Evaluating Surface Water Salinity Indicators from Landsat-8 OLI Imagery Using Machine Learning

The Mekong Delta region is important to Vietnam’s economy, agriculture, and ecosystem. Salinity has significantly impacted the region’s living environment, livelihoods, and productive activities. Therefore, salinity prediction maps are important in agriculture, environmental management, water resource planning, and ecosystem conservation since they can provide information about the spatial distribution and variability of salinity levels in a specific area. Landsat 8 satellite image data allows for the extraction of many salinity indicators. However, not all these indicators are necessary for salinity prediction studies. This study aims to contribute to selecting suitable input indicators to enhance the accuracy and precision of salinity prediction modelling. In paper utilize advanced machine learning techniques, including Bayesian Model Averaging, Extreme Gradient Boosting, Bagging, and Random Forest, to select the salinity prediction model and evaluate the relative importance of salinity indicators. The results obtained from the XGBoost model indicate that 18 out of the 20 input variables in the first optimal model made a significant contribution. These variables include two coordinate variables, 12 time-related variables, and four variables derived from the Landsat 8 OLI images. The performance evaluation of the selected salinity prediction model was conducted using statistical indicators of Root Mean Square Error (RMSE), determination coefficient (R2), and Mean Absolute Error (MAE) for both Random Forest and Bagging models. The salinity intrusion map for the lower Mekong Delta was provided using the Random Forest model.