SPATIAL DOWNSCALING OF LAND SURFACE TEMPERATURE MAPS – A CASE STUDY FOR THE CITY MUNICIPALITY MEDIJANA, THE CITY OF NIŠ, SERBIA

Abstract

Urban heat islands are becoming an increasingly important challenge in the 21st century. Land surface temperature is a key factor in the urban heat island risk assessment. Its monitoring is possible through satellite land surface temperature (LST) detection. However, satellite images with insufficient spatial resolution are available to monitor temperature changes in urban areas, which arise due to the technical limitations of satellite thermal infrared sensors. Numerous algorithms have been proposed to solve the problem of the coarse spatial resolution of LST. This study explores the application of a machine learning algorithm based on a Random Forest (RF) regression model between LST and predictor variables such as aspect, digital elevation model (DEM), hillshade, normalized difference vegetation index (NDVI),  building heights, digital height model (DHM), and land cover. The study focuses on the municipality of Medijana, located in the City of Niš in the Republic of Serbia. The spatial resolution of MODIS LST was improved from 1 km to 250 m. The results indicate that the applied machine learning method can predict potential temperatures at a finer scale with high accuracy, with NDVI indicating a significant local influence on LST. The results indicated that the RF approach demonstrated a robust and high-performance methodology. The Mean Square Error (MSE) values ranged from 0.730 °C² to 1.028 °C², while the Root Mean Square Error (RMSE) values varied from 0.854 °C to 1.100 °C across the 84 models generated.