Combining physical modelling and AI for removing sun glint from atmospherically corrected imagery

Abstract

Specular reflections of the sun on the water surface (sunglint) can be of comparable intensity or even much higher than the water leaving radiance, even if observation in the direct direction of the sun’s specular reflection is avoided. Apart from a perfectly plane water surface, ripples and waves can reflect light from the sun in sensor direction, with a probability that increases with wind speed and in the direction of the reflected sun. The combination of a bio-optical model of aquatic reflectance and an analytic model of the downwelling irradiance, implemented for more than ten years in the publicly available WASI software, has long been shown to be well suited to correct sunglint from above-water field spectrometer measurements and atmospherically corrected multispectral satellite imagery. However, inverse modelling of each individual pixel is too computationally intensive for operational image processing. Since the variability within an image is governed by only a few environmental parameters, it is justified to apply the physical modelling only to a small subset of representative image pixels and process the entire image with a statistical method based on the inversion results of the physical model, for example a neural network. The new artificial intelligence module WASI-AI implements such a method in WASI. Results of the application of WASI-AI for sunglint correction are presented for a number of multispectral (Sentinel-2, Landsat-8/9) and hyperspectral (DESIS, EnMAP) images.