Object detection in disaster images is often difficult because of class imbalance and a lack of labeled samples. To solve this problem, we introduce Semantic-Guided Generative Augmentation (SGGA), a new method that uses semantic masks and textual prompts to generate more samples for the rare classes. SGGA creates new images by changing clean road areas into Road-Blocked areas using mask-based sampling and prompt-guided inpainting, making sure the new objects appear in the right places. We filter the new images using CLIP similarity and LPIPS distance, ensuring high semantic and visual quality. Experiments on the RescueNet dataset show that SGGA improves Road-Blocked detection by +26.2% mAP@0.5 and +29.7% recall, beating other augmentation methods. Cross-domain validation on FloodNet further demonstrates SGGA’s generalizability across disaster scenarios. Furthermore, t-SNE analysis confirms strong semantic alignment between real and SGGA-generated images. SGGA offers significant advantages, including spatial precision, contextual realism, and low annotation overhead, making it particularly suitable for practical deployment in disaster scenarios and other domains where spatial priors are available.
