Using High-resolution Aerial Photography and Neural Networks to Inventory Properties at Risk of Earthquakes

Earthquakes are one major cause for the loss of many lives and properties in many regions around the world. In this project we develop algorithms based on aerial images and neural networks to better assess the risk of earthquakes for buildings. The algorithm consists of two major parts: extraction of buildings footprints from aerial photos and classification of structure type with neural network. The neural network model generates the structure type based on parameters such as usage, size, height, shape and year of construction. Many research projects in the past attempted to automatically or semi-automatically extract buildings from aerial and satellite images. In this research, we suggest the use of the parcels layer in order to simplify the building extraction task. Simple clipping of the digital image to parcel boundaries can dramatically reduce the amount of information that needs to be processed when searching for a building footprint. The current research shows preliminary results of the building extraction task and an automated building classification. The project scope also involves future use of LIDAR images to either refine or completely extract the buildings.