Dynamical encoding models characterize neural activity with low-dimensional hidden states that dynamically evolve in time and gienerate behavior. Current methods have identified these models from single-scale activity, either spikes or fields. However, behavior is simultaneously encoded across multiple spatiotemporal scales of activity, from spikes of individual neurons to neural population activity measured through fields. Identifying a multiscale dynamical model to extract hidden states that simultaneously describe spike-field activities is challenging because of their fundamental differences. Spikes are binary-valued with fast millisecond time-scales while fields are continuous-valued with slower time-scales. Here, we develop a novel multiscale dynamical modeling and identification algorithm to simultaneously characterize multiscale spike-field dynamics and extract multiscale hidden states. We also devise a modal approach to dissociate task-relevant and task-irrelevant dynamics. Using extensive simulations, we show that the algorithm accurately identifies a multiscale dynamical model to simultaneously describe spike-field dynamics. Furthermore, the algorithm extracts hidden states that are multiscale, i.e., contain information from both spikes and fields and accurately predict behavior. Finally, the algorithm detects which of the identified dynamics are task-relevant and to what extent. This multiscale dynamical modeling and identification framework can help study neural dynamics across spatiotemporal scales and may facilitate future neurotechnologies.