Evaluation of SRGAN Algorithm for Superresolution of Satellite Imagery on Different Sensors

Abstract

Abstract. In recent years, deep learning has quickly evolved to be the go-to solution for any kind of analysis of non-linear data. One such use has been that of Generative Adversarial Networks (GAN) in the field of Computer Vision. GAN models have a variety of applications for image processing, specifically, super-resolution of images. A lot of work has been done to enhance or upscale generic RGB imagery such as the ones taken from a mobile or digital camera. However, in the field of remote sensing, it presents challenges like preserving the spatial resolution of the sensor, which is affected by a wider pixel value range and relation of a pixel to ground sampling distance (GSD). From data preparation to enhancing a complete set of tiles at scale, the upsampling/downsampling requires the ratio of number pixels to the actual area in geography to be preserved. SRGAN model has been proven to be effective for interpolating the pixels based on context. However, it was observed that the same algorithm with or without parameter tuning behaves differently based on the sensor source and target resolution. We evaluate the performance of the model from 10m to 2.5m and 2.4m to 0.6m resolution. The comparison will enable better decision making when using the enhanced images for LULC classification, segmentation, and object detection.