Monitoring Total Phosphorus and Total Nitrogen Using Hybrid Machine Learning

The nutrients in surface waters have poor optical properties, and it is challenging to characterize the relationship between concentration and their reflectance. Accuracy of non-optical water quality prediction by machine learning models were affected by the feature used to train the models. However, studies on the prediction of non-optically surface water parameters using hybrid machine learning (ML) were limited. Therefore, this study was aimed to predict the concentration of Total Phosphorus (TP) and Total Nitrogen (TN) in Koka Reservoir and Ziway Lake using a hybrid ML model. Random Forest Regression (RFR), Back Propagation Neural Network (BPNN), Support Vector Regression (SVR), and Enhanced Whale Optimization Algorithm (LSEWOA) features optimization models are analyzed by ascertaining the correlation between on-site data and features reflectance. At equal overpass period, Sentinel-2 (S2) images and on-site data were collected from July 2015 to February 2024. Demonstrates that the LSEWOA-RFR model exhibits better TN and TP prediction performance than the other models in both study area. This study also revealed that, the TN and TP concentrations were estimated high in end of summer and a low in the middle of the dry seasons. The spatial variations of nutrient concentrations indicated high in areas near river inflow and agricultural activities. The finding suggests that LSEWOA-RFR model is a new approach to estimate nutrient in the lakes and reservoirs that achieves more accuracy, efficient and operationally feasible. Furthermore, this study recommend to use this hybrid ML model for monitoring the non-optically active surface water parameters in both study area.