Post-Earthquake Candidate Building-Loss Screening from Bitemporal Google Earth Imagery: An Object-Level Framework for the 2023 T¨urkiye Earthquakes

When a major earthquake strikes, the immediate priority for rescue teams is to know where buildings have collapsed. In the chaotic hours and days following a disaster, engineers cannot physically visit every street to inspect structures; the roads may be blocked, and the danger is too high. Instead, they rely on images taken from above, looking down at the city like a map. For decades, scientists have tried to teach computers to look at these images and spot the difference between a standing house and a pile of rubble. The challenge is that these images are rarely perfect. They are taken from different angles, at different times of day, and often show the same street in different seasons. A building might look different simply because the sun is in a different spot, not because it has been destroyed. Furthermore, many regions lack detailed digital maps of where every house is located, making it difficult for computers to know what they are looking for in the first place.

This difficulty is exactly what a new study from Turkey aims to solve. The researchers focused on the devastating earthquakes that struck Türkiye in February 2023. Their goal was to create a system that could automatically scan historical satellite images from before the quake and compare them with images taken after, using only the pictures themselves. They did not rely on pre-existing maps of building locations or complex 3D data that might not be available in a crisis. Instead, they built a method to find the buildings first, and then check if they were still there later. The result is a tool that acts as a first filter, highlighting areas where buildings seem to have vanished so that human experts can focus their detailed inspections on the most likely spots.

To build this system, the researchers first had to teach a computer how to recognize buildings in satellite photos. They gathered a massive collection of images, combining a public dataset of satellite photos with 150 specific pairs of images from Google Earth taken over the affected Turkish cities. One set of these images was from 2022, before the disaster, and the other was from 2024, after the destruction. Because the computer needed to learn what a building looks like, the researchers used a semi-automatic process. They let an initial computer program mark the buildings, and then humans carefully checked every single mark, correcting mistakes and adding buildings the computer missed. This created a high-quality training set where the computer could learn to identify houses, apartments, and other structures even when they were small or crowded together.

Once the training data was ready, the team tested six different types of computer vision models to see which one was best at finding buildings. These models are like different sets of eyes, each with its own way of processing visual information. Some are designed to be very fast, while others are designed to be extremely precise. The researchers ran them all under the same conditions to see which one offered the best balance between speed and accuracy. They found that a model called YOLOv12n performed the best. It was able to find the most buildings while keeping the number of false alarms low, and it did so quickly enough to be useful for large areas. After selecting this winner, they retrained it on the entire collection of images to make it even sharper. When tested on new, unseen images, this final model was able to correctly identify buildings with a high degree of reliability, though it still missed some smaller or heavily damaged structures, which is a known limitation when dealing with such complex scenes.

With the best building detector in hand, the researchers moved to the second stage: comparing the before and after images. This step was tricky because the two images of the same place were not perfectly aligned; one might be slightly shifted or rotated compared to the other. To fix this, the system ignored the buildings themselves, which might have been destroyed, and instead locked onto stable features like roads, intersections, and the edges of fields. Using these fixed points, the computer adjusted the 2024 image to match the 2022 image perfectly. Once the images were aligned, the system looked at every building it had found in 2022 and tried to find its twin in the 2024 image.

The results of this comparison were striking. The system processed 145 pairs of images covering various districts in the hardest-hit regions. In the 2022 images, it identified 2,872 buildings. In the 2024 images, it found 2,039 buildings. When it tried to match them up, it successfully linked 842 buildings that appeared in both years. However, 2,030 buildings that were present in 2022 had no matching counterpart in 2024. The system flagged these 2,030 missing buildings as "candidate losses." This does not mean they were definitely destroyed by the earthquake; it simply means they disappeared from the view between the two dates. They could have been demolished, rebuilt, or perhaps the computer just missed them in the second image. The system is designed to catch all these possibilities so that humans can look closer.

The study found that the rate of these candidate losses varied significantly by location. In the district of Antakya, for example, the system flagged 715 buildings as missing, which represented over 86% of the buildings it had seen there before the quake. In Kahramanmaraş, the central district showed a loss rate of nearly 79%. Other areas had lower rates, but the pattern was clear: the system could pinpoint exactly where the most dramatic changes had occurred. The researchers also noted that the buildings flagged as missing were not just the ones the computer was unsure about; many had high confidence scores, meaning the system was quite certain it had seen them before and was certain it did not see them after.

This work demonstrates that it is possible to conduct a large-scale preliminary screening of earthquake damage using only publicly available images, without needing specialized maps or expensive sensors. The system does not claim to know for sure which buildings collapsed; rather, it provides a prioritized list of locations that need human attention. By narrowing down the search from thousands of buildings to a few thousand likely candidates, the method saves time and resources for rescue teams and engineers. It turns a massive, overwhelming task into a manageable one, allowing experts to focus their energy where the evidence of loss is strongest. The study concludes that while the computer cannot replace the human eye, it can serve as a powerful guide, pointing the way to the most critical areas in the aftermath of a disaster.