A Comparative Analysis of Machine Learning Models for Simulating, Classifying, and Assessment River Inflow

Accurately classifying river inflow is crucial for understanding river dynamics and ecosystem health. This study evaluates the performance of seven machine learning models, including Support Vector Machines (SVM), Logistic Regression (LR), K-Nearest Neighbors (KNN), Decision Trees (DT), Random Forests (RF), Adaptive Boosting (AdaBoost), and Multi-Layer Perceptron (MLP), for streamflow classification. One of the key challenges in this task is the imbalance in class distributions, which can negatively impact model performance. To address this, we apply the Synthetic Minority Over-sampling Technique (SMOTE) to improve classification outcomes for minority classes. Furthermore, we investigate the impact of four proposed feature selection methods, including mutual information (MI-FS), linear kernel SVM (SVM-FS), random forest (RF-FS), and multi-criteria selection (MC-FS) on model performance by identifying optimal lag values. Model hyperparameters are fine-tuned using GridSearchCV technique, and evaluation step is assessed across seven performance metrics. Experimental results show that MLP and SVM consistently outperform other models, making them the most suitable choices for streamflow classification. Among the FS techniques, MC-FS demonstrates superior performance by effectively reducing dimensionality while preserving predictive power. However, our findings indicate that SMOTE enhances classification for minority classes but reduces accuracy for majority classes, highlighting a trade-off in handling imbalanced data. Additionally, we observe that the linear assumption in SVM-FS can negatively impact model performance when it fails to detect all relevant input features. These insights provide valuable guidance for future streamflow classification tasks.