Deep internal learning for single SWIR satellite image super resolution

The compact dimensions of CubeSats limit the optical equipment they can carry, which in turn affects the spatial resolution of the images they capture. BGUSAT, a 3U CubeSat, gathers Short Wave Infra-Red (SWIR) images between 1.55-1.7 micrometers with a spatial resolution of 600 meters per pixel. Leveraging deep learning techniques for enhancing satellite imagery resolution, particularly for CubeSats like BGUSAT, offers significant improvements of data analysis for remote sensing applications. Traditional deep learning super-resolution algorithms require a large amount of training data. However, if there is a shortage of available satellite imagery or if a single existing image requires enhancement through super-resolution techniques, it may not be sufficient to rely on traditional methods. Single image super-resolution methods, such as bicubic interpolation, do not consider the complexity of features within the image and provide very limited enhancement results. Satellite imagery characteristics vary significantly across sensors, altitudes, and spectral bands. Pre-trained models in supervised learning may yield inaccurate predictions with new sensor data. Thus, a self-supervised method, Zero Shot Super Resolution (ZSSR), which focuses on the unique internal features of each image to extract latent information, was adopted. Our proposed approach using the ZSSR algorithm operates without reference data, ensuring high-quality, image-specific data enhancement using a single image. A single BGUSAT image was super-resolved using our approach, with scale factors ranging from 2 to 9. several evaluation methods were applied to compare the quality of super-resolved images using ZSSR against traditional bicubic interpolation: visual interpretation, and three non-reference evaluation methods.