Akademik Veri Yönetim Sistemi
INTERNATIONAL JOURNAL OF REMOTE SENSING, 2025 (SCI-Expanded)
Earthquakes are one of the natural disasters that are common in the world and profoundly influence human life. On 6 February 2023, eleven provinces in T & uuml;rkiye were most affected by two major earthquakes, the Mw 7.7 event in Pazarc & imath;k (Kahramanmara & scedil;) and the Mw 7.6 event in Elbistan (Kahramanmara & scedil;). Over 50,000 lives were lost, and more than 84,000 buildings were damaged due to the earthquakes. An efficient post-disaster intervention necessitates the interpretation of information from various data sources. Satellite images are among these sources, offering detailed insights into the region impacted by the earthquake before and after the event. Despite the extensive detail in satellite images making human interpretation challenging, deep learning architectures have been developed to address this issue. These advanced models utilize sophisticated computer vision techniques for the semantic segmentation of satellite imagery. However, training these models requires a time-consuming process due to the need for labelled images. This study tackled the issue of scarce labelled data in deep learning by using open-source building footprint data as training labels and conducted change detection between pre- and post-earthquake building conditions. The study focused on the city centre of Antakya (Hatay), which was heavily impacted by the Kahramanmara & scedil; earthquakes. Three different deep learning architectures, SegNet, PSPNet, and ResUNet, known for their effectiveness in semantic segmentation, were employed. Among the models evaluated on pre-earthquake images, ResUNet achieved the highest F1-score average at 92.27%, while SegNet and PSPNet followed with similar scores of 90.47% and 90.26%, respectively.