The sound field in coupled rooms or rooms with non-uniform absorptive material distributions can be considerably anisotropic. In such scenarios, the sound energy decays with more than one decay rate, thus making it practical to use a decay model that consists of multiple exponential decays and a noise term. In this work, we use a recently proposed neural-network-based approach for estimating the underlying model parameters from sound energy decay curves. Introducing a spatial filter bank allows for a directionally constrained analysis of anisotropic late reverberation, resulting in a set of multi-exponential decays with corresponding decay parameters. Our work shows that the proposed analysis framework is suitable for modelling anisotropic sound fields with multi-exponential decays and that it can be used for challenging acoustic problems, such as the denoising of spatial room impulse responses.