Development of Multi-Source Weighted-Ensemble Precipitation: Influence of bias correction based on recurrent convolutional neural networks

•Best available precipitation products from multiple sources for Taiwan were merged.•Influence of deep learning-based bias correction on the merged products was examined.•Our tailor-made merged product outperformed the original product.•The merged product can be more accurate than gauge-based data over ungauged areas. Accurate precipitation information is the cornerstone of regional hydroclimatic studies, and merging precipitation data from various sources provides a means to enhance data accuracy. This study aims to apply the technique referred to as Multi-Source Weighted-Ensemble Precipitation (MSWEP) to merge the gauge-, satellite-, and model-based precipitation products for Taiwan (MSWEP_TW). To correct the known biases in satellite precipitation, the long short-term memory (LSTM) and the emerging convolutional long short-term memory (ConvLSTM) networks are employed. Afterward, how the correction of satellite data influences the performance of merged precipitation is assessed. The correction of satellite data reveals that LSTM with the spatial coherence scheme can show similar effectiveness as ConvLSTM in increasing the correlations of satellite data with gauge-based data by ∼10 %. MSWEP_TW is proven to outperform the original product (i.e., MSWEP version 2.8), and higher weights for satellite- and model-based products over gauge-scarce regions are verified. Further, this study confirms that the merged product can provide more accurate precipitation information than gauge-only interpolation, promoting the advantage of merged precipitation for ungauged areas. Lastly, the enhancement made in the correction of satellite data directly contributes to the development of satellite-only merged products for Taiwan with low latency, suggesting their usefulness in near real-time applications.