Daily prediction of Urmia Lake water level using remote sensing data and honey badger optimization-based data-driven models

Artificial neural networks (ANNs), support vector regression (SVR) and CatBoost regression (CBR) machine learning methods have been combined with the honey badger optimization algorithm (HBA) and metaheuristic optimization algorithm to accurately and reliably predict lake water level (LWL), which is of great importance for the management and planning of water resources. In this study, meteorological and hydrological parameters, including temperature (T), precipitation (P), date (D), surface soil moisture (SSW), root zone moisture (RZW) and water level (WL), were employed as input data for predicting the LWL of Urmia Lake. The input data were employed to develop six different prediction scenarios. This study not only examined the impact of meteorological and hydrological parameters on LWL prediction but also compared the performance of individual models and hybrid models. The Akaike information criterion (AIC) index was used to ascertain the optimal machine learning model and to evaluate the six prediction scenarios. The results of the study indicate that, according to the AIC index, the data regarding the water level (WL) were significant in the prediction models. However, it should be noted that satisfactory results could also be obtained without using the WL data in certain scenarios. In scenario 4 (input data: D, T, P, SSW, RZW), where the WL variable was not included, the HBA-CBR hybrid model was the best model with the lowest AIC value (Train: -63,735, Test:-4693). In prediction scenario 6 (input data: D, T, P, SSW, RZW, WL), which included the WL data, the HBA-SVR hybrid model demonstrated high performance with the lowest AIC value (Train: -102,358, Test:-27,233). Accordingly, it was recommended to use lagged WL values as input in WL prediction because the prediction accuracy of the models significantly improved. Furthermore, hybrid models were found to perform better than individual models due to their more consistent results.