3D semantic segmentation network for post-disaster assessment with unmanned aerial vehicles

Natural disasters are occurring with increasing frequency and severity, highlighting the need for rapid and accurate damage assessment. Traditional assessment methods based on 2D ground imagery often lack the spatial context necessary for precise evaluation. To overcome this limitation, we developed an Enhanced Fast Point Transformer (eFPT), a Transformer-based 3D semantic segmentation network. eFPT builds upon the original Fast Point Transformer by introducing a refined position encoding mechanism with learnable multipliers and bias terms to enable the model to better capture complex spatial relationships in 3D point clouds. To evaluate the effectiveness of eFPT, we constructed a novel 3D benchmark dataset specifically designed for postdisaster assessment. The dataset is built from unmanned aerial vehicles (UAVs)-captured aerial imagery and reconstructed using Structure-from-Motion and Multi-View Stereo techniques. Experimental results show that eFPT outperforms the original FPT in key metrics, including mean class accuracy (mAcc) and mean intersection over union (mIoU). Our contributions advance the state of 3D semantic segmentation for disaster response and lay the groundwork for integrating AR/VR and digital twin technologies for enhanced situational awareness and emergency planning.