Global transferability of local climate zone models

Using the cloud-computing resources of Google's Earth Engine (EE) and a range of satellite sensors (input features) this paper for the first time explores the potential of up-scaling the current Local Climate Zone mapping efforts to regional and global scales. Using a transferability framework, we test whether information from one city contains valuable information to categorise a different city, simultaneously exploring the role of the input features and the characteristics of individual cities. It was found that the accuracies of the EE approach are comparable to the standard WUDAPT method, making EE a viable alternative approach. The results from the city-to-city experiments are generally poor when compared to the single city benchmark experiments, indicating that the collection of site-specific training areas remains relevant. However, LCZ mapping accuracies are considerably improved when a) the source of the training data is from a city in the same ecoregion as the city of interest and b) if the training areas from several cities are combined. These results support the claim that the LCZ framework is a universal urban typology and indicate that, provided a continued optimisation of input features and quality of training areas, up-scaling to regional or global levels is feasible.