Machine learning-driven analysis of Urban Heat Island dynamics in Coimbatore: a two-decade spatio-temporal assessment using XGBoost and Random Forest.
Saved in:
| Title: | Machine learning-driven analysis of Urban Heat Island dynamics in Coimbatore: a two-decade spatio-temporal assessment using XGBoost and Random Forest. |
|---|---|
| Authors: | Gadekar, Kajesh1 (AUTHOR), Mathew, Aneesh1 (AUTHOR) aneesh52006@gmail.com |
| Source: | Journal of Earth System Science. Jun2026, Vol. 135 Issue 2, p1-34. 34p. |
| Subject Terms: | *Land surface temperature, *Machine learning, *Climate change, *Random forest algorithms, *Spatiotemporal processes, *Boosting algorithms, *Urban heat islands, *Cities & towns |
| Geographic Terms: | India, Coimbatore (India) |
| Abstract: | This study investigates the spatio-temporal dynamics of Land Surface Temperature (LST) in Coimbatore, India, from 2001 to 2022, with a focus on assessing the Urban Heat Island (UHI) effect and evaluating the performance of machine learning models— Extreme Gradient Boosting (XGBoost) and Random Forest (RF)—for LST prediction. A significant warming trend is observed, particularly in urban regions. The maximum annual LST peaked at 26.8 °C in 2016, while the minimum was recorded at 8.5 °C in 2001. The mean night-time LST increased from 20.5 °C in 2006 to 22.6 °C in 2019. Seasonal trends showed that summer LST ranged from 20.7 °C to 26.7 °C, with pronounced warming in 2016. In contrast, winter LST varied between 16.7 °C and 24.7 °C, with notable cooling observed in rural areas. The urban–rural LST differential increased over time, reaching approximately 1 °C. Urban areas consistently exhibited higher temperatures, peaking at 23.7 °C in 2019. The warming trend in urban zones was statistically stronger, with Kendall's τ = 0.593 compared to τ = 0.429 for rural areas, indicating a significant and growing UHI effect. These findings underscore the urgent need for climate-responsive urban planning and green infrastructure to mitigate thermal stress in rapidly urbanizing regions. Correlation analysis was conducted to understand the influence of environmental parameters on LST. In winter, the Normalized Difference Built-up Index (NDBI) showed a moderate positive correlation with LST (r = 0.31), indicating the contribution of built-up areas to surface warming. The Enhanced Vegetation Index (EVI) and elevation showed negative correlations with LST (r = -0.46 and -0.44, respectively), reflecting the cooling effects of vegetation and higher altitudes. Aerosol Optical Depth (AOD) and the Modified Normalized Difference Water Index (MNDWI) showed weak positive correlations (r = 0.14 and 0.13). In the summer, similar patterns emerged, with NDBI (r = 0.15), AOD (r = 0.21), EVI (r = -0.26), and elevation (r = -0.21) maintaining their respective trends. The Moisture Built-up Index (MBI) exhibited a negligible correlation (r = 0.04), and MNDWI remained weakly positive (r = 0.07), suggesting a limited role in modulating LST. To model and predict LST, both XGBoost and RF were trained and tested using 21 years of data (2001–2021), with 2022 as the validation year. In winter, RF slightly outperformed XGBoost in error metrics (RMSE = 0.7925, MSE = 0.6281, MAE = 0.699), yet XGBoost achieved better generalization with a higher R2 value of 0.8616 compared to RF's 0.7961. In the summer, XGBoost demonstrated superior performance with RMSE = 0.9405, MAE = 0.9164, and R2 = 0.7529, outperforming RF across all metrics. XGBoost closely matched observed values in 2022, predicting annual LST values of 23.93 °C (max), 18.06 °C (min), and 21.53 °C (mean), compared to actuals of 25.02 °C, 18.45 °C, and 22.84 °C. Although RF had a slightly lower error value of 1.16 °C, XGBoost better captured extremes and maintained spatial fidelity, with most errors within ±1 °C in urban zones. Overall, XGBoost proved to be a more robust and accurate model for predicting LST across different seasons and landscapes. The findings validate the utility of machine learning—especially ensemble methods like XGBoost—for urban climate monitoring, long-term LST modeling, and informed decision-making for sustainable urban development and UHI mitigation strategies. Research highlights: Urban LST rose by ~1 °C (2001–2022), confirming a significant and growing UHI effect. Urban areas showed a stronger warming trend (Kendall's τ = 0.593) than rural (τ = 0.429). XGBoost achieved higher R2 (0.8616 winter; 0.7529 summer) than Random Forest. Vegetation (EVI r = -0.46) and elevation (r = -0.44) had strong cooling effects on LST. XGBoost predicted 2022 LST with errors mostly within ±1°C in urban zones. [ABSTRACT FROM AUTHOR] |
| Database: | Energy & Power Source |
|
Full text is not displayed to guests.
