A Deep Learning Model for Road Damage Detection After an Earthquake Based on Synthetic Aperture Radar (SAR) and Field Datasets

Abstract

This article is a new assessment of damaged roads after the Kumamoto earthquake in southern Japan (2016) using remotely sensed synthetic aperture radar (SAR) data, field data and deep learning. Three SAR images from descending orbits of Sentinel-1 in vertical-vertical polarizations are considered for radiometric calibration, geocoding and interferometric analyses. Field data in terms of the international roughness index (IRI) were gathered over more than 530 km using a smartphone accelerometer and the BumpRecorder application. The relationship between SAR data and IRI data was investigated in a binary (0 and 1) mode to establish a multilayer perceptron model of damaged and intact roads. We found the remote sensing SAR datasets suitable, not only for the detection of damaged roads, but also as an indicator of road roughness changes. The classification results for damaged and intact roads indicated that our datasets (SAR and field measurements), together with a deep learning model, yielded acceptable overall accuracy (87.1%).

Keywords

• Deep learning
• international roughness index (IRI)
• Kumamoto
• synthetic aperture radar (SAR)