Toward reliable post-disaster assessment: advancing building damage detection using you only look once convolutional neural network and satellite imagery

Natural disasters continuously threaten populations worldwide, with hydrometeorological events standing out due to their unpredictability, rapid onset, and significant destructive capacity. However, developing countries often face severe budgetary constraints and rely heavily on international support, limiting their ability to implement optimal disaster response strategies.

This study addresses these challenges by developing and implementing YOLOv8-based deep learning models trained on high-resolution satellite imagery from the Maxar GeoEye-1 satellite. Unlike prior studies, we introduce a manually labeled dataset, consisting of 1400 undamaged and 1200 damaged buildings, derived from pre- and post-Hurricane Maria imagery. This dataset has been publicly released, providing a benchmark for future disaster assessment research. Additionally, we conduct a systematic evaluation of optimization strategies, comparing SGD with momentum, RMSProp, Adam, AdaMax, NAdam, and AdamW.

Our results demonstrate that SGD with momentum outperforms Adam-based optimizers in training stability, convergence speed, and reliability across higher confidence thresholds, leading to more robust and consistent disaster damage predictions. To enhance usability, we propose deploying the trained model via a REST API, enabling real-time damage assessment with minimal computational resources, making it a low-cost, scalable tool for government agencies and humanitarian organizations. These findings contribute to machine learning-based disaster response, offering an efficient, cost-effective framework for large-scale damage assessment and reinforcing the importance of model selection, hyperparameter tuning, and optimization functions in critical real-world applications.