An evaluation of random forest based input variable selection methods for one month ahead streamflow forecasting.

Area of Science:

• Hydrology and Water Resources
• Machine Learning in Environmental Science
• Data-driven Environmental Modeling

Background:

• Accurate streamflow forecasting relies heavily on selecting optimal input variables for data-driven models.
• While Random Forest (RF) is used for input variable selection (IVS), a comprehensive comparison of different RF-IVS methods is lacking.
• Existing methods like partial Pearson correlation and conditional mutual information have limitations in complex hydrological systems.

Purpose of the Study:

• To comparatively analyze the performance of five RF-IVS methods across four distinct data-driven models.
• To evaluate the effectiveness of RF-IVS in enhancing streamflow forecasting accuracy.
• To identify the most effective RF-IVS strategy and model combination for hydrological applications.

Main Methods:

• Investigated five RF-IVS techniques applied to four data-driven models: RF, Support Vector Regression (SVR), Gaussian Process Regression (GP), and Long Short-Term Memory (LSTM).
• Conducted a case study for one-month-ahead streamflow forecasting in the contiguous United States.
• Compared RF-IVS methods against traditional approaches like partial Pearson correlation and conditional mutual information.

Main Results:

• RF-IVS methods significantly improved streamflow forecasting performance compared to conventional techniques.
• Performance-based RF-IVS methods outperformed test-based methods, which tended to select redundant variables.
• The combination of RF with a forward selection strategy demonstrated superior performance when integrated with the GP model.

Conclusions:

• RF-IVS is a valuable approach for optimizing input variable selection in streamflow forecasting.
• Performance-based RF-IVS strategies are more effective than test-based ones, avoiding variable redundancy.
• The RF forward selection strategy coupled with the GP model presents a promising combination for accurate streamflow prediction.