A hybrid machine learning approach to investigate the changing urban thermal environment by dynamic land cover transformation : a case study of Suwon, republic of Korea

Abstract

Urban thermal environment should be analyzed by considering the dynamic structural changes as cities grow both horizontally and vertically. Local Climate Zone (LCZ) scheme can describe built-up areas in detail, mainly based on density and height; however, the low overall accuracy of LCZ urban classes (OAurb) remains a notable limitation that requires improvement. This study proposes a hybrid analytical method considering bidirectional urban expansion and low OAurb. Temporal LCZ maps were constructed using a convolutional neural network to observe the dynamic urban growth between 2004 and 2021 in Suwon, South Korea. Unlike previous LCZ mapping studies, we utilized the additional information provided by deep learning through softmax-based probability maps. Random forest-based downscaling models were developed by combining various auxiliary variables related to the Land Surface Temperature (LST) to observe the detailed surface energy flux. A filtering method was then employed by eliminating areas where LCZs were identified with a low confidence level using extracted probability maps. Finally, thermal variability was investigated by overlaying the filtered LCZ maps and the corresponding LST. The produced LCZ maps and spatially downscaled LSTs accurately depicted dynamic urban form changes, with the LCZ maps exhibiting an average overall accuracy of approximately 90% and downscaled LSTs showing an average coefficient of determination of ∼ 0.9 and a root mean square error of 0.7 °C. Thermal variability occurring due to structural transitions varied in magnitude depending on the height and density of the buildings, while exhibiting a maximum and minimum value of 2.8 °C and − 2.2 °C, respectively. By selecting reliably classified areas, the proposed filtering method produced more rational results than the original non-filtering method, resulting in higher variability from − 0.4 °C to 0.6 °C.