Abstract The 2018 natural disaster in Palu, Indonesia, presented significant challenges in identifying damaged buildings at scale. Assessing structural integrity and reconstruction feasibility manually is a slow, labor-intensive process. This research presents an automated object recognition system utilizing the Scale Invariant Feature Transform (SIFT) algorithm implemented in Python. By comparing pre- and post-disaster imagery, the system identifies damage levels with high precision. Experimental results demonstrate that the SIFT-based approach achieves an average accuracy of 98.24% in distinguishing building states, providing a robust tool for government agencies and search-and-rescue teams.
Introduction: The Challenge of Rapid Assessment
In the immediate aftermath of a major earthquake or tsunami, rapid situational awareness is critical. For organizations like Indonesia's National Search and Rescue Agency, the sheer volume of destroyed infrastructure makes manual inspection a bottleneck. Traditional image comparison methods often fail due to variations in camera angle, lighting, and scale between historical photos and post-disaster aerial or ground-level shots. This study explores the application of SIFT to overcome these geometric and photometric variances, enabling automated damage detection.
Methodology
The proposed system leverages the SIFT algorithm to detect and describe local features in images. Unlike simple pixel-to-pixel comparisons, SIFT identifies "keypoints" that are invariant to image scaling and rotation.
The workflow consists of several primary stages as outlined in the system architecture:
- Feature Detection: Utilizing the Difference of Gaussians (DoG) to locate stable points of interest.
- Feature Description: Generating descriptors based on local gradient orientations.
- Feature Matching: Identifying correspondences between the "before" and "after" image sets.
- Dimensional Analysis: Measuring changes in the object's pixel dimensions (height and width) to detect structural collapse or shifting.
- Damage Quantification: Calculating a final damage level based on the ratio of successful matches and dimensional errors.
Experimental Setup and Dataset
The system was tested using a curated dataset of JPEG images from the 2018 Palu disaster. Five major landmarks were analyzed: the Globe Monument, Mercure Hotel, Roa-roa Hotel, Tatura Mall, and Sya Hotel. These sites provided diverse examples of structural damage, from partial degradation to total collapse.
Key Insights and Results
The SIFT algorithm proved exceptionally capable of handling the chaotic visual environment of a disaster zone.
Feature Matching and Accuracy
The system achieved an impressive average accuracy of 98.24%. A critical finding was the behavior of feature matches in destroyed structures. For heavily damaged buildings, the percentage of successful feature matches dropped to between 0.83% and 2.28%, providing a clear mathematical threshold for catastrophic failure.
The "Rubble Effect"
Interestingly, the research noted that damage does not always lead to a decrease in detected keypoints. In several instances, the number of keypoints increased significantly post-disaster. This is attributed to the complex textures created by debris and rubble, which generate numerous edges and corners that the algorithm detects as features, even if they no longer match the original structure's descriptors.
Dimensional Distortion
By incorporating "Width Error" and "Height Error" metrics, the system could quantify physical deformation. For example, the Roa-roa Hotel—which collapsed entirely—showed a 23.79% Width Error, highlighting how structural shifts can be used as a secondary confirmation of damage.
Conclusion and Future Directions
The Python-based SIFT implementation offers a high-accuracy, automated alternative to manual building inspections. While highly effective on the tested landmarks, future iterations should focus on scaling the system to handle larger, more diverse datasets including residential areas. Additionally, integrating this logic into real-time drone or mobile platforms would significantly enhance its utility for field-based rescue operations.
Summary updated