DWT-3DRec: DeepJSCC-based wireless transmission for efficient 3D scene reconstruction using CityNeRF

The Unmanned Aerial Vehicle (UAV)-assisted sensing–transmission–computing integrated system plays a vital role in emergency rescue scenarios involving damaged infrastructure. To tackle the challenges of data transmission and enable timely rescue decision-making, we propose DWT-3DRec—an efficient wireless transmission model for 3D scene reconstruction. This model leverages MobileNetV2 to extract image and pose features, which are transmitted through a Dual-path Adaptive Noise Modulation network (DANM). Moreover, we introduce the Gumbel Channel Masking Module (GCMM), which enhances feature extraction and improves reconstruction reliability by mitigating the effects of dynamic noise. At the ground receiver, the Multi-scale Deep Source–Channel Coding for 3D Reconstruction (MDS-3DRecon) framework integrates Deep Joint Source-Channel Coding (DeepJSCC) with Cityscale Neural Radiance Fields (CityNeRF). It adopts a progressive close-view training strategy and incorporates an Adaptive Fusion Module (AFM) to achieve high-precision scene reconstruction. Experimental results demonstrate that DWT-3DRec significantly outperforms the Joint Photographic Experts Group (JPEG) standard in transmitting image and pose data, achieving an average loss as low as 0.0323 and exhibiting strong robustness across a Signal-to-Noise Ratio (SNR) range of 5–20 dB. In large-scale 3D scene reconstruction tasks, MDS-3DRecon surpasses Multum in Parvo Neural Radiance Fields (Mip-NeRF) and Bungee Neural Radiance Field (BungeeNeRF), achieving a Peak Signal-to-Noise Ratio (PSNR) of 24.921 dB and a reconstruction loss of 0.188. Ablation studies further confirm the essential roles of GCMM, DANM, and AFM in enabling high-fidelity 3D reconstruction.