Evaluating the performance of CHIRPS and CPC precipitation data for streamflow forecasting using multiple linear...

Area of Science:

• Hydrology
• Environmental Science
• Data Science

Background:

• Accurate streamflow forecasting is crucial for water resource management but traditionally relies on costly, coarse-resolution ground-based weather stations.
• Machine learning offers cost-effective streamflow prediction with minimal input data, addressing limitations of traditional methods.

Purpose of the Study:

• To evaluate the effectiveness of gauge-based (CPC) and satellite-based (CHIRPS) rainfall data for streamflow forecasting in the Wolf River watershed.
• To compare the performance of Multiple Linear Regression (MLR) and Long Short-Term Memory Neural Network (LSTM-NN) models for streamflow prediction.

Main Methods:

• Utilized daily precipitation data from CHIRPS and CPC (1991-2021) for the Wolf River watershed.
• Developed and compared MLR and LSTM-NN models using daily streamflow data from USGS gauge 07031650.
• Assessed model performance using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE).

Main Results:

• The LSTM-NN model using CHIRPS data achieved lower RMSE (15.02) and MAE (21.53) compared to CPC data.
• LSTM-NN models outperformed MLR models in streamflow prediction accuracy.
• CHIRPS data demonstrated superior performance over CPC data when paired with the LSTM-NN model.

Conclusions:

• CHIRPS satellite-based rainfall data is a viable and effective alternative to gauge-based CPC data for streamflow forecasting in the study area.
• The LSTM-NN model is a more effective tool for streamflow prediction than MLR, offering improved accuracy.
• These findings support the use of satellite data and machine learning for cost-effective and reliable water resource management.