Semantic segmentation of imagery is critical for remote sensing feature extraction and image generation, but faces challenges like microscopic structure recognition in high-resolution images and ambiguous semantic boundaries in complex scenes—leading to suboptimal precision, detail loss, and insufficient frequency-spatial feature integration. To address these, we propose a multiscale spatio-temporal boundary-aware semantic segmentation framework inspired by biological neural responses, consisting of three core modules: dual-path feature extraction (using spiking convolution kernels, DCT multi-band enhancement, and cross-domain RGB concatenation for info complementarity), multi-scale feature pyramid fusion (via bidirectional cross-scale interaction to integrate high-level semantics and low-level details), and boundary perception (three-stage architecture to extract edge heatmaps and reinforce weak boundaries). Experiments on WHU Satellite Datasets I/II and Qinqiang Dataset show superior performance over existing methods, effectively capturing boundary details in dense built-up and small-target regions while mitigating the “coarse shape” issue.