Deep learning–based individual segmentation and damage assessment of earthquake-damaged buildings: a case study of the old town of Beichuan county earthquake ruins

Introduction Addressing the technical demand for rapid and precise building segmentation and intelligent damage assessment after earthquakes, this study proposes a collaborative method based on improved YOLOv8 for individual building segmentation and damage assessment. Methods Using UAV imagery of the Beichuan old town ruins, we build a dual_task sample database and separately train segmentation and classification models via network improvements, under an integrated framework of image segmentation, individual extraction, and multi_view assessment. The segmentation model extracts building footprints and combines with DSM to build 3D constraint boxes for isolating individual buildings. We then use five_direction virtual camera imaging and the most severe damage criterion for automatic damage grading. Results Validation via high_accuracy 3D models and field surveys yields: (1) Under 300_epoch progressive training, both models converge stably. building.pt achieves mAP50(B)=0.934 and mAP50_95(B)=0.803, and assessment.pt attains Accuracy of 0.64, Precision of 0.58, Recall of 0.75, and F1_Score of 0.65, respectively, demonstrating effective handling of irregular building outlines and complex damage morphologies (2) A total of 64 target buildings are present in the study area, of which 61 are successfully identified and segmented by the algorithm, yielding an overall recognition success rate of 89.1% (= 57/64 × 100%), a segmentation accuracy of 92.2% (= 59/64 × 100%), and a classification accuracy of 93.4% (= 57/61 × 100%). Discussion These results confirm that the proposed method can accurately perform individual building extraction and multi_view damage assessment.Regional analysis indicates that the Wangjiayan landslide west of the town is the main cause, with its frontal accumulation burying structures and impact loads progressively damaging central and peripheral buildings, creating a graded pattern of burial, impact, and disturbance, making the old urban area the worst-hit zone.

Authors

Institutions

Publication Details

Journal
Frontiers in Remote Sensing
Published
2026-11-17
DOI
https://doi.org/10.3389/frsen.2026.1963497
Primary Topic
Infrastructure Maintenance and Monitoring
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Deep learning–based individual segmentation and damage assessment of earthquake-damaged buildings: a case study of the old town of Beichuan county earthquake ruins

Zongpan Bian, Chuanjin Liu, Hongbin Cheng, Hongtao Tang et al.
4 links
Frontiers in Remote Sensing
Infrastructure Maintenance and Monitoring
article

Deep learning–based individual segmentation and damage assessment of earthquake-damaged buildings: a case study of the old town of Beichuan county earthquake ruins

Zongpan Bian, Chuanjin Liu, Hongbin Cheng, Hongtao Tang, Zhuoli Bai, Jianhua Tang, Liangyu Zhu, Lingyun Ji, Shuai Kang, Jing Xu
article en

Abstract

Introduction Addressing the technical demand for rapid and precise building segmentation and intelligent damage assessment after earthquakes, this study proposes a collaborative method based on improved YOLOv8 for individual building segmentation and damage assessment. Methods Using UAV imagery of the Beichuan old town ruins, we build a dual_task sample database and separately train segmentation and classification models via network improvements, under an integrated framework of image segmentation, individual extraction, and multi_view assessment. The segmentation model extracts building footprints and combines with DSM to build 3D constraint boxes for isolating individual buildings. We then use five_direction virtual camera imaging and the most severe damage criterion for automatic damage grading. Results Validation via high_accuracy 3D models and field surveys yields: (1) Under 300_epoch progressive training, both models converge stably. building.pt achieves mAP50(B)=0.934 and mAP50_95(B)=0.803, and assessment.pt attains Accuracy of 0.64, Precision of 0.58, Recall of 0.75, and F1_Score of 0.65, respectively, demonstrating effective handling of irregular building outlines and complex damage morphologies (2) A total of 64 target buildings are present in the study area, of which 61 are successfully identified and segmented by the algorithm, yielding an overall recognition success rate of 89.1% (= 57/64 × 100%), a segmentation accuracy of 92.2% (= 59/64 × 100%), and a classification accuracy of 93.4% (= 57/61 × 100%). Discussion These results confirm that the proposed method can accurately perform individual building extraction and multi_view damage assessment.Regional analysis indicates that the Wangjiayan landslide west of the town is the main cause, with its frontal accumulation burying structures and impact loads progressively damaging central and peripheral buildings, creating a graded pattern of burial, impact, and disturbance, making the old urban area the worst-hit zone.

Frontiers in Remote SensingVol. 7
Seismological Bureau of Shanghai (CN), Beijing Seismological Bureau (CN), China Earthquake Administration (CN)
National Natural Science Foundation of China, Natural Science Foundation of Shaanxi Province, China Earthquake Administration, National Key Research and Development Program of China
Sustainable cities and communities, Industry, innovation and infrastructure
Infrastructure Maintenance and Monitoring
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon