A Machine Learning-Based Regional Hybrid Model for Remote Retrieving Turbidity From Landsat Imagery

Turbidity [nephelometric turbidity unit (NTU)] monitoring is of great interest to water quality stakeholders. Traditional monitoring programs are limited in time and space, are expensive, and do not reflect the true extent of NTU. In contrast, remote sensing data are able to model the NTU, to monitor its spatial expansion, and are cost-effective. Models developed are usually a single-based function. This study presents a simple machine learning-based Regional hybrid model (R-HM) for NTU retrieval. The R-HM allows prior recognition of the NTU level concentration (high or low) before estimation. The calibration step highlighted that low and high NTUs are sensitive to different spectral regions, but mainly controlled by the red part. Validation was satisfactory with
$R^{2} = 0.99$
, although high NTUs tend to be underestimated (BIAS = −14%). Landsat (LS) NTU products derived from R-HM were found to be only sensitive to turbidity, even under conditions of high algal blooms.