Utilizing Neural Networks to Predict Land Surface Temperature in Urban Areas Using Satellite Imagery in the Czech Republic

Rising average temperatures in cities are reshaping the way we view urban structure and development. Average July land surface temperatures (LST) can vary by more than 10 °C within a single city, and thus, the quality of life is increasingly influenced by the specific district in which residents live. This thesis aims to develop a neural network model for predicting land surface temperature (LST) in Prague and its surrounding areas, with potential applications in urban planning. The thesis also presents a data processing and modeling methodology that can be adapted for LST prediction in other cities. The methodology combines publicly available data from various sources, such as NASA, ESA, and OSM, and incorporates RBF and other methods to enrich the data with spatial dependencies. The resulting model achieves a mean absolute error (MAE) of 0.2095 °C and explains 96.53 % of the variability in LST. The model’s predictive performance was further examined through a ceteris paribus experiment, which ruled out excessive influence of spatially dependent variables on the LST prediction, supporting the model’s suitability for further application.