Detecting and correcting sky and Sun Glint effects in Sentinel-2 images over tropical reservoir

Abstract

Sun and sky glint pose significant challenges in remote sensing when evaluating water quality via satellite images in aquatic systems. These glint effects raise the water-leaving radiance, resulting in an overestimation of biogeochemical and optical parameters. Since conventional techniques, mainly designed for ocean and coastal waters, are less effective with high-spectral-resolution satellite data like Sentinel-2/MSI, it is crucial to enhance methods for masking and correcting these effects in satellite products. This study developed a glint-detection classifier using machine learning-based models and two new spectral indices with Sentinel-2 MSI imagery. The proposed spectral indices are called Normalized Difference Glint Index (NDGI) and Normalized Difference Glint Index 2 (NDGI2), and they are based on deep-blue (443 nm) and shortwave infrared (> 1600 nm) ranges. The machine learning model used a Random Forest Classifier (RF) and was trained on a total of 10,000 samples, achieving an accuracy of approximately 80% for glint detection. A comparison of four different glint correction algorithms was conducted and compared with in-situ measurements. The glint algorithms included SWIR-Subtraction (SubSWIR), Atmospheric Correction for OLI ’lite’ (ACOLITE), POLYnomial-based algorithm applied to MERIS (POLYMER), and Sun Glint Removal of Sentinel-2-like images (GRS). ACOLITE and GRS performed the best, reducing MAPE by about 59.86%. Among the glint algorithms, ACOLITE demonstrated a strong ability to preserve regions free of glint in the Sentinel-2/MSI imagery. The combined use of the mask and glint-correction algorithms can enhance time-series analysis, providing an effective solution for inland water monitoring using satellite images and supporting various applications.

Publication
International Journal of Remote Sensing