Abstract

AbstractPrimary productivity expressed as the abundance of phytoplankton measured by the chlorophyll-a concentration (Chl-a), and water clarity in terms of suspended particulate matter, are considered as key indicators defining the water quality of any aquatic system. To maintain the good water quality, it is important to continuously monitor the spatio-temporal variability of these key indicators. Optical satellite remote sensing techniques in the visible spectral range are well known for their cost-effectiveness in estimating the water quality features on sufficient spatial and temporal scale with better radiometry. To overcome that, level 1C images from the multi-spectral instrument (MSI) onboard Sentinel 2 (S2), a medium to high resolution satellite sensor, were used in the present study. Even though there has been a radical improvement in the development of semi-analytical optical algorithms especially using band ratio methods, they need accurate spectral and specific absorption characteristics which are challenging to obtain for many inland water bodies. Machine learning algorithms, on the other hand can statistically derive the spatio-temporal distribution of chlorophyll-a and suspended matter from explicit optical relationships without the complexities of conventional empirical or semi-analytical algorithms. In this study, the best suitable machine learning (ML) algorithm using S2-MSI data to retrieve (Chl-a) and total suspended matter (TSM) for tropical lakes and inland waters were identified from the available machine learning models. The ML prediction models were trained using the surface reflectance together with the vegetation and water indices that are sensitive to Chl-a and TSM obtained from Sentinel-2 data. In situ Chl-a values for validation of the machine learning models were obtained from multiple field surveys conducted along the inland water bodies (Vellar river in Tamilnadu, and Paleru and Karedu inland tributaries of Krishna River in Andhra Pradesh) and a tropical coastal region (Palk Bay) in the south east coast of India (Palk Bay). From the validation analysis it was evident that Support Vector Machine (SVM) performed better in deriving the Chl-a (R2 = 0.81; RMSE = 0.19) and Radom Forest (RF) model performed better in modeling TSM distribution along the studied water bodies (R2 = 0.98; RMSE = 1.46). Validation of ML-based models for optically different water bodies proved the efficiency of the SVM and RF models in estimating the optical constituents in inland water bodies and tropical coastal waters with optical complexities from mixed composition of water constituents. The capability of medium resolution satellite like Sentinel 2 can hence provide means to establish tools to monitor the biophysical conditions of small inland water system effectively when coupled with machine learning methods.KeywordsWater qualityRemote sensingSentinel-2Machine learningChlorophyll-aSuspended matter

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call