Spatiotemporal Fourier neural operator-empowered super-resolution traffic flow field reconstruction from sparse observations

An ideal traffic state reconstruction (TSR) framework should be able to generate traffic state values at any resolution level, thus achieving super-resolution. However, most existing studies are limited to fixed resolutions, which greatly limits the applicability of TSR methods in real-world scenarios. To overcome this bottleneck, this paper proposes a super-resolution TSR framework based on the spatiotemporal Fourier Neural Operator, named STFNO. The central concept of this framework is the development of mapping operators between functions within the frequency domain based on STFNO. This approach overcomes the constraints inherent in conventional pointwise function learning techniques, thereby facilitating the comprehension of fundamental universal principles governing traffic flow. The framework consists of two core tasks: (1) Resolution-fixed TSR: the model takes spatiotemporal coordinates as input and outputs traffic state values of the corresponding resolution; (2) Zero-shot super-resolution TSR: the model is trained with low-resolution traffic state data and then directly inferred based on high-resolution spatiotemporal coordinates to achieve zero-shot super-resolution reconstruction of traffic state. This paper conducts extensive experiments on multiple datasets at different spatiotemporal scales. In the resolution-fixed TSR task, STFNO effectively reconstructing traffic field evolution using sparse observations, with advantages in accuracy and a good balance between memory usage and computational efficiency. In the zero-shot super-resolution task, the powerful generalization capability of STFNO is fully demonstrated from three perspectives: temporal generalization, spatial generalization, and joint spatiotemporal generalization.