Background: In most existing EEG-based emotion recognition studies, the spatial configuration of electroencephalography (EEG) electrodes is not adequately preserved, and conventional single-scale convolutional neural networks (CNNs) have limited ability to capture both local and global spatial patterns. To address these limitations, an EEG emotion recognition framework integrating azimuthal equidistant projection (AEP) with a multiscale residual temporal network (MRTNet) is proposed. Methods: Differential entropy (DE) features are first extracted, and the three-dimensional electrode coordinates are projected onto a two-dimensional plane to construct AEP-based DE (AEP-DE) feature maps that preserve the spatial topology of the electrode layout. MRTNet then combines multiscale depthwise separable convolutions with a long short-term memory (LSTM) network to extract spatial features at multiple receptive-field scales and model temporal dependencies across EEG segments. Results: Experimental results show that MRTNet achieves average accuracies of 95.1% for four-class classification on the Database for Emotion Analysis using Physiological Signals (DEAP) and 93.1% for three-class classification on the SJTU Emotion EEG Dataset (SEED). Conclusions: Grad-CAM visualizations further indicate that the model responses are concentrated in spatially localized frontal and temporal regions, providing qualitative support for the spatial interpretability of the proposed method.