Deep Learning Approaches for Regional Rainfall Time Series Prediction Using ERA5 Dataset

Abstract

Precise monthly rainfall forecasting is crucial for Gujarat, India, where agriculture depends heavily on monsoon patterns. This study compares three deep learning architectures—Convolutional Neural Network (CNN), Convolutional Long Short-Term Memory (Conv-LSTM), and hybrid Discrete Wavelet Transform-ConvLSTM (DWT-ConvLSTM)—using 10 years (2010-2020) of high-resolution ERA5 meteorological data processed via Google Earth Engine across approximately 14,600 grid points. The CNN baseline achieved a Mean Absolute Error (MAE) of 0.0471 and Root Mean Square Error (RMSE) of 0.0747. Conv-LSTM improved performance with MAE 0.0421 and RMSE 0.0723. The proposed DWT-ConvLSTM hybrid model excelled, attaining the lowest errors of MAE 0.0370 and RMSE 0.0570—a 12% improvement over Conv-LSTM. This superior performance stems from wavelet decomposition, which isolates complex climate signals into frequency components for enhanced pattern recognition. These findings demonstrate that integrating frequency-domain analysis with deep learning effectively uncovers hidden spatiotemporal climate patterns. Though computationally intensive, the hybrid model offers significant potential for agricultural planning, water resource management, and climate adaptation in monsoon-dependent regions.