A Multi-source Precipitation Data Fusion Model for Qinghai Province Based on 3D CNN and Bidirectional ConvLSTM

Abstract Accurate precipitation estimation is crucial for water resource management, ecological modeling, and natural disaster prevention. Although there are various methods for integrating precipitation data from different sources with observed data, many of them only focus on spatial or temporal correlations, neglecting the potential improvement in precipitation estimation brought about by considering spatio-temporal correlations. In this study, we propose a deep learning-based multi-source precipitation fusion model (MPFM), which simultaneously utilizes the spatial-temporal correlations of ERA5, IMERG, and GLDAS for fusion. Unlike existing methods, our model integrates both spatial and temporal correlations, enhancing fusion by utilizing spatio-temporal information more fully. The MPFM comprises a 3D Convolutional Neural Network (3D CNN) to extract the spatio-temporal features of precipitation data and a bidirectional Convolutional Long Short-Term Memory Neural Network (ConvLSTM) to capture the temporal and spatial dependencies. Compared to other models, including deep learning methods (3D CNN and bidirectional ConvLSTM), traditional machine learning methods (Multi-Layer Perceptron, eXtreme Gradient Boost, and Random Forest), and linear fusion methods (one outlier-removed-average, optimized weighted average, and inverse error variance weighting), the MPFM significantly outperforms in independent validation tests. The generated high-precision daily precipitation dataset for 2013–2017 at a 0.01°resolution exhibits substantially improved accuracy over raw data, with a 17.16% and 25.66% reduction in mean absolute error (MAE) and root mean square error (RMSE), respectively, and a 20.35% increase in correlation coefficient (CC) compared with the best-performing raw data (IMERG). This dataset can offer dependable data support for hydrological or climate ecology-related studies in Qinghai Province.