The increasing frequency of natural disasters necessitates rapid and reliable methods for assessing building damage to aid timely disaster response. This thesis investigates the potential for deep learning models to generalize across diverse geographical and environmental contexts for building damage detection, a critical requirement for models deployed in real-world scenarios. The primary objective is to evaluate biases within these models, specifically those that may limit performance in out-of-domain datasets, and to propose methodologies to enhance model robustness. This study leverages two datasets, xBD and Ida-BD, utilizing both in-domain and out-of-domain data to analyze model biases. We introduce a novel Fusion Augmentation technique designed to enhance the model’s ability to capture building edges, thereby improving classification of damage levels, especially in regions with dense vegetation. A series of supervised and unsupervised domain adaptation techniques, including CORAL, were applied to improve model generalizability across varied disaster scenarios without requiring target labels. Grad-CAM visualization techniques further support explainability by offering insights into the areas of focus in model predictions. Results demonstrate that combining Fusion Augmentation with domain adaptation significantly improves model accuracy, especially in damage classes, and reduces location specific biases. This research contributes a practical framework for developing reliable and generalizable building damage detection models, which can serve as valuable tools in post-disaster assessment, ultimately supporting faster and more effective humanitarian responses.