Kaleybar et al. (2025) CNN-LSTM-RF integration for predicting Mississippi River discharge dynamics

Identification

• Journal: Acta Geophysica
• Year: 2025
• Date: 2025-10-11
• Authors: Fariborz Ahmadzadeh Kaleybar, Ahad Molavi
• DOI: 10.1007/s11600-025-01719-x

Research Groups

• Department of Water Sciences and Engineering, Ta.C., Islamic Azad University, Tabriz, Iran

Short Summary

This study developed and evaluated hybrid machine learning models (CNN-LSTM, RF-LSTM, CNN-RF-LSTM) to predict Mississippi River discharge at the Memphis station. The CNN-LSTM model with a 3-day lag interval demonstrated the highest accuracy (NRMSE = 0.0165), providing an efficient tool for water resource management.

Objective

• To predict the daily discharge of the Mississippi River at the Memphis station using hybrid machine learning models (CNN-LSTM, RF-LSTM, CNN-RF-LSTM) with varying lag intervals (3–18 days).

Study Configuration

• Spatial Scale: Mississippi River at the Memphis station, Memphis, Tennessee, USA (Latitude: 35° 8′ 58.2″ N, Longitude: 90° 2′ 56.4″ W).
• Temporal Scale: Daily discharge data from 1990 to 2024, with prediction lag intervals ranging from 3 to 18 days.

Methodology and Data

• Models used: Long Short-Term Memory (LSTM), Random Forest (RF), Convolutional Neural Network (CNN), and hybrid models: CNN-LSTM, RF-LSTM, CNN-RF-LSTM.
• Data sources: Long-term, high-quality historical daily discharge records from the Mississippi River at the Memphis station.

Main Results

• All models exhibited outstanding predictive capabilities, with R² values consistently exceeding 0.9982 and Nash–Sutcliffe efficiency (NSE) values above 0.998 across all time-lag intervals.
• The CNN-LSTM model with a 3-day lag interval (CNN-LSTM-3) achieved the best overall performance, with an R² of 0.9991, Root Mean Square Error (RMSE) of 284.10 m³/s, Normalized Root Mean Square Error (NRMSE) of 0.017, Mean Absolute Percentage Error (MAPE) of 1.422%, NSE of 0.9991, and Willmott's Index of Agreement (WI) of 0.9998.
• The standalone LSTM-3 model also performed strongly (R² = 0.9990, RMSE = 296.41 m³/s, NRMSE = 0.0172, MAPE = 1.4341%).
• Performance generally degraded with increasing lag intervals, but hybrid models maintained stronger stability.
• Under extreme flow conditions, the CNN-RF-LSTM system performed optimally with an NSE value of 0.9512, demonstrating enhanced generalizability compared to standalone LSTM (NSE = 0.8915) and other hybrids.

Contributions

• Pioneering application of a CNN–RF–LSTM hybrid framework for Mississippi River discharge forecasting, integrating spatial feature extraction, temporal sequence learning, and ensemble-based regularization.
• Development of a resilient and exceptionally precise predictive model for a large-scale river system.
• Advances methodological development in hydrological forecasting and provides practical insights for sustainable river basin management, enhancing efficiency in forecasting and reducing labor costs.

Funding

• The authors did not receive support from any organization for the submitted work.

Citation

@article{Kaleybar2025CNNLSTMRF,
  author = {Kaleybar, Fariborz Ahmadzadeh and Molavi, Ahad},
  title = {CNN-LSTM-RF integration for predicting Mississippi River discharge dynamics},
  journal = {Acta Geophysica},
  year = {2025},
  doi = {10.1007/s11600-025-01719-x},
  url = {https://doi.org/10.1007/s11600-025-01719-x}
}