An Effective Deep Learning Model for Pan-Sharpening of Satellite Images

Image fusion techniques are widely used to enhance images by combining two or more remote sensing images. The fusion task of "pan-sharpening" is to merge low resolution Multispectral (MS) and High resolution Panchromatic (PAN) satellite images of the same scene obtained by the same satellite. This paper presents proposed an effective deep learning model leveraging a combination of novel techniques for feature enhancement and aggregation. The proposed model named as Efficient Non-local Feature Enhancement Network (ENFE-Net) integrates the PAN guided band-aware feature enhancement module with an Efficient Non-local Attention (ENLA) mechanism and Spectral Aggregation Module (SpecAM). The PAN guided band-aware feature enhancement module facilitates effective feature extraction, leverages PAN features to conduct band-aware multi-spectral feature modulation, selectively enhancing the information of each spectral band. Additionally, the integration of the ENLA mechanism enables the model to capture similar contextual dependencies in the input data efficiently, enhancing its discriminative power. Furthermore, the SpecAM is employed to aggregate spectral information effectively, improving the model’s effectiveness to adjust the spectral information. Performance of proposed ENFE-Net model is evaluated on PAirMax datasets and demonstrate its superior performance compared to existing traditional and recent deep learning methods. Experimental results of proposed ENFE-Net model show significant improvements over existing pan-sharpening methods.