Comparing the machine learning and empirical algorithms on remote sensing estimates water quality parameters of constructed wetland

Abstract

The water quality parameters (WQPs) at the constructed wetland of the lakeside zone are fluctuating. In order to accurately estimate the water purification efficiency of the lakeside belt, the remote sensing technology by satellite has a great advantage to complete such a task. In this study, back-propagation neural network and polynomial regression are compared in remote sensing estimate Total Nitrogen (TN), Total Phosphorus (TP), Nitrate Nitrogen (NN), and Ammonia Nitrogen (AN) concentrations in constructed wetland water quality. The result shows that the BP neural network algorithms outperformed the polynomial regression algorithms in the estimate AN and TN. However, the polynomial regression algorithms have achieved better performance in the estimate NN and TP. Moreover, the best algorithms produce about 60% of rRMSE in all WQPs in this study. As to mean normalized bias (MNB) result, the overall estimate by the BP neural network algorithms is lower than the measured data. In addition to TP, the empirical model is the opposite. This study could provide some reference for the remote sensing estimate of the water purification efficiency in constructed wetlands. Furthermore, BP neural network performance is more stable than the polynomial regression.