| Summary: | The spectroscopic capability of the satellite observation of ocean colour contributes to the estimation of the concentration of Chlorophyll-a (Chl-a) on the ocean surface. Chl-a can be a proxy in the determination of the phytoplankton biomass distribution, which indicates the trophic status of the water body. Long-term records of ocean colour data at lower spatial resolution of 1 km has been widely used in the derivation of various ocean colour algorithms. Although most of the algorithms perform well in clear water state, the significant uncertainty is evident when algaeprone areas near the coast and shallow water are mapped at the 1-km resolution. Therefore, the current study designed a methodology for new estimation of Chl-a and nutrient concentration in coastal water from multi-platform satellite imageries at medium spatial resolution (10 to 30 m) with systematic accuracy assessment using collocated sea-truth. In particular, Artificial Neural Network (ANN), Support Vector Machine (SVM), and Random Forest (RF) techniques were used to establish the complex relationships of collocated remote sensing reflectance from the consecutive Landsat 8 Operational Land Imager (OLI) and Sentinel 2 MultiSpectral Instrument (MSI) images and in-situ parameters. Using these machine learning methods, this study also demonstrated the estimation of nutrients (nitrate and phosphate). The radiometric resolution of OLI in this study allowed higher overall accuracy of Chl-a estimates in the West Johor Straits (WJS) water. Meanwhile, the ANN recorded higher accuracy of Chl-a and nitrate estimates than that of the SVM and RF variants. Using the ANN, the Chl-a estimates at lower root-mean-square error (RMSE < 6 mg/m3) and APD of lower than 35% were mapped. The regression between Chl-a and nutrients was remarkably low (R2 < 0.2) on OLI and MSI. However, Fine Tree RF and ANN models improved the precision (RMSE) of nitrate (< 12 µmol/L) and phosphate (< 3 µmol/L). The absence of direct relationships of optical properties and spectral characteristics with nutrients led to higher uncertainties (> 100%), and this made phosphate content estimates in shallow water dubious, resulting in the need for extensive in-situ validation. Machine learning offers powerful estimation capability on Chl-a and nutrient concentration, especially for the higher spatio-temporal variability optical parameter of coastal waters, which was successfully demonstrated in this study through the discussed application in WJS.
|
|---|