Deep Learning Based Exposure Analysis of LandslideProne Areas in Medellín, Colombia

Kurzfassung

In the last century, Medellín grew into one of Colombia's largest cities. Today, the city continues to grow primarily due to the influx of internally displaced people (IDP’s), who have been forced to leave their homes at the country side due to natural disasters or drug-related violence. Since the internally displaced are mostly lowincome farmers and peasants, they are migrating to the larger cities in search of greater security and jobs. In Medellín, the new residents mostly settle informally on the steep slopes to the east and west of the city. Due to limited space and steep topography, such settlements are often built in areas with medium and high probability of landslides. However, not only free land area within the municipal boundaries are exploited by the build-up of new settlements, but also free land beyond the border of the municipality, which causes the city to grow into the rural area. The study therefore seeks to find out how many residents are prone to potential landslide activity in the context of the pattern of migration. To analyze the exposure, the population is disaggregated down to the individual building block level. Such an approach requires precise building footprints to locate the population in relation to landslide-prone areas. Although the city has a cadaster including building footprints, it is more imprecise and incomplete towards the outskirts of the city, where landslide susceptibility is most pronounced. The incompleteness is due to the high population dynamics, which makes it quite difficult to maintain an up-to-date cadaster. But since Medellín's geospatial data service provides an orthophoto from 2019, a deep learning-based building extraction is applied to generate a more comprehensive building footprint dataset. This will be the main data source for the exposure analysis. The respective deep learning architecture is a U-Net has been refined with the EfficientNetB2 as a backbone and eventually fine-tuned. It could show very accurate results, while still facing some challenges, like the field-of.view of the image tiles, that is sometimes too small for the vast rooftop landscapes, which leads to misclassifications. The exposure analysis of population exposed to landslide hazard could prove the importance of having a more up-to-date data basis. While the number of residents living in formal settlements is not to different from the cadaster and the deeplearning derived building footprints, those numbers of residents of the informal settlements are much higher in the more actual deep learning derived dataset. A strong increase could also be found in the population exposed to medium and high landslide hazard. Further analyses facilitate the impression, that the poorer and the more vulnerable population has distinctively higher exposure to considerable landslide hazard, when using the deep-learning derived dataset. These findings show the strength of remote sensing techniques in order to retrieve actual building footprint data, that is clearly important for the estimation of potential consequences of landslide-prone areas.