A Multi-Layer Perceptron Regression and Variant Windowing for Estimating Rainfall Based on Weather Radar Data

Accurate rainfall information is crucial for various applications, including river flow estimation, water resource management, and flood warning system development. Traditional rain gauge networks, however, suffer from limited spatial coverage, leading to incomplete and biased data for large areas. This study proposes a novel approach for surface rainfall estimation using weather radar data and a MultiLayer Perceptron (MLP) Regressor machine learning model. Grid search was employed to explore model performance across different windowing configurations: no windowing, n-1 windowing, and n-2 windowing. The results demonstrate that n-1 windowing outperforms other configurations, achieving an average RMSE of 0.987, MAE of 0.263, and R-squared of 0.242 across five locations. This suggests that n-1 windowing effectively captures the temporal dynamics of rainfall patterns while improving the model's sensitivity to regularization. However, a tendency for underestimating high-intensity rainfall events remains. This research highlights the effectiveness of n-1 windowing with MLP Regressors for enhanced surface rainfall estimation using weather radar data. Further investigation is needed to address the underestimation bias, particularly for high rainfall events.