal processes and acting as an effective proxy for climatic variability.

Funding

• Fonds France-Canada pour la Recherche (FFCR)
• French government through the France 2030 investment plan, managed by the National Research Agency (ANR) (reference ANR-15-IDEX-01)
• Université Côte d’Azur Center for High-Performance Computing (OPAL infrastructure)

Citation

@article{Kporha2026Comparing,
  author = {Kporha, Vannesah K. and Fox, Dennis M. and Banitalebi, Mostafa and Bouroubi, Yacine and Fournier, Richard},
  title = {Comparing daily and 8-day MODIS land surface temperature data for urban heat island assessment using random forest modeling in data-limited regions},
  journal = {Remote Sensing Applications Society and Environment},
  year = {2026},
  doi = {10.1016/j.rsase.2026.101904},
  url = {https://doi.org/10.1016/j.rsase.2026.101904}
}