Abstract: Most of the existing Synthetic Aperture Radar （SAR） image despeckling models have been derived from related generic vision fields and generally lack the capability to handle the spatially heterogeneous distribution characteristics of speckle noise. This paper proposes a Frequency-Driven Heterogeneous Model for SAR image despeckling， which explicitly separates high- and low-frequency features using pooling layers and processes them separately based on their distinct noise-distribution characteristics. First， a Dconv Cross-Covariance Attention mechanism was designed to model the spatial-channel global characteristics of high-frequency features， enhancing the representation of critical edge information and removing complex speckle effects. Then， a Large Kernel Convolutional Residual Module was introduced to effectively suppress the simple speckle interference in low-frequency features by leveraging the receptive field advantage. Finally， to strengthen the complementary interaction between frequency features and further improve the despeckling performance， a Selective Frequency Fusion Module was used to adaptively fuse the heterogeneously processed high- and low-frequency features with minimal parameters. The results of experiments on both simulated and real SAR image datasets demonstrated that the proposed model outperformed some existing despeckling methods in both quantitative metrics and visual quality. Compared to the suboptimal method， it achieved average PSNR and SSIM increases of 0.14 dB and 0.01， respectively， while the inference time was reduced by at least 99%. Compared to classical lightweight methods， it achieved average PSNR and SSIM increases of 0.78~1.12 dB and 0.02~0.06， respectively.