Collapsed Building Detection from Multi-Temporal Pléiades Imagery: Comparing CNN Architectures and Input Data Strategies
- 1Department of Geomatics, Faculty of Civil Engineering, Czech Technical University in Prague, Thákurova 7/2077, 166 29 Prague, Czech Republic
- 2School of Earth Sciences, D. Serikbayev East Kazakhstan Technical University, Ust-Kamenogorsk 070000, Kazakhstan
Abstract
A rapid and spatially explicit assessment of structural damage is essential following major disasters, particularly where access to the field is limited and reliable information on damage is scarce. This study evaluates how semantic segmentation architecture, temporal input design, and class formulation influence the detection of collapsed buildings in very high-resolution Pléiades imagery from Jindires, Syria, following the earthquakes of 6 February 2023. The U-Net, SegNet, DeepLabv3+, and fully convolutional network (FCN) models were compared together with a reference random forest model. Methodologically, this study employs a controlled factorial experiment designed to isolate the effects of using either post-event images alone or a combination of pre- and post-event imagery, across both class-based and binary label formulations. Performance was evaluated in a spatially separate validation region using Dice loss, overall accuracy, and pixel-level F1-score, supplemented by location-level evaluation for 102 visually interpreted sites of collapsed buildings. The FCN with multitemporal binary input achieved the highest F1-score at the pixel level (79.1%), followed by the U-Net network with multitemporal three-class input (76.0%) and post-event binary input (74.8%). At the site level, both the U-Net network with post-event binary input and the FCN with multitemporal binary input achieved an F1-score of 73.0%; the U-Net network prioritized recall, while the FCN prioritized precision. The results reveal that the optimal configuration depends heavily on the chosen architecture’s inductive biases. Therefore, these findings highlight the importance of selecting not only the right architecture, but also the appropriate temporal input and class formulation. The results provide reproducible information for screening; although they do not constitute a definitive inventory of damage, they offer a truly independent perspective on the consequences of the earthquake in a region from which verified data cannot be obtained.
Related
- Building resilient and sustainable project outcomes through mitigation strategies and project characteristics
- Antibiotic resistance genes in aquaculture: Molecular detection, environmental dynamics, and biological control strategies
- Optimising Survey Strategies for the Endangered Romer's Tree Frog (Liuixalus romeri): Effects of Diel, Seasonal and Weather Variation on Acoustic Detection
- Comparing the pristine biochar and Zn-biochar composite for stabilizing Cd in a contaminated calcareous soil
- Comparing soil-health- and species-based sensitivity distributions for metal risk assessment in Tianjin soils
BibTeX
@article{boucek2026collapsed,
title = {{Collapsed Building Detection from Multi-Temporal Pléiades Imagery: Comparing CNN Architectures and Input Data Strategies}},
author = {Tomáš Bouček and Ondřej Pešek and Karel Pavelka and Eva Matoušková and Marzhan Yessenbekovna Rakhymberdina and Zhanna Aleksandrovna Assylkhanova and Karel Pavelka},
journal = {Geomatics},
year = {2026},
month = oct,
doi = {10.3390/geomatics6050114},
url = {https://doi.org/10.3390/geomatics6050114},
note = {CC BY},
}
RIS
TY - JOUR
TI - Collapsed Building Detection from Multi-Temporal Pléiades Imagery: Comparing CNN Architectures and Input Data Strategies
AU - Tomáš Bouček
AU - Ondřej Pešek
AU - Karel Pavelka
AU - Eva Matoušková
AU - Marzhan Yessenbekovna Rakhymberdina
AU - Zhanna Aleksandrovna Assylkhanova
AU - Karel Pavelka
PY - 2026
DA - 2026/10/26
JO - Geomatics
DO - 10.3390/geomatics6050114
UR - https://doi.org/10.3390/geomatics6050114
AB - A rapid and spatially explicit assessment of structural damage is essential following major disasters, particularly where access to the field is limited and reliable information on damage is scarce. This study evaluates how semantic segmentation architecture, temporal input design, and class formulation influence the detection of collapsed buildings in very high-resolution Pléiades imagery from Jindires, Syria, following the earthquakes of 6 February 2023. The U-Net, SegNet, DeepLabv3+, and fully convolutional network (FCN) models were compared together with a reference random forest model. Methodologically, this study employs a controlled factorial experiment designed to isolate the effects of using either post-event images alone or a combination of pre- and post-event imagery, across both class-based and binary label formulations. Performance was evaluated in a spatially separate validation region using Dice loss, overall accuracy, and pixel-level F1-score, supplemented by location-level evaluation for 102 visually interpreted sites of collapsed buildings. The FCN with multitemporal binary input achieved the highest F1-score at the pixel level (79.1%), followed by the U-Net network with multitemporal three-class input (76.0%) and post-event binary input (74.8%). At the site level, both the U-Net network with post-event binary input and the FCN with multitemporal binary input achieved an F1-score of 73.0%; the U-Net network prioritized recall, while the FCN prioritized precision. The results reveal that the optimal configuration depends heavily on the chosen architecture’s inductive biases. Therefore, these findings highlight the importance of selecting not only the right architecture, but also the appropriate temporal input and class formulation. The results provide reproducible information for screening; although they do not constitute a definitive inventory of damage, they offer a truly independent perspective on the consequences of the earthquake in a region from which verified data cannot be obtained.
SN - 2673-7418
ER -