School of Science Department of Mathematics 41 A Machine Learning Approach to Study the Relationship between Urban Morphology and Urban Heat Island Supervisor: FUNG Jimmy Chi Hung / MATH Student: HE Qihao / DSCT Course: UROP2100, Fall Neural Network (NN) and Random Forest Regressor (RFR) models are common tools in machine learning. Based on the result of our previous study, this study will examine the models’ performance in terms of accuracy and efficiency in predicting the Land Surface Temperature (LST) of the Hong Kong Islands. To improve the performance of the models, we designed a feature extraction method, tried a special way of train-test set splitting, and corrected a minor shift of the temperature image compared with other feature images using QGIS. The RFR has outperformed the NN with the optimal network structure among the tested configurations with two inner layers. A Machine Learning Approach to Study the Relationship between Urban Morphology and Urban Heat Island Supervisor: FUNG Jimmy Chi Hung / MATH Student: MOK Wan Hin / MATH-IRE Course: UROP1100, Fall UROP2100, Spring UROP3100, Summer In the previous semester, we built machine learning models that study the quantitative relationship between urban morphology and the heat island effect by relating the former with Land Surface Temperature (LST) by using MODIS LST products and Random Forest models. We found that the model had a high correlation (more than 0.9) between the predicted LST and the measured LST over the whole year, but some of this came from the seasonal variations that the model observed. For regional variations in LSTs due to urban morphology, we found that the performance varied from day to day. This semester, similar models were hoped to be explored for Sentinel images. A Machine Learning Approach to Study the Relationship between Urban Morphology and Urban Heat Island Supervisor: FUNG Jimmy Chi Hung / MATH Student: SHEN Che / DSCT Course: UROP1100, Fall The goal of this research is to study the urban heat island effect, which is a commonly observed meteorology phenomenon. Since the traditional three-dimensional model is computationally costly with many uncertainties, we attempted to build machine learning models for predicting the land surface temperature (LST) by some possible influencing factors, including average building height (ah), total building surface (lb), population, terrain height, tree height and impervious surface percentage of corresponding locations. We gather the data of the land surface temperature (LST) and predictors with different formats, transform them into Excel form and process them to make the sample for regression. After the sample is accomplished, we use it to train different regression models, evaluate their performance and analyze the possible causes of deviation.