Due to the inherent volume conduction effect, EEG-based motor imagery (MI) decoding is limited by the low spatial resolution of scalp recordings. To improve MI decoding, we propose a brain atlas-constrained EEG source-space spatial-temporal cross-fusion framework assisted by functional near-infrared spectroscopy (fNIRS). In the source-imaging stage, task-relevant spatial priors derived from fNIRS are used to regularize EEG source imaging, yielding more reliable source-space signals. In the spatial-temporal fusion stage, a lightweight temporal branch is constructed to capture multi-scale temporal dynamics. Spatially, ROI nodes are defined under a brain-atlas constraint, and a correlation-based functional connectivity graph is built among nodes. To better model MI-related spatial laterality and inter-regional cooperation, a spatial branch that combines approximate personalized propagation (APPNP) and a graph attention network (GAT) is employed to learn long-range dependencies and adaptive connection weights. Finally, a cross-attention fusion module is used to enable dynamic interactions between temporal features and spatial node representations. The proposed method achieves an accuracy of 72.81% ± 7.23 on Dataset 1 and 34.97% ± 5.39 on Dataset 2, outperforming representative EEG–fNIRS multimodal baselines. The code is available at https://github.com/useflf/MI-ESI.