An Efficient Crop-Based Digital Twin Network Leveraging a Novel Texture-Enabled Neural Radiance Field

The Agricultural Digital Twin Network (ADTN) represents one of the key enabling technologies of Agriculture 4.0. However, the unstructured nature of agricultural scenarios poses significant challenges to the geometric fidelity and scalability of ADTNs, as the morphological diversity and entity uniqueness of crops further reduce the rendering accuracy of virtual entities and increase modeling costs. Existing research lacks low-cost solutions for digital twin generation in unstructured agricultural environments, as implicit-based models struggle to capture global invariant features critical to crop structure and texture. To address these challenges, a crop-based ADTN (CBDTN) framework is designed, which incorporates low-cost image acquisition terminals, edge devices, and cloud platforms, thereby optimizing resource utilization through implicit modeling and minimizing costs in resource constrained scenarios. Furthermore, to improve the accuracy of virtual entity generation, a texture prior feature enabled neural radiance field (TPFNeRF) is proposed, inspired by the human visual system. It emulates the processing mechanism of the visual cortex to extract consistent prior features from multi-view images, thereby providing robust rendering guidance that enhances perceptual fidelity in 3D reconstruction tasks. The experimental results show that the CBDTN with TPFNeRF exhibits superior performance in generating objects with complex textures and structures. In synthetic dataset, it significantly improves reconstruction quality, achieving 6.52 dB PSNR and 4.3% SSIM improvements over the baseline, and delivers overall performance superior to recent methods such as GFB-NeRF and DiSRNeRF. In real-world scenarios, TPFNeRF effectively reconstructs object geometry and color from fixed viewpoints using only 5 MB of model weights, significantly reducing CBDTN deployment costs.<p></p>