Feature Selection in Machine Learning for Wetland Water Level Prediction

Water levels in wetlands are crucial for preserving the environment, maintaining biodiversity, controlling floods, and regulating water quality. Effective water level prediction in wetlands leads to proper and efficient management of wetlands, thus supporting the factors mentioned above. Although previous work can be found in the context of water level prediction in lakes and wetlands, a proper feature selection of the input features for predicting wetland water levels can be identified as a research gap. This study addresses this gap using machine learning techniques and verifies the results through a cost analysis of the input features. Four wetlands in the Colombo Flood Detention Center: Kimbulawala Bridge, Kotte Canal, Heen Canal, and Parliamentary Canal were chosen as the locations to conduct the study. Feature selection under ensemble techniques, namely Random Forest and Gradient Boosting, were used in the study to reduce the number of features used to predict water levels. Fivefold cross-validation was used for model selection and to obtain evaluation measures: root mean squared error (RMSE) and coefficient of determination (R² value), for the models trained for each combination of selected features. Six machine learning (ML) algorithms: Linear Regression, Random Forest, Gradient Boosting, XGBoost, LightGBM (Light Gradient Boosting Machine), and Support Vector Regression were used to train models incorporating the groups of selected features. The integration of the results of feature selection and cost analysis showed that a cost-effective solution can be derived using three features instead of seven. This reduction emphasizes that feature selection using machine learning for wetland water level prediction would lead to reducing costs associated with the collection of input parameters and this method can be adopted in similar fields of study.