Pendekatan GeoAI untuk pemodelan spasial suhu permukaan dan urban heat island di kota cimahi

Abstract

Urban heat island (UHI) intensity in rapidly urbanizing Indonesian cities is commonly assessed through land surface temperature (LST) derived from satellite thermal sensors, yet fine-resolution, machine-learning-based spatial prediction for small and medium cities in the Bandung Raya region remains limited. This study applies a geospatial artificial intelligence (GeoAI) workflow to model and map LST in Kota Cimahi, West Java, using four cloud-screened Landsat 8/9 Collection 2 Level-2 thermal scenes acquired between July and August 2025, resampled to the 10 m grid of Sentinel-2-derived NDVI, NDBI, and NDWI and Dynamic World built-up density and land-cover layers. Random Forest (RF) and Support Vector Regression (SVR) models were trained on 627,687 pixels and evaluated on 269,009 independent pixels. RF achieved a marginally higher test R² (0.528; RMSE 2.17 °C; MAE 1.63 °C) than SVR (R² 0.525; RMSE 2.18 °C; MAE 1.61 °C). Built-up density and NDBI dominated under RF impurity-based importance, whereas NDVI dominated under SVR permutation importance, reflecting strong multicollinearity among the optical predictors. Spatial prediction confirmed a consistent thermal gradient, with the highest LST concentrated in the built-up southern core and the lowest in the vegetated northern highlands, although one acquisition date relied on cross-date gap-filling for over 90% of its pixels because of persistent cloud cover — a limitation reported transparently in this article rather than concealed.