Monday, 29 January 2024: 4:45 PM
338 (The Baltimore Convention Center)
Super-resolution is the general objective of artificially increasing the spatial resolution of an imaging system. Recently, there have been many promising ML-based approaches for satellite image super-resolution and pan-sharpening. In this abstract, we present a convolutional neural network (CNN) we developed to super-resolve the 1-km and 2-km bands on the GOES-R series Advanced Baseline Imager (ABI) to a common high resolution of 0.5 km. The super-resolution of the lower-resolution bands is aided by the native 0.5-km observations from Band-2 available during the daytime only. We first train the CNN on a proxy task that involves degrading the resolution of ABI bands and training the CNN to restore the original imagery. Once trained, this model can be applied at ABI’s native resolution to super-resolve all channels to 0.5-km resolution. Access to native 0.5-km visible imagery from ABI Band-2 enables the CNN to realistically sharpen lower-resolution bands without significant blurring. Comparisons at reduced resolution and at full resolution with Landsat-8/9 observations illustrate that the CNN produces images with realistic high-frequency detail that is not present in a bicubic interpolation baseline. These comparisons include evaluations of pixel-wise error, overall spatial structure, the preservation of spectral relationships and the presence of accurate high-frequency detail. This general approach is readily extensible to other remote sensing instruments that have bands with different spatial resolutions and requires only a small amount of data and knowledge of each channel's spatial response.

