An Improved Pan-Sharpening Method Based on Convolutional Autoencoder and Guided Filter (CAE-GF) on Quickbird Images

Pan-sharpening is the process of combining high-resolution panchromatic and low-resolution multispectral images to generate a high-resolution multispectral image. However, some pan-sharpening methods have poor preservation of spectral and spatial information. In this context, the aim of this study is to improve the pan-sharpening method based on a convolutional autoencoder and a guided filter (CAE-GF), by developing a new Convolutional Autoencoder (CAE) network. A CAE network was designed and trained to generate original panchromatic images from their spatially degraded versions. The trained network was used to improve the spatial details of the intensity component of the multispectral image obtained through the adaptive intensity-hue-saturation (AIHS) method. The pan-sharpening process is achieved by applying the multi-scale guided filter to improve the original PAN image using the enhanced intensity component and by injecting the detail map into the multispectral image. To analyze the preservation of the spectral and spatial information in the proposed method, full-reference and no-reference quality indices were calculated, and a visual analysis was performed. These analyses were compared with traditional pan-sharpening methods of component substitution. The results showed the developed method’s efficacy in generating Pan-sharpened QuickBird images.