Evaluating land surface temperature trends and environmental interactions through machine learning and wavelet analysis

Accurate land surface temperature (LST) assessment is crucial for comprehending and reducing the impacts of climate change and understanding land use evolution. This study presents an innovative method by utilizing ensemble models, advanced correlation analysis, and trend analysis to investigate its environmental influences. Google Earth Engine (GEE) was utilized to process the datasets from Landsat-7 and Landsat-8 for the five big cities of Punjab, Pakistan, from 2001 to 2023. Results from this study show significant urban warming trends, and a strong correlation between environmental variables and LST was identified. The ensemble-based three machine learning models, including XGBoost, AdaBoost, and random forest (RF), were adopted to improve the accuracy of LST evaluation. Although XGBoost and AdaBoost attained modest levels of accuracy, with R2 values of 0.767 and 0.706, respectively, the RF model outperformed them by achieving an exceptional R2 of 0.796 and RMSE of 0.476. Moreover, Pearson correlation analysis revealed a negative relationship between LST and normalized difference latent heat index (NDLI) with r=−0.67, normalized difference vegetation index (NDVI) with r=−0.6, and modified normalized difference water index (MNDWI) with the value of r as −0.57. In addition, wavelet analysis showed that vegetation and water offer long-term LST cooling, lasting up to 64 months, while built-up areas and bare soil contribute to short-term warming, lasting 4 to 8 months. Latent heat indicated variable cooling periods, surpassing 60 months in cities. These findings enhance the understanding of LST changes and the impact of climate change on the environment.