Assessment of Feature Selection Methods for Machine Learning-Based Chlorophyll-a Retrieval Across Optical Water Types

Accurate retrieval of Chlorophyll-a (Chla, mg m−3) from ocean color reflectance remains a challenging issue due to spectral redundancy, nonlinear optical interactions, and water type variability. This study develops and evaluates a globally representative feature selection (FS) and machine learning (ML) framework to improve Chla estimation from multispectral reflectance. Using a quality-controlled in situ global dataset aggregated to Medium Resolution Imaging Spectrometer (MERIS) bands, we evaluate seven FS methods and five ML architectures across four optical water types (OWTs). The corresponding FS-ML models for each OWT are trained and validated based on partitioned subsets of in situ data using a novel data-partitioning scheme, ‘Spatially Blocked Stratified Monte-Carlo Split’. A three-stage model evaluation and a robustness filter are utilized. Importance-driven FS methods consistently produce compact, physically interpretable predictor sets and yielded the best generalization. Robust FS–ML combinations achieve high validation performance with minimal training–validation gaps. Cross-sensor transfer tests indicate that the MERIS-trained models could be generalized to independent Moderate Resolution Imaging Spectroradiometer (MODIS) and GlobColour matchups. Distributional uncertainty diagnostics further characterize retrieval confidence spatially and temporally. Overall, the OWT-adaptive FS provides a physically grounded, statistically robust, and operationally scalable approach for global Chla retrieval from multispectral ocean color reflectances.