Abstract

Predicting streamflow in ungauged basins is a challenging hydrological issue that requires accurate estimation for effective water resource management. This article aims to evaluate the effectiveness of five different Machine Learning (ML) models (i.e., M5 model tree (M5), Random Forest (RF), Support Vector Regression-polynomial kernel (SVR-poly), Support Vector Regression-radial basis function kernel (SVR-rbf), and Multilayer Perceptron (MLP)) for predicting monthly streamflow in ungauged basins. The proposed models were compared with the method of GR2M's regionalized model parameters. Data was collected from 37 streamflow stations in the southern basin of Thailand. The data utilized included hydrological information like monthly rainfall, potential evapotranspiration, and streamflow, as well as physical watershed characteristics such as basin size, river length, distance from the hydrometric station to the area's centroid, and slope. The study evaluated these methods for two distinct scenarios, namely (a) estimating average monthly streamflow and (b) estimating monthly streamflow. The study was conducted in four phases: selection of input data, hyperparameter tuning, performance comparison of different models, and assessment of the chosen model's suitability for predicting monthly streamflow in ungauged basins. Five-fold cross-validation with four statistical indicators, namely, the Nash-Sutcliffe Efficiency (NSE), Overall Index (OI), Coefficient of Determination (r²), and Combined Index (CI), were utilized for the model's performance comparison. The results showed that the RF model produced the best performance compared to other ML models and outperformed the GR2M's regionalized model parameters in both scenarios, achieving performance indicators with NSE >0.6, OI > 0.6, r² > 0.6, and CI > 2.0.