<p>Decoding emotional states from electroencephalogram (EEG) signals is challenging due to the complex and nonlinear nature of brain activity. To address this, we propose a novel multiscale convolution-spike coupling network with Haar wavelet transform, termed CSCN-WT. The framework leverages Haar wavelet decomposition to enhance informative components and reduce signal non-stationarity, while integrating convolutional and spiking neural networks to exploit their complementary strengths in spatial and temporal feature learning. A spatiotemporal coupling module is introduced to enable multi-level fusion of spatial features and wavelet-decomposed subbands through interactive integration, thereby improving the model’s ability to capture discriminative time-frequency patterns with enhanced temporal resolution and frequency localization. Additionally, an adaptive weighted multi-path loss is designed to guide parameter optimization, enabling the network to dynamically adapt to EEG data distribution and varying task demands. Experiments are conducted on both a collected dataset (UEED) and widely used public datasets (SEED, SEED-IV, DEAP). Results show that CSCN-WT achieves state-of-the-art performance, demonstrating its effectiveness.</p>