Tyagi et al. (2025) Machine Learning-Based Forecasting of Wet-Bulb Temperature and Two-Decade Urban Climate Shifts

Identification

• Journal: Earth Systems and Environment
• Year: 2025
• Date: 2025-12-08
• Authors: Aryan Tyagi, Sagar Tomar, A. Raut, Kishor S. Kulkarni, Shilpa Sharma, Tarig Ali, Jerry Wayne Nave, Rabin Chakrabortty
• DOI: 10.1007/s41748-025-00911-9

Research Groups

• Netaji Subhas University of Technology, Delhi, India
• CSIR- Central Building Research Institute, Roorkee, India
• CSIR-Central Electronics Engineering Research Institute, Pilani, India
• Academy of Scientific and Innovative Research, Ghaziabad, India
• Department of Civil Engineering, American University of Sharjah, Sharjah, United Arab Emirates
• Department of Built Environment, College of Science and Technology, North Carolina A&T State University, Greensboro, NC, US

Short Summary

This study investigates the spatio-temporal dynamics of Wet-Bulb Temperature (WBT) and Land Surface Temperature (LST) in Delhi, India, from 2005 to 2024, using satellite data and an LSTM model to forecast future WBT trends. The research reveals a significant rise in WBT, projected to exceed 35 °C by 2030 in densely built-up areas, highlighting the urgent need for climate adaptation strategies.

Objective

• To investigate the spatio-temporal dynamics of Wet-Bulb Temperature (WBT) and Land Surface Temperature (LST) in Delhi, India, from 2005 to 2024.
• To project future WBT trends using a Long Short-Term Memory (LSTM) model.
• To integrate temporal prediction and spatial correlation analysis for WBT and LST using satellite-derived data and machine learning for urban resilience planning and early warning systems.

Study Configuration

• Spatial Scale: Delhi, India (a metropolitan region), focusing on urban centers and surrounding areas.
• Temporal Scale: Historical data analysis from 2005 to 2024, with future projections extending to 2030 (and 2031 for WBT).

Methodology and Data

• Models used: Long Short-Term Memory (LSTM), Deep Neural Network (DNN), Random Forest (RF), Decision Tree (DT), Seasonal Autoregressive Integrated Moving Average (SARIMA), Gated Recurrent Unit (GRU), Naïve Bayes Regression (as a baseline).
• Data sources: NASA POWER (for Wet-Bulb Temperature and meteorological variables like air temperature, relative humidity, atmospheric pressure, wind speed, and solar radiation), LANDSAT imagery (for Land Surface Temperature and urban expansion data), India Meteorological Department (IMD) station data (for spot comparisons/validation of temperature and humidity).

Main Results

• Wet-Bulb Temperature (WBT) is projected to consistently rise, exceeding 35 °C during extreme heat events by 2030, particularly in densely built-up zones of Delhi.
• A strong positive correlation was observed between urbanization (increased built-up areas and diminished vegetation) and elevated WBT and Land Surface Temperature (LST) levels, intensifying the Urban Heat Island (UHI) effect.
• Pearson correlation analysis of May LST values revealed strong associations between recent years (e.g., r = 0.74 for 2023–2024, r = 0.71 for 2015–2023), indicating persistent warming trends and prolonged thermal stress.
• Weaker correlations between early and recent years (e.g., r = 0.20 for 2005–2021) highlight a substantial shift in LST patterns over the past two decades.
• The LSTM model demonstrated superior predictive accuracy for WBT, achieving an R² value of 0.9645 (or 0.98 in a comparative assessment), with a Mean Absolute Error of 1.07 °C and a Root Mean Squared Error of 1.36 °C, outperforming other tested models.
• Spatial measurements confirmed that populated city centers recorded elevated WBT and LST data, reflecting the impact of urbanization and damaged vegetated areas.

Contributions

• Integration of high-resolution satellite-based climate data (NASA POWER, LANDSAT LST) with advanced machine learning (LSTM) for comprehensive spatio-temporal WBT forecasting and urban climate shift analysis.
• Development of a robust framework for understanding urban heat stress patterns, enabling data-driven climate adaptation and urban planning strategies.
• Bridging a significant gap in the literature by combining spatial LST data with temporal WBT forecasts to identify both where and when heat stress peaks occur, specifically in the context of rapidly urbanizing Indian cities.
• Providing critical insights for policymakers to formulate evidence-based heat mitigation strategies, including AI-based early warning systems, green infrastructure development, and resilient urban policies.
• Application of Pearson correlation to measure spatial LST consistency across two decades (2005–2024), offering a valuable diagnostic tool to identify heat trend persistence and abnormal climatic events.

Funding

• CSIR-HRDG (fellowship support)
• CSIR-CBRI, Roorkee (computational facilities)

Citation

@article{Tyagi2025Machine,
  author = {Tyagi, Aryan and Tomar, Sagar and Raut, A. and Kulkarni, Kishor S. and Sharma, Shilpa and Ali, Tarig and Nave, Jerry Wayne and Chakrabortty, Rabin},
  title = {Machine Learning-Based Forecasting of Wet-Bulb Temperature and Two-Decade Urban Climate Shifts},
  journal = {Earth Systems and Environment},
  year = {2025},
  doi = {10.1007/s41748-025-00911-9},
  url = {https://doi.org/10.1007/s41748-025-00911-9}
}