Source-linked AI summary
Building Damage Detection in Satellite Imagery Using Convolutional Neural Networks
Joseph Z. Xu, Wenhan Lu, Zebo Li, Pranav Khaitan, Valeriya Zaytseva
TL;DR
Humanitarian response requires timely building-damage information, but manual satellite-image assessment is slow and labor-intensive. The paper builds CNN-based damage detectors, compares four architectures, and evaluates transfer across disasters. TTS performs best on Haiti, while cross-region performance improves with more diverse training data and small amounts of local fine-tuning.
Problem
Manual damage assessment is slow and labor-intensive, while future disasters may affect regions absent from training data, motivating generalizable automated detection.
Method
The paper constructs a three-disaster dataset, automatically generates undamaged examples, compares four CNN architectures, and tests cross-region training and fine-tuning settings.
Results
Twin-tower subtraction achieved the best Haiti result with 0.8302 validation AUC, while cross-region AUC improved with two-dataset training and local fine-tuning.
Takeaways & Limitations
The model can generalize to new regions and disasters when fine-tuned on a small set of examples from the target region.
Takeaways & Limitations
Future work is needed to cover additional disaster types and improve robustness to data flaws such as pre/post-image misalignment.
Abstract
from arXiv · showhide
In all types of disasters, from earthquakes to armed conflicts, aid workers need accurate and timely data such as damage to buildings and population displacement to mount an effective response. Remote sensing provides this data at an unprecedented scale, but extracting operationalizable information from satellite images is slow and labor-intensive. In this work, we use machine learning to automate the detection of building damage in satellite imagery. We compare the performance of four different convolutional neural network models in detecting damaged buildings in the 2010 Haiti earthquake. We also quantify how well the models will generalize to future disasters by training and testing models on different disaster events.
1 Introduction
Humanitarian agencies need rapid, reliable information on affected populations, but manual damage mapping is slow and labor-intensive. This paper investigates CNN generalizability for detecting damaged buildings and evaluates cross-region transfer because future disasters may affect unseen regions.
- Motivation: Manual digitization of damaged structures is reliable but labor-intensive, unsuitable for large areas, and prone to fatigue-related inconsistencies.Automating damage assessment could reduce the time required to produce reports.
- Motivation: Remote sensing offers wide-area coverage for identifying damaged buildings, which serve as a proxy for locating affected populations.
- Background: Deep neural networks have achieved strong performance on computer-vision tasks and are suitable for automatically extracting information from satellite images.
- Related work: Prior damage-detection studies used machine learning with pre- and post-event imagery or post-event imagery, including a 78.6% overall accuracy for collapsed-building identification in Haiti.
- Motivation: Cross-region transfer learning matters because future disasters may affect regions absent from the training data.
- Contribution: The paper compares CNN architectures on one dataset and then evaluates the best architecture across different transfer-learning contexts to investigate generalizability.
2 Data Generation Pipeline
The authors construct a multi-disaster dataset because no comprehensive dataset was available, combining pre- and post-disaster imagery with automated negative-example generation and standardized patches.
- Dataset scope: The dataset spans the 2010 Haiti, 2017 Mexico City, and 2018 Indonesia disasters.
- Obtain Satellite Images: Satellite images were collected before and after each disaster and resampled to 0.3 meter resolution for consistent pixel scales.
- Identify Undamaged Buildings: Undamaged examples were generated by detecting all buildings and filtering out those marked damaged by UNOSAT analysts.The building detector had precision 0.64 and recall 0.75 at confidence threshold 0.5.
- Labeling: Damage labels were simplified by grouping “Severe Damage” and “Destroyed” into a single “Damaged” class.
3 Model Architecture Comparisons
Four AlexNet-style CNN variants differ in how they combine pre- and post-disaster imagery. On Haiti, twin-tower models outperformed single-tower models, with twin-tower subtraction performing best.
- Single-Tower Models: CC concatenates pre- and post-disaster RGB channels, whereas PO uses only the post-disaster image.
- Twin-Tower Models: TTC and TTS process pre- and post-disaster images through separate feature extractors before combining them by concatenation or element-wise subtraction.
- Architecture Overview: All four models use convolutional layers, fully connected layers, and a sigmoid output, while differing in pre/post-image processing.
- Evaluation: AUC was used with 5-fold cross-validation because it is robust to class imbalance and more indicative of model quality than conventional accuracy.
- Results: 0.8302 validation AUC was achieved by TTS, the best-performing architecture; twin-tower models outperformed single-tower models.
- Results: The authors used TTS in all subsequent experiments.
4 Cross-Region Generalization
The study evaluates whether building-damage models generalize across disasters and regions when training data have limited variability. Cross-region performance improves with more diverse training data and with local fine-tuning.
- Motivation: Limited variation in building characteristics, lighting, terrain, image quality, and camera angles can cause models to overfit and perform poorly out-of-sample.The paper frames generalization to future disasters as a practical requirement for damage detection.
- Experimental conditions: The experiments compare same-dataset evaluation, Haiti-only training with testing on other datasets, two-dataset training with third-dataset testing, and local fine-tuning.The same-dataset condition provides a best-case baseline, while Haiti-only training represents a low-diversity setting.
- Evaluation design: Longitude-based folds reduce the chance that nearby buildings from training and test sets create information leakage.Each fold contains buildings within a longitude interval rather than a random assignment.
- Results: Same-dataset evaluation achieves the highest AUCs, while training on one dataset and testing on another produces poor AUC.The authors attribute the same-dataset advantage to greater consistency in building characteristics, image quality, and annotation standards.
- Results: Training on two datasets improves cross-region AUC, and the best cross-region results follow fine-tuning on a small neighborhood.The paper suggests that training-data region variability helps even when it does not come from the test region.
5 Conclusion
The paper presents a CNN-based pipeline for automated building-damage detection and identifies the two-tower subtract model as its best architecture. It reports generalization to new regions and disasters after fine-tuning on a small regional sample, while leaving broader disaster types and robustness to data flaws for future work.
- Contributions: The pipeline automatically generates large numbers of negative training examples and uses CNNs to detect damaged buildings in satellite images.This summarizes the paper’s stated method and data-generation contribution.
- Contributions: The two tower subtract variant performs best among the evaluated model architectures.
- Generalization: The model can generalize to new regions and disasters when fine-tuned on a small set of examples from the target region.
- Future work: Future work includes additional disaster types and methods for robustness to flaws such as pre- and post-disaster image misalignment.The authors specifically mention hurricanes, armed conflicts, and random translations as examples.
Appendix A
The appendix includes figures showing pre- and post-disaster imagery and a sample building-detection model output on Haiti.
- Figure 3: Figure 3 presents pre-disaster images on the left and post-disaster images on the right for Haiti, Mexico City, and Indonesia earthquakes.
- Figure 4: Figure 4 shows a sample output from the building detection model on Haiti.