Urban Temperature Dynamics in Sri Lanka: A Comparative Machine Learning and Time-Series Analysis Across Wet, Intermediate, and Dry Climate Zones
Authors:
Preprint
- 8 May 2026
- preprint
- Published by Springer Nature in Research Square
Abstract
Rapid urbanization and land-use change have intensified temperature variability in tropical urban environments. Sri Lanka is a tropical country that is experiencing heat stress across its urban areas in all three climate zones. Daily temperature forecasting is essential for planning in agriculture, energy, and public health. Daily meteorological data (2010–2023) were obtained for three representative urban stations: Colombo (wet zone), Kurunegala (intermediate zone), and Puttalam (dry zone), to analyze daily mean temperatures in urban areas, with 4,916 daily observations per station. Outliers were handled using the IQR-based winsorization method, except for wind direction and rainfall. The rainfall was handled using the log(x + 1) transformation. Autoregressive Integrated Moving Average with Fourier terms (ARIMAF) was applied as a time series, along with Multiple Linear Regression (MLR), Polynomial Regression, Support Vector Regression (SVR), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) as machine learning algorithms. The best predictive performance was achieved by SVR models with the Radial Basis Function (RBF) kernel. The coefficient of determination (R²) values exceeded 0.98 in all climate zones, and the MAE ranged from 0.11°C to 0.13°C, while the RMSE ranged from 0.14°C to 0.17°C. All other machine learning techniques showed strong, stable performance in urban regions, whereas the ARIMAF showed significantly lower accuracy than the machine learning techniques. The feature importance analysis revealed that apparent temperature mean, temperature max, temperature min, and evapotranspiration were the dominant predictors. The proposed models effectively capture the temperature variations and are suitable for urban temperature forecasting and climate-related applications.Keywords
All Related Versions
This publication has 31 references indexed in Scilit:
- Modeling climate change impacts and predicting future vulnerability in the Mount Kenya forest ecosystem using remote sensing and machine learningEnvironmental Monitoring and Assessment, 2025
- Projecting and Downscaling Future Temperature and Precipitation Based on CMIP6 Models Using Machine Learning in Hatay Province, TürkiyePure and Applied Geophysics, 2025
- Leveraging Groundwater Dynamics to Improve Predictions of Summer Low‐Flow DischargesWater Resources Research, 2023
- Protein-Specific Signal Peptides for Mammalian Vector EngineeringACS Synthetic Biology, 2023
- An outliers detection and elimination framework in classification task of data miningDecision Analytics Journal, 2023
- Urban heat islands: a review of contributing factors, effects and dataIOP Conference Series: Earth and Environmental Science, 2023
- An accurate and fully-automated ensemble model for weekly time series forecastingInternational Journal of Forecasting, 2022
- Hyperparameters and tuning strategies for random forestWIREs Data Mining and Knowledge Discovery, 2019
- XGBoost2016
- Identifying appropriate spatial scales of predictors in species distribution models with the random forest algorithmMethods in Ecology and Evolution, 2012