Categorizing Urban Structural Types using an Object-Based Local Climate Zone Classification Scheme in Medellín, Colombia

Climate change is reshaping societies. As we see more and more people moving to urban areas an ever-increasing number settles in low-cost and more hazardous areas. However, due to the rapid growth and sheer scale of informal settlements, knowledge gaps often exist on location or quantity. In this sense, Earth Observation combined with machine learning techniques allows to generate reliable geo-information. In this study, we classify the morphologically heterogeneous entire urban area of Medellín, Colombia into urban structural types. We do this by the Local Climate Zone (LCZ) scheme. Our specific focus is on one structural type, i.e. informal settlements. We test whether it is feasible by the LCZ concept to localize and quantify these vulnerable areas. The LCZ scheme is generic, replicable, neutral, and has become widespread in urban studies. We use urban blocks to perform a scene-based image classification into nine LCZs. We refer to multi-modal remotely-sensed data: high-resolution multispectral image data and elevation data. We apply an optimized random forest algorithm using shape metrics, as well as spectral and texture features. In general, we find the LCZ classification, measured with an overall accuracy of 82%, shows a reliable representation of urban typologies and functions across the city. Specifically, we compare the urban blocks classified as the LCZ lightweigth low-rise to the informal settlements provided by the city of Medellín. Here we reach an agreement of 86%. Besides, our approach complements the official dataset by including recently developed areas which are not yet considered by the city