Author: Saut Sagala, Dekka Putra, Benaya Ginting (Intern)
Disasters are volatile, and despite efforts in disaster risk reduction, they still occur. Traditional damage assessment methods are often slow, delaying resource allocation and infrastructure recovery. Rapid and accurate evaluations are crucial for effective response and rebuilding—advanced techniques like deep learning address these challenges. Deep learning, a subset of machine learning, uses artificial neural networks to analyse complex data for automated damage detection and classification (Mustafa et al., 2024). Frameworks such as ChangeOS integrate object-based image analysis and remote sensing techniques, leveraging high-resolution satellite and drone imagery to accurately assess damage and localise buildings (Zheng et al., 2021).
Figure 1. Framework of the ChangeOS Method
Source: Zheng et al., 2021
Advanced Damage Assessment Methods
As demonstrated in the research by Zheng et al. (2021), ChangeOS was applied to assess building damage from the 2020 Beirut Port Explosion and the 2021 Bata Military Barracks Explosion. This framework combines the semantic consistency of conventional object-based image analysis (OBIA) with the robust feature representation of Siamese fully convolutional networks (FCN). Using satellite images from the xBD dataset (WorldView-2 and GeoEye-1), ChangeOS achieved significant accuracy improvements over other frameworks like UNet+ResNet. For densely populated areas like Beirut, pre-disaster building localisation achieved an F1 score of 64.5%, while post-disaster building damage classification reached 46.9%. The framework processed data efficiently, demonstrating its effectiveness even in less densely populated areas with low-rise buildings, such as Bata (Zheng et al., 2021).
Despite its strengths, ChangeOS faces challenges, particularly with low-quality imagery often resulting from catastrophic events. Furthermore, it excels at assessing human-induced disasters but requires integration with broader hazard databases to assess climate-related disasters effectively.
Another promising approach is crowdsourcing and social media data for early damage assessments. By integrating crowdsourced data with frameworks like PAGER (Prompt Assessment of Global Earthquakes for Response) by USGS, these methods provide near-real-time loss estimations. Recent advancements with Large Language Models (LLMs) have enhanced the interpretation of multilingual and noisy data, as demonstrated in the 2021 Haiti Earthquake (Wang et al., 2024). This integration improved data accuracy and timeliness, reducing misinformation and enabling faster disaster response.
Source: Wang et al., 2024
- Critical Infrastructure Resilience
High-resolution damage assessments can inform the repair and retrofitting of critical infrastructure, including hospitals, schools, transportation networks, and power grids. For instance, detailed damage maps generated through ChangeOS or UAV-based imagery can prioritise resources for repairing essential facilities, ensuring minimal disruption to vital services. - Infrastructure Risk Mapping
AI and Machine learning-powered frameworks can identify vulnerabilities in infrastructure before disasters occur, allowing for proactive measures. Combining pre-disaster assessments with hazard maps clarifies areas requiring enhanced resilience measures, such as seismic retrofitting or flood-proofing. - Rapid Post-Disaster Reconstruction
Damage and loss assessments provide critical data for planning reconstruction efforts. AI-powered systems can simulate reconstruction scenarios, optimise the allocation of materials, and ensure that rebuilt infrastructure adheres to higher safety and sustainability standards. An example is Japan’s use of advanced algorithms following the 2011 Tohoku Tsunami. Soil stability analyses guided infrastructure rebuilding in Japan, mitigating future liquefaction risks (Cong & Inazumi, 2024).
Challenges and Opportunities for Infrastructure-Centric Tools
While advanced tools like ChangeOS and crowdsourced frameworks enhance damage assessment, their application to infrastructure requires addressing certain challenges:
- Data Availability and Quality: High-resolution satellite or drone imagery is essential for detailed assessments, but such data may be limited in rural or disaster-struck areas.
- Integration with Existing Systems: Infrastructure assessments must integrate with national disaster management frameworks to ensure actionable insights. For instance, combining AI models with Indonesia's InaTEWS could enhance tsunami early warning systems and infrastructure protection.
- Scalability: Large-scale assessments for urban and rural infrastructure demand high computational resources, necessitating cloud-based solutions to ensure scalability.
In conclusion, computational tools for damage and loss assessments offer the transformative potential to enhance disaster management. By addressing vulnerabilities in infrastructure and incorporating AI-driven methodologies, governments and institutions can build resilience against future hazards, ensuring efficient recovery and long-term sustainability.
Table 1 Comparison of each Method Framework
Aspect | Crowdsourced Rapid Assessments | Deep learning Satellite Images Rapid Assessments |
---|---|---|
Objective and Scope | Near-real-time estimation of fatalities in earthquakes for rapid disaster response. | Assessing building damage after disasters using deep learning remote sensing images. |
Key Technologies | Few-shot learning, Large Language Models (LLMs), Bayesian updating, and truth discovery algorithms. | Deep learning, Siamese fully convolutional networks (FCNs), and object-based image analysis (OBIA). |
Data Sources | Crowdsourced multilingual social media and existing rapid assessment systems such as PAGER | High Spatial Resolution (HSR) bitemporal satellite images. |
Strengths | Automates casualty estimation with timeliness and accuracy handles multilingual and noisy data, and integrates with PAGER. | Achieves instance-level accuracy, combines localisation and damage classification, and is suitable for global and local scales. |
Limitations | Relies on the accuracy and availability of crowdsourced data, which is susceptible to misinformation. | It requires high-quality pre- and post-disaster imagery and is computationally intensive for large datasets. |
Building Resilience with Computational Methodology
As a global think tank, the Resilience Development Initiative (RDI) collaborated with GIZ to evaluate frameworks for calculating damage and loss due to climate change in Southeast Asia. This includes assessing the damage and impact of critical infrastructure and identifying strategies to strengthen it against future disasters. Computational methodology will develop faster assessment due to automated data collection and processing using satellite imagery, drones, machine learning, remote sensing, and geospatial data. However, initial costs can be high, as they require advanced technology, software, and specialized training. However, they can save costs in the long term due to faster assessments and better resource management for Damage and Loss Assessment.
Aligned with the Smart System Centre (SSC) roadmap, this initiative emphasises integrating AI-driven tools into infrastructure resilience planning. By combining damage assessments with infrastructure data, RDI is working toward a more systematic approach to disaster risk reduction and climate adaptation. Advanced methodologies like ChangeOS and crowdsourced frameworks are essential for timely recovery, improved resource allocation, and building safer, more sustainable infrastructure.
References
Cong, S., & Inazumi, T. (2024). Post-Tsunami Infrastructure Recovery: Lessons from the 2011 Tohoku Tsunami. Journal of Disaster Resilience Studies, 18(2), 110–123.
Mustafa, R., et al. (2024). Deep Learning in Disaster Damage Assessment: Current Practices and Future Directions. Disaster Science and Technology, 12(1), 45–62.
USGS. (2017). PAGER: Prompt Assessment of Global Earthquakes for Response. U.S. Geological Survey.
Wang, L., et al. (2024). Integrating Crowdsourcing with LLMs for Real-Time Earthquake Damage Assessment. Earth Science Informatics, 15(3), 210–234.
Zheng, Q., et al. (2021). ChangeOS: A Framework for Object-Based Disaster Damage Assessment Using Satellite Imagery. Remote Sensing Letters, 12(6), 456–472.