Learning low dimensional latent dynamics in population spiking and field potential activities together can reveal the relationship between different spatiotemporal scales of population activity and can improve performance of brain-machine interfaces (BMIs). But developing a multiscale learning algorithm is challenging because spikes are discrete-valued signals while field potentials are continuous-valued signals with slower time-scales than spikes. Recently, a multiscale learning method based on expectation maximization was developed (multiscale-EM) for multimodal discrete-continuous data, which maximizes the joint likelihood of spikes and fields iteratively, thus can be computationally expensive. In some applications, such as tracking neural plasticity in basic science studies or real-time learning in BMIs, a desirable feature for a learning method is computational efficiency. Therefore, it is important to develop a more efficient multiscale learning algorithm than the iterative multiscale-EM. Here, we develop a more efficient multiscale learning algorithm based on subspace identification (multiscale-SID) and validate it using numerical simulations and a non-human primate (NHP) neural dataset during a random target reaching task. We show that multiscale-SID accurately learns multiscale dynamical models for spiking and field potentials and extracts the low dimensional dynamics/modes. Also, multiscale-SID combines information across spiking and field potential population activity, allowing for more accurate identification of dynamics/modes compared to single-scale SID. Critically, we demonstrate that multiscale-SID is much faster than multiscale-EM while having similar or better accuracy when provided with enough training samples. Taken together, multiscale-SID provides a new tool for studying multimodal population dynamics across different spatiotemporal scales and for real-time learning of these dynamics.