Instance Segmentation of Informal Buildings in Medellín for Assessing Population at Risk from Landslides

In informal settlements, accurate population data, such as the number of people per building, is often lacking, outdated, or incomplete, yet such information is crucial for assessing the population at risk in hazard-prone areas, such as parts of Medellín, Colombia. This study presents a methodology to estimate the population at risk from landslides in informal areas by combining deep learning-based building detection on aerial remote sensing images with a population estimation. A Mask R-CNN instance segmentation model is trained to detect individual buildings in orthophotos of informal settlements from 2019, using official cadastral data from 2018. Population estimates are made based on the size of the detected buildings. When applied to informal settlements within landslide-prone regions, our methodology identifies 10,180 more buildings and 28,575 more people compared to official cadastral data, revealing an underestimation of people at risk in these areas. This approach demonstrates the value of combining instance segmentation with population estimation to enhance risk assessment in data-limited settings.