Presentation Information
[U16-P03]Analysis of Building Damage caused by the 2011 off the Pacific Coast of Tohoku Earthquake using Optical Satellite Imagery and Deep Learning focused on Subdivision of Damage Classification Categories
*Ryoki Jinoka1, Takashi Nonaka1, Tomohito Asaka1 (1.Nihon University)
Keywords:
deep learning,high resolution optical satellite imagery,building damage
Japan is frequently struck by natural disasters; in particular, large-scale earthquakes cause immense damage to human life, property, and the economy. A notable recent example is the 2011 Tohoku Earthquake, which caused extensive damage across a wide area—including the Tohoku region—due to both seismic motion and tsunamis. With major events such as the Nankai Megathrust and Tokyo Metropolitan earthquakes anticipated in the near future, it is essential to quickly and accurately assess building damage immediately after a disaster to facilitate prompt rescue operations and recovery activities. Currently, damage assessment primarily relies on on-site visual inspections, which are time-consuming, labor-intensive, and pose risks of secondary disasters from aftershocks. While research has explored using Synthetic Aperture Radar (SAR) imagery to detect changes, this approach is constrained by the need for multi-temporal data from the same orbit.
To address these challenges, this study utilizes satellite remote sensing and deep learning to classify building damage. While previous studies often used three categories—“washed away,” “damaged,” and “undamaged”—actual disaster response requires more granular data to prioritize rescue efforts. Therefore, this study investigates the relationship between category subdivision and classification accuracy. Using the ResNet-34 model and high-resolution optical imagery from the GeoEye-1 satellite, we classified damage in Ishinomaki City following the 2011 earthquake into four categories: “washed away,” “severely damaged,” “minor damaged,” and “undamaged.” Training data (200 samples per category) was sourced from the Reconstruction Support Survey Archive. The results indicated that the recall rate for “damaged” in the three-category model and “severely damaged” in the four-category model remained low. Furthermore, subdividing the categories led to a 6% decrease in recall and a 10% decrease in precision for the “washed away,” category. While more granular classification helps clarify disaster conditions, the findings suggest that incorporating the surrounding environmental context into training data is crucial for improving accuracy.
To address these challenges, this study utilizes satellite remote sensing and deep learning to classify building damage. While previous studies often used three categories—“washed away,” “damaged,” and “undamaged”—actual disaster response requires more granular data to prioritize rescue efforts. Therefore, this study investigates the relationship between category subdivision and classification accuracy. Using the ResNet-34 model and high-resolution optical imagery from the GeoEye-1 satellite, we classified damage in Ishinomaki City following the 2011 earthquake into four categories: “washed away,” “severely damaged,” “minor damaged,” and “undamaged.” Training data (200 samples per category) was sourced from the Reconstruction Support Survey Archive. The results indicated that the recall rate for “damaged” in the three-category model and “severely damaged” in the four-category model remained low. Furthermore, subdividing the categories led to a 6% decrease in recall and a 10% decrease in precision for the “washed away,” category. While more granular classification helps clarify disaster conditions, the findings suggest that incorporating the surrounding environmental context into training data is crucial for improving accuracy.