Resnet50 to Detect Landslides, Damaged Infrastructures and Crumbled Houses from Haiti 2010 and 2021 Earthquakes

Abstract

In this paper, the residual convolutional neural network Resnet50 is applied to satellite imagery collected on the 12 January 2010 earthquake with a moment magnitude (Mw) of 7.0 that struck the western cities of Haiti such as Port-au-Prince, Leogane and Jacmel, and on 14 August 2021 another earthquake with moment magnitude (Mw) of 7.2 that hit the southwestern peninsula of Haiti with its epicenter located not too far from the main city of Les Cayes. Meta data that provide geolocations of landmark buildings, residential quarters, road infrastructures and landslide areas are used to partition the post-earthquake satellite images and create three class databanks that allow training and testing of the Resnet50 architecture to establish similarities between western and southwestern areas of the country in land topography, housing quarters and road networks. In a first experiment, datasets derived from the post-earthquake image of 12 January 2010 are used to train the network while the datasets from the post-earthquake of 14 August 2021 are reserved for testing; the network architecture Resnet50 exhibits an average performance of about 88% on testing. Using data augmentation by 8 fold on the training set with datasets from the 14 August 2021 earthquake, testing performance on the 12 January 2010 earthquake improves by 4% with the network trained on the original datasets. Therefore, Resnet50 appears to be a well suited network architecture to detect and locate land areas, houses and roads severely impacted by an earthquake.