Assessment of Urban Heat Island Intensity and Key Drivers in Three Major Indonesian Cities Using Machine Learning
Wulan Salle Karurung 1*, Lee Ki Rim 2, Lee Won Hee 3

1 Graduate Student, Department of Convergence and Fusion System Engineering, Kyungpook National University, Republic of Korea. *wulansalle[at]gmail.com
2 Research Visiting Professor, Research Institute of Artificial Intelligent Diagnosis Technology for Multi-Scale Organic and Inorganic Structure, Kyungpook National University, Republic of Korea
3 Professor, Department of Location-Based Information System, Kyungpook National University, Republic of Korea

Abstract

Urban heat island (UHI) effects in tropical cities are increasingly severe due to urban expansion, reduced green cover, and high population density. Environmental and anthropological factors interact to increase surface temperatures, necessitating studies that examine these factors in detail areas to anticipate further long-term impacts. This study aims to quantify and compare UHI characteristics and driving factors in three metropolitan cities in Indonesia, such as Jakarta, Surabaya, and Makassar. This study integrates the multivariate spatiotemporal data monthly in 2019-2020, covering the natural and anthropogenic factors. The datasets were collected from satellite derived datasets, including urban heat index intensity (UHII), land surface temperature (LST), vegetation indices (NDVI), built-up indices (NDBI), PM2.5, NO2, precipitation, population density, and land use/land cover sourced from MODIS, Landsat 8, Sentinel 5P, and other open-access platforms. All the datasets were standardized into a 100 m spatial resolution. Three machine learning models, Random Forest (RF), XGBoost, and LightGBM (LGBM), were compared to evaluate the prediction accuracy. Based on the final accuracy results, RF outperformed other models with R2 values of 0.92 for Jakarta, 0.86 for Surabaya, and 0.82 for Makassar. SHAP and partial dependence plots (PDP) were used to interpret the importance of features and their effects on interaction, using the results of RF models. The results show Jakarta had the highest UHI intensity, and the top influencing factors were population density, PM2.5, and LST. In contrast, Surabaya and Makassar were more influenced by population and precipitation. This study demonstrates the importance of integrating multi-feature geospatial data with machine learning to improve our understanding of data interactions in prediction models. The results can support regional climate resilience actions by identifying vulnerable areas and key factors for sustainable urban planning.

Keywords: Urban Heat Island (UHI), Machine Learning, Spatiotemporal Analysis, Remote Sensing

Topic: Topic B: Applications of Remote Sensing