Transform low resolution, low-quality satellite images into high-resolution, high-quality satellite images

BAE Systems GXP^® has developed an advanced Artificial Intelligence (AI) model, Super Resolution and Quality (SRQ), designed to enhance satellite images through super-resolution, de-noising, and de-clouding techniques. Satellite images often suffer from low-quality due to the high altitudes at which they are captured, with cloud cover posing a significant challenge. The SRQ model addresses these issues by improving both image resolution and quality, which are critical for various geospatial intelligence applications. High-quality, high-resolution images, including aerial imagery, are utilized to train the SRQ model. To simulate low-quality and low-resolution satellite images, artificial noise and thin clouds are introduced into the ground truth training images. The training process involves minimizing the differences between the ground truth images and their degraded counterparts. This paper presents the use cases and test results that validate the effectiveness of the SRQ model. To train the SRQ model, GXP employs three preprocessing algorithms: (1) scale augmentation, (2) pixel value normalization, and (3) training tile randomization. The training dataset comprises 1170 images, including satellite panchromatic images from various sensors, satellite multi-spectral images, and aerial frame images in both panchromatic and color formats. Some of these training images are significantly large, resulting in a total of over 2 million training image tiles, each measuring 512 x 512 pixels. The ground sample distances (GSD) of the training images range from 0.20 meters to 0.60 meters. The SRQ model is capable of generating high-resolution, highquality images from both satellite (panchromatic and multi-spectral) and aerial (panchromatic and color) images. The training process, conducted on a high-end desktop workstation equipped with three NVIDIA RTX 6000 GPUs, spans six months and involves 60 epochs.