Meta-classification of remote sensing reflectance to estimate trophic status of inland and nearshore waters

Abstract

Common aquatic remote sensing algorithms estimate the trophic state (TS) of inland and nearshore waters through the inversion of remote sensing reflectance (Rrs ()) into chlorophyll-a (chla) concentration. In this study we present a novel method that directly inverts Rrs () into TS without prior chla retrieval. To successfully cope with the optical diversity of inland and nearshore waters the proposed method stacks supervised classification algorithms and combines them through meta-learning. We demonstrate the developed methodology using the waveband configuration of the Sentinel-3 Ocean and Land Colour Instrument on 49 globally distributed inland and nearshore waters (567 observations). To assess the performance of the developed approach, we compare the results with TS derived through optical water type (OWT) switching of chla retrieval algorithms. Meta-classification of TS was on average 6.75% more accurate than TS derived via OWT switching of chla algorithms. The presented method achieved 90% classification accuracies for eutrophic and hypereutrophic waters and was 12% more accurate for oligotrophic waters than derived through OWT chla retrieval. However, mesotrophic waters were estimated with lower accuracy from both our developed method and through OWT chla retrieval (52.17% and 46.34%, respectively), highlighting the need for improved base algorithms for low - moderate biomass waters. Misclassified observations were characterised by highly absorbing and/or scattering optical properties for which we propose adaptations to our classification strategy.