Pan-sharpening reconstructs high-resolution multispectral （HRMS） images by integrating the spatial details of panchromatic （PAN） images and the spectral information of multispectral （MS） images. However， achieving a balance between spatial detail enhancement and spectral fidelity remains challenging due to the complex spatial-spectral relationships and frequency characteristics of PAN and MS images. This study proposes a pan-sharpening method based on Spectral Swin Transformer and frequency enhancement to improve spatial-spectral feature representation and reconstruction quality. The proposed method enhances the interaction between heterogeneous PAN and MS features， captures spatial structures and spectral correlations， and reduces spectral distortion during detail reconstruction. Experimental evaluations demonstrate that the proposed framework improves fusion performance by jointly optimizing spatial detail recovery and spectral information preservation.An end-to-end deep learning framework is developed with an enhanced dual-branch U-Net backbone （EUB）， a Spectral Swin Transformer module （SSM）， and a frequency enhancement module （FEM）. The proposed framework achieves effective fusion of PAN spatial information and MS spectral information through modality-specific feature extraction， spatial-spectral interaction modeling， and frequency-aware reconstruction. The EUB adopts a dual-branch encoder-decoder architecture， where PAN and MS images are processed through independent encoding pathways. The PAN branch focuses on extracting spatial structures， edge information， and texture details， while the MS branch preserves spectral distributions and inter-band characteristics. This design enables heterogeneous inputs to be represented according to their intrinsic characteristics and reduces feature interference before fusion. During decoding， multi-scale features from different levels are progressively integrated to combine shallow spatial details with deep semantic information for HRMS image reconstruction. To enhance spatial-spectral feature representation， the SSM is introduced to separately model spatial contextual dependencies and spectral-related feature interactions. The spatial relation modeling branch adopts the shifted-window attention mechanism to capture local and cross-window spatial relationships， improving the representation of complex structures and texture variations. The spectral relation modeling branch further learns correlations among deep feature channels through cross-channel interaction， thereby enhancing spectral information exchange and maintaining spectral consistency during spatial detail enhancement. The joint modeling of spatial and spectral relationships improves reconstruction capability for complex scenes while preserving spectral information. Furthermore， the FEM is designed to optimize frequency information during the reconstruction process. The input features are adaptively decomposed into high-frequency and low-frequency components， and their contributions are dynamically adjusted according to image characteristics. High-frequency components are enhanced to recover missing spatial details， whereas low-frequency components are preserved to maintain structural information and spectral stability. The reconstructed frequency features are further integrated with spatial-spectral representations to achieve a better balance between detail enhancement and spectral preservation. A composite loss function containing spatial reconstruction constraints， frequency consistency constraints， and spectral similarity constraints is adopted to guide network optimization. The proposed model is implemented using PyTorch and evaluated on four public satellite datasets， including WorldView-2， WorldView-3， GeoEye-1， and Pléiades-1.The proposed method is evaluated on four public remote sensing datasets， including WorldView-2， WorldView-3， GeoEye-1， and Pléiades-1， and compared with representative traditional and deep learning-based pan-sharpening methods. Multiple full-reference metrics， including correlation coefficient （CC）， spectral angle mapper （SAM）， relative dimensionless global error in synthesis （ERGAS）， quality index （Q4/Q8）， and peak signal-to-noise ratio （PSNR）， are used for quantitative evaluation. The experimental results demonstrate that the proposed method achieves stable and competitive performance across different satellite datasets. In terms of spatial fidelity， the proposed method achieves competitive CC and Q4/Q8 values on most datasets， indicating improved preservation of spatial structures and image details. Regarding spectral consistency， the proposed method achieves lower SAM and ERGAS values on multiple datasets， demonstrating its capability to reduce spectral distortion during detail enhancement. The PSNR results further indicate that the proposed method provides improved overall reconstruction quality compared with existing approaches. Visual comparisons show that the proposed method produces clearer edges， richer textures， and more consistent spectral representations， particularly in complex scenes with dense buildings， roads， and heterogeneous land covers. Residual analysis reveals more uniform error distributions， indicating stable reconstruction performance in different regions. Spectral response analysis further confirms that the proposed method better maintains the original spectral characteristics of multispectral images. Ablation experiments verify the effectiveness of each component， demonstrating that the enhanced dual-branch backbone improves heterogeneous feature extraction， the Spectral Swin Transformer module strengthens spatial-spectral feature interaction， and the frequency enhancement module contributes to detail recovery and spectral consistency. In addition， evaluations using both reference-based and non-reference indicators show that the proposed method maintains a favorable balance between spatial enhancement and spectral preservation under different sensor conditions. The results on multiple datasets demonstrate that the proposed framework has strong robustness when processing remote sensing images with different spatial resolutions， spectral characteristics， and scene complexities.The proposed framework demonstrates reliable generalization ability across different satellite datasets and provides an effective approach for high-quality HRMS image reconstruction in remote sensing applications. The results indicate that the proposed approach has strong adaptability for remote sensing image fusion tasks involving complex spatial structures and diverse spectral characteristics.