A transferable Sentinel-2 Salinity Water Index (S2SWI) for saline lakes

Authors:
Preprint
Publisher website
Google Scholar
Abstract
Saline lakes exhibit complex spectral behaviour due to the coexistence of shallow brine, wet sediments and evaporitic crusts. Instead of a clearly defined shoreline, these environments exhibit hydro-evaporitic transition zones in which surface states vary gradually, producing continuous spectral variations within individual pixels. As a consequence, conventional water and salinity indices often show limited transferability across saline systems with contrasting surface conditions. This study proposes a Sentinel-2 Salinity Water Index (S2SWI) derived from optimised spectral band combinations. Index coefficients were derived using three different approaches: Particle Swarm Optimisation (PSO), Ridge Regression (RR), and a hybrid U-Net + Ridge Regression (UNET+RR) framework. Reference surface classes (water, wet saline margin and crust) were generated through Random Forest classification. The index was developed using three saline lakes within a Leave-One-Lake-Out (LOLO) validation framework to promote spectral formulations that generalize across different environments. Its transferability was subsequently evaluated in three independent lakes by generating S2SWI maps and applying an unsupervised K-means clustering approach; the results were compared with Random Forest (RF) reference classifications. Results indicate that the S2SWI provides a consistent ordering of surface conditions across the six lakes, ranging from liquid hypersaline water to dry evaporitic crusts. Among the three formulations evaluated, the hybrid UNET+RR approach showed the strongest overall transferability, particularly in spectrally challenging systems characterized by extensive transition zones.The proposed S2SWI offers a computationally simple and transferable tool for Sentinel-2 monitoring of saline lakes. Moreover, the comparison between optimization, regression, and hybrid spatial-spectral approaches helps clarify the environmental conditions under which spectral indices remain robust, as well as those in which spatial information can improve the interpretation of surface conditions.

This publication has 43 references indexed in Scilit: