Abstract: Seismic multi-attribute analysis enables a multi-dimensional interpretation of subsurface structures. However, conventional point-to-point fusion methods often fail to capture their continuous nature. To address this limitation, we propose a novel multi-scale fusion method based on multi-scale decomposition and convolutional sparse representation. First, a Gaussian-Laplacian pyramid is constructed in the frequency domain to achieve multi-scale representation of the selected seismic attributes with high sensitivity to the target layer. Then, the detail layers are fused using convolutional sparse representation, while the base layer is integrated via a gradient-based fusion algorithm. Finally, the fused Gaussian-Laplacian pyramid undergoes inverse transformation to reconstruct the integrated attribute. While preserving high-frequency details, this approach effectively decomposes individual attributes across multi-scale spatial-frequency bands. It enables a comprehensive characterization of geological targets from both multi-scale and global perspectives, and plays a key role in accurately depicting reservoir distribution. Tests on an overthrust structural model demonstrate its successful integration of global features across input data and its distinct advantages in delineating channel boundaries and reservoir distribution. A field data application to a beach reservoir yields a multi-attribute fusion outcome highly consistent with drilling results. The outcome aligns well with established geological understanding and demonstrates promising application performance.