DeepGlint S2: A Deep Learning Method for Sun Glint Correction of Atmospherically Corrected Sentinel 2 Imagery

Kurzfassung

The specular reflection of sunlight on the water surface, i.e., sun glint, can vary strongly from pixel to pixel in optical imagery, and its intensity may be comparable to or even exceed the water leaving radiance. This greatly increases measurement uncertainty and errors in retrieved biophysical parameters, making sun‑glint correction essential for accurate retrieval. Here, we introduce DeepGlint S2, a neural network–based method for sun glint correction of atmospherically corrected Sentinel 2 imagery. The training data for DeepGlint S2 are generated based on a widely tested three component radiance model implemented in the WASI software for simulating water surface reflections. The key parameter for estimating the sun glint contribution to the remote sensing reflectance (Rrs) is gdd, which represents the fraction of sky radiance originating from direct solar radiation. WASI and its recently AI enhanced module (WASI AI) invert gdd simultaneously with other fit parameters (e.g., water constituents). However, this inversion is image specific and requires careful parametrization of the physical model for the given image. To overcome this limitation, DeepGlint S2 is trained on a dataset consisting of 11k samples of Sentinel 2 Rrs spectra paired with their corresponding gdd values, derived from WASI AI processing of 11 images acquired over diverse inland and coastal waters. The transferability of the model is evaluated using more than 38k samples extracted from three Sentinel 2 images collected over two previously unseen water bodies. The results demonstrate strong transferability and generalization of the proposed DeepGlint S2, yielding R² = 0.93 and NRMSD < 4% when comparing its gdd estimates with those from image‑specific WASI‑AI processing.