Hybrid dual-transformer pansharpening network for enhanced spatial-spectral fidelity

Pansharpening plays an important role in improving the spatial resolution of multispectral images while preserving their spectral information. It enables more detailed and accurate analysis in various applications, such as remote sensing and environmental monitoring. Recent advances in deep learning-based pansharpening models have resulted in substantial improvements in performance. However, these models still suffer from the balancing of spectral accuracy and spatial detail, which can lead to artifacts, quality degradation, and overfitting problems. To overcome these limitations, an efficient pansharpening model is proposed. Initially, a dual transformer block is designed which integrates Swin and DeiT transformers to improve both local and global feature extraction. These features are then processed through a proposed U-shaped encoder–decoder network. This network utilizes the dual transformer block in both encoding and decoding stages. Finally, a customized multi-aspect pansharpening loss (MAPL) is introduced to preserve spectral fidelity, enhance spatial resolution, and improve perceptual quality. Extensive experimental results demonstrate that the proposed model significantly outperforms competitive models on various performance metrics. Thus, compared to competitive models, the proposed model shows significant improvements in preserving fine spatial details and maintaining spectral accuracy.