Three-dimensional confocal Raman spectroscopic imaging is widely used in materials science and biomedicine due to its unique molecular selectivity and non-destructive, in-situ analytical capabilities. However, its inherent limitation of inadequate geometric resolution hinders the development of more advanced applications. In this study, a super-resolution mapping method that employs a Residual Dense Channel Attention Network （RDCAN） to transform 3D confocal Raman spectral images into geometric topography is proposed. An end-to-end deep learning model is designed based on a Residual Dense Network （RDN）, which effectively integrates both local and global multi-scale features. The model incorporates a channel attention mechanism to dynamically optimize feature weight allocation, thereby enhancing feature extraction capability. A self-developed laser differential confocal Raman spectroscopy system is used for data acquisition. In this system, 3D spectral data obtained by the confocal Raman module serve as input to the network, while the geometric topography data captured by the differential confocal module serve as ground truth for training. The reconstruction of geometric topography is ultimately realized. Experimental results show that the reconstructed geometric topography achieves a peak signal-to-noise ratio（PSNR） of 27.4dB and a structural similarity index （SSIM） of 0.98, closely matching the measurements obtained by differential confocal microscopy. Compared with traditional methods, the proposed approach enhances high-resolution geometric imaging based on conventional 3D confocal Raman spectroscopy without requiring additional hardware. It also enables spectral-geometric homologous fusion, offering an innovative solution to overcome the limitations in geometric resolution.