Emotion recognition based on electroencephalogram (EEG) is a key interdisciplinary research area bridging brain science and artificial intelligence. Despite advancements, the core challenge remains in extracting complex spatio-temporal information from high-dimensional 3D EEG topographies for precise, concurrent multi-dimensional emotion prediction. To address that challenge, we propose SFHTNet, a novel EEG emotion recognition framework featuring an innovative Spatial Feature Encoding Module (SFEM) and a Hierarchical Temporal Attention Module (HTAM). The framework takes 3D EEG topographies derived from Differential Entropy (DE) and Power Spectral Density (PSD) features as input. The SFEM uses multi-layer, multi-scale 3D convolutions to capture detailed spatial patterns of brain activity. For temporal analysis, the HTAM extracts hybrid temporal features, adaptively focusing on the most discriminative segments and dynamics from diverse sources for emotion classification. Experiments on proprietary and public datasets (SEED and DEAP) show SFHTNet outperforming methods by up to 4.5% in average classification accuracy. SFHTNet offers novel insights for efficient and precise EEG emotion recognition, with potential applications in mental health, intelligent education, and immersive human-computer interaction.