Improving seaweed cover estimation in high latitudes: a focus on kelp classification

Abstract. Satellite remote sensing offers the most practical means of monitoring Macrocystis pyrifera kelp forests across the remote Strait of Magellan, but standard atmospheric correction workflows assume a plane-parallel atmosphere that becomes inaccurate at the high solar zenith angles (SZA) characteristic of subantarctic latitudes, introducing wavelength-dependent radiometric biases into spectral unmixing products. This study develops and validates a Kelp Fractional Cover (KFC) product for the full 70,000 km² extent of the Strait of Magellan by applying an empirical SZA correction to Sentinel-2 Level-1C imagery prior to endmember extraction, using bilinear spectral unmixing implemented through the R-based Endmember Composition Algorithm (RECA) and Google Earth Engine. Pure pixel spectra for kelp, water, and land were extracted from over a decade of cloud-filtered imagery across ten sampling zones, yielding 51,002 pure kelp pixels, and endmembers were applied via a pseudoinverse unmixing model to generate per-pixel fractional cover with an associated RMSE layer. Validation against 60 independently paired, tidally and temporally controlled SuperDove image comparisons across ten additional zones showed strong model performance, with a peak Mean Jaccard Index of 0.50, an area-based R² of 0.69, and a polygon-based Overall Accuracy of 0.94, converging within an optimal binarization threshold range of 0.27–0.35. A second analysis comparing corrected and uncorrected models in optically deep, kelp-free reference zones confirmed that uncorrected fractional cover estimates exhibited spurious seasonal oscillations in kelp and land fractions tracking the annual SZA cycle, an artifact that the correction substantially suppressed in both deep-water and coastal validation zones. These results demonstrate that accounting for SZA-driven radiometric distortion is a prerequisite, rather than an optional refinement, for producing temporally consistent kelp cover estimates at high latitudes, and that failing to do so risks misattributing illumination-geometry artifacts to genuine ecological change. The resulting decade-long KFC time series, publicly available at https://doi.org/10.5281/zenodo.21877772 (Iribarren et al., 2026), provides a spatially consistent, regionally continuous monitoring tool for one of the most extensive and least disturbed kelp forest systems on Earth, supporting future integration with in situ and environmental covariate data.