Image deraining aims to eliminate precipitation-induced degradation and reconstruct physically consistent visual representations. Existing methods show progress in spatial-domain modeling but remain limited by single-domain frameworks that cannot resolve rain degradation’s dual challenge involving simultaneous frequency-domain energy imbalance and spatial structural distortion. This coupling inherently forces a trade-off between global noise suppression and local detail preservation. Through Fourier spectrum analysis, we observe that rain artifacts concentrate in mid-high frequency amplitude spectra while structural information remains intact in phase spectra. To bridge this gap, we propose MambaSpectra, a dual-domain network integrating frequency-domain decoupling and spatial-domain optimization. The core Frequency-Aware Dual State Space Block employs Fourier transform in its frequency branch to isolate global degradation features and utilizes Mamba-based selective scanning in the spatial branch for long-range dependency modeling. The Dynamic Gated Cross-Conv Module achieves adaptive kernel deformation guided by GRU-driven rain direction continuity, eliminating spatial information loss from pooling operations while enhancing local detail recovery. Extensive experiments demonstrate our method’s superior performance across multiple metrics compared to state-of-the-art models.