Assessment of Deep Learning Models Trained Using Global Remote Sensing Imagery in Real-Context Emergency Response

Remote sensing and deep learning have been integrated to solve multiple problems, including building damage assessment. With rapid development in both fields, deep learning and remote sensing can play a greater role in damage mapping, specifically in rapid damage assessment, to support emergency response efforts. Deep learning model evaluation is generally based on a statistical split separating training and testing sets of the same data distribution. Although this enables the evaluation of the model performance, this scheme does not disclose the ability of the model to perform in data obtained from different distributions, which is often the case in real-context disaster emergency response. This study evaluates the model generalization in emergency response scenarios. The results show that the current deep learning model has a high performance in in-domain testing yet experiences a drop of up to 53(%) in F₁ in realistic applications. Future studies should focus on enhancing the model transferability, including using domain adaptation techniques and harnessing multi-modal features.