Login for full access.
|
|
| Abstract: | This study investigates the spatio-temporal dynamics of Land Surface Temperature (LST) in Coimbatore, India, from 2001 to 2022, with a focus on assessing the Urban Heat Island (UHI) effect and evaluating the performance of machine learning models— Extreme Gradient Boosting (XGBoost) and Random Forest (RF)—for LST prediction. A significant warming trend is observed, particularly in urban regions. The maximum annual LST peaked at 26.8 °C in 2016, while the minimum was recorded at 8.5 °C in 2001. The mean night-time LST increased from 20.5 °C in 2006 to 22.6 °C in 2019. Seasonal trends showed that summer LST ranged from 20.7 °C to 26.7 °C, with pronounced warming in 2016. In contrast, winter LST varied between 16.7 °C and 24.7 °C, with notable cooling observed in rural areas. The urban–rural LST differential increased over time, reaching approximately 1 °C. Urban areas consistently exhibited higher temperatures, peaking at 23.7 °C in 2019. The warming trend in urban zones was statistically stronger, with Kendall's τ = 0.593 compared to τ = 0.429 for rural areas, indicating a significant and growing UHI effect. These findings underscore the urgent need for climate-responsive urban planning and green infrastructure to mitigate thermal stress in rapidly urbanizing regions. Correlation analysis was conducted to understand the influence of environmental parameters on LST. In winter, the Normalized Difference Built-up Index (NDBI) showed a moderate positive correlation with LST (r = 0.31), indicating the contribution of built-up areas to surface warming. The Enhanced Vegetation Index (EVI) and elevation showed negative correlations with LST (r = -0.46 and -0.44, respectively), reflecting the cooling effects of vegetation and higher altitudes. Aerosol Optical Depth (AOD) and the Modified Normalized Difference Water Index (MNDWI) showed weak positive correlations (r = 0.14 and 0.13). In the summer, similar patterns emerged, with NDBI (r = 0.15), AOD (r = 0.21), EVI (r = -0.26), and elevation (r = -0.21) maintaining their respective trends. The Moisture Built-up Index (MBI) exhibited a negligible correlation (r = 0.04), and MNDWI remained weakly positive (r = 0.07), suggesting a limited role in modulating LST. To model and predict LST, both XGBoost and RF were trained and tested using 21 years of data (2001–2021), with 2022 as the validation year. In winter, RF slightly outperformed XGBoost in error metrics (RMSE = 0.7925, MSE = 0.6281, MAE = 0.699), yet XGBoost achieved better generalization with a higher R2 value of 0.8616 compared to RF's 0.7961. In the summer, XGBoost demonstrated superior performance with RMSE = 0.9405, MAE = 0.9164, and R2 = 0.7529, outperforming RF across all metrics. XGBoost closely matched observed values in 2022, predicting annual LST values of 23.93 °C (max), 18.06 °C (min), and 21.53 °C (mean), compared to actuals of 25.02 °C, 18.45 °C, and 22.84 °C. Although RF had a slightly lower error value of 1.16 °C, XGBoost better captured extremes and maintained spatial fidelity, with most errors within ±1 °C in urban zones. Overall, XGBoost proved to be a more robust and accurate model for predicting LST across different seasons and landscapes. The findings validate the utility of machine learning—especially ensemble methods like XGBoost—for urban climate monitoring, long-term LST modeling, and informed decision-making for sustainable urban development and UHI mitigation strategies. Research highlights: Urban LST rose by ~1 °C (2001–2022), confirming a significant and growing UHI effect. Urban areas showed a stronger warming trend (Kendall's τ = 0.593) than rural (τ = 0.429). XGBoost achieved higher R2 (0.8616 winter; 0.7529 summer) than Random Forest. Vegetation (EVI r = -0.46) and elevation (r = -0.44) had strong cooling effects on LST. XGBoost predicted 2022 LST with errors mostly within ±1°C in urban zones. [ABSTRACT FROM AUTHOR] |
|---|---|
| ISSN: | 02534126 |
| DOI: | 10.1007/s12040-026-02771-x |
