Machine Learning Methods for Water Surface Salinity Estimation in the Georgian Estuaries from Satellite-based Reflectance and Temperature Data

Keywords: coastal salinity, cdom, remote sensing, machine learning
Session Type: Virtual Paper Abstract
Day: Tuesday
Session Start / End Time: 3/1/2022 11:20 AM (Eastern Time (US & Canada)) - 3/1/2022 12:40 PM (Eastern Time (US & Canada))
Room: Virtual 67

Abstract

Water surface salinity is an important parameter to understand the biophysical processes within estuarine regions, tidal wetlands, and other habitats thriving in the transition zone between freshwater and oceanic waters. Especially in the coastal salt marshes water salinity is crucial to understand the flow of organic and inorganic carbon, nutrients sources and sinks, and species distribution. Remote sensing of water surface salinity is challenging due to the lack of a characteristic reflectance feature, voluminous and consistent in-situ salinity data, and limited spatial, spectral, and temporal resolutions of satellite sensors. This study proposes to overcome existing challenges in surface salinity estimation, especially for the estuarine region of coastal Georgia, by extensive machine learning-based modeling and validation on in-situ data and ultimately scaling it up to the satellite data. Our model will be trained and tested from temperature and salinity data collected within the Georgia Coastal Ecosystems (GCE-LTER) study domain near Sapelo Island, GA, and other locations near coastal Georgia from the R/V Savannah. Satellite data includes remote sensing reflectance (Rrs) from Sentinel-2 at 10m resolution and land surface temperature (LST) from Landsat-8 at 30m resolution. This study proposes to model surface salinity using Rrs, LST, spectral indices proxy for nutrient content such as NDCI, NDVI, absorption by Colour Dissolved Organic Matter (aCDOM) as variables. An ensemble approach would be used to investigate multiple machine learning models to estimate a surface salinity map of the study area. The best-performing model would be tested on similar estuarine sites using Sentinel-2, Landsat 8/9 data.