Evaluation of spatial patterns accuracy in identifying built-up areas within risk zones using deep learning, RGB aerial imagery, and multi-source GIS data

Abstract

In the presence of natural disasters that increasingly affect urban centers, innovative methodologies that can support all the subjects and bodies involved in the disaster management system are increasingly important. This task can be enhanced in urban settings by automatically assessing at-risk buildings through satellite and aerial imagery. However, creating and implementing models with robust generalization capabilities is crucial to achieving this goal. Based on these premises, the authors proposed a deep learning approach utilizing the U-Net model to map buildings within known landslide-prone areas. They trained and validated the U-Net model using the Dubai Satellite Imagery Dataset. The model's prediction accuracy in adapting its results to urban environments in Italy, different from those involved in the training and validation stages, was tested using natural color orthoimages and diverse geographic information system (GIS) data sources. The outcomes indicate that the model's predictions are better in contexts with denser urban fabric. The level of accuracy in dispersed urban shapes worsens as building footprints cover a small portion of the total image area. Overall, the results demonstrate that the suggested methodology can effectively identify buildings in landslide risk zones, demonstrating noteworthy adaptability, making the proposed platform a tool that can be instrumental for decision-makers and urban planners in pre-disaster and post-disaster stages.