The radiosity is one of the popular algorithms for radiation simulation within a virtual canopy. However, owing to the complexity of plant architecture, enormous computation on computing the form factors has become a serious burden. Thus, a new optimized strategy of radiosity based on CUDA and 3D voxel traversal was developed to improve computation efficiency. Taking simulation of radiation transfer within a virtual Loquat canopy as an example, our proposal is based on uniform partition of bounding box and voxels traversing along a 3D line to identify occlusion between the light source and tree model facets, combining GPU to compute form factor in parallel using CUDA. Furthermore, we adopted reduction algorithm and shared memory to optimize the radiation flux calculation. Compared with serial implementation on CPU, the results are good in terms of execution times with speed-ups about 150. By comparative analysis with ray tracing and traditional radiosity model (progressive refinement radiosity algorithm), the simulation results of PAR distribution are similar and consistent. The results of comparison show that the new method not only improve computation efficiency, but also insure the accuracy.