Examination of training data for analysis of damage to buildings during the 3.11 North-East Japan Earthquake 2011 using deep learning

Abstract

Numerous destructive earthquake have occurred in Japan, and particularly, it's still fresh in our memory that large number of buildings were washed away or completely destroyed over a wide area in the 3.11 North-East Japan Earthquake 2011. It is essential to assess the damage to buildings in a wide area in order to promptly perform rescue and restoration work immediately after the earthquake. In this study, we classify the damage to buildings using satellite remote sensing, which can observe a wide area of the earth's surface, and deep learning, which can learn and extract features from the data by itself. And also high-resolution optical satellite images of Ishinomaki City after the earthquake were used to classify the damage to buildings into "washed away, " "damaged, " and "undamaged" using a Convolutional Neural Network (CNN), and the damage was evaluated quantitatively using reference data. We focused on the number of training data, rotation angle, tile size, and CNN hierarchy to obtain quantitative knowledge of the effects of these parameters on the classification results. The results showed that the combination of rotation angle and CNN hierarchy was as effective as increasing the number of training data.