Station-centric geographically weighted regression (SCGWR): a novel framework for chlorophyll-a prediction with Sentinel-2

Chlorophyll-a (Chl-a) is a key indicator of phytoplankton biomass and serves as an essential parameter for evaluating marine pollution. As an optically active constituent (OAC), its spatial distribution can be effectively monitored using multi-spectral satellite imagery, which captures variations in water-leaving reflectance associated with phytoplankton dynamics. While traditional predictive modeling approaches, including machine learning techniques, primarily use reflectance spectra, this study advances the analysis by incorporating both spectral and spatial characteristics through Geographically Weighted Regression (GWR). However, GWR’s local modeling framework often presents limitations, including increased computational demands and challenges in unbiased accuracy evaluation. To overcome these issues, this study proposes a novel Station-Centric GWR (SCGWR) approach. SCGWR constructs individual GWR models for each in-situ measurement station, capturing localized relationships between spectral bands and Chl-a concentrations. The predictive performance of SCGWR was evaluated in the Gulf of İzmit, using Sentinel-2 A Level-2 A surface reflectance data and concurrent in-situ Chl-a measurements. The results demonstrated that SCGWR significantly enhanced the balance between model simplicity and sensitivity to data variations compared to both Multiple Linear Regression (MLR) and GWR. This improvement led to greater accuracy and robustness in predictive modeling of spatial relationships, as reflected in substantially lower RMSE values for SCGWR (0.41–0.57 for 2021–2023) compared to MLR (1.51–30.97) and GWR (1.47–2.75), together with higher correlation and better agreement with in-situ measurements. Qualitative analysis also confirmed that SCGWR generated more spatially consistent Chl-a’s categorical distributions, closely aligned with in-situ measurements. These findings underline the potential of SCGWR as a robust spatial predictive modelling framework for enhancing remote sensing-based assessment of coastal water quality.