BP-LSTM-ATTENTION: A Regional Precipitation Multisource Data Processing Fusion Model

Qinghai Province shoulders the significant mission of maintaining ecological security, protecting the three-river source region, and safeguarding the ‘Chinese Water Tower’. To improve the accuracy of precipitation data is a great importance for studying regional ecological changes. The paper takes Qinghai Province as the research area and develops a multi-source precipitation data fusion model: BP-LSTM-ATTENTION, which is based on BP network, LSTM network and attention mechanism. The model uses satellite radar precipitation data and auxiliary data such as vegetation normalized difference index as feature data, and ground observation station data as label values, and through training and evaluation, improving the accuracy of data fusion. Then, a regional long-term and high spatial resolution precipitation fusion dataset with a spatial resolution of 1Km, a temporal resolution of 3 hours, and a time series from 2000 year to 2020 year, was generated by the data fusion model. Taking the fusion processing of precipitation data from 2016 to 2020 as an example, the fusion model successfully reduced the average RB of precipitation fusion data in the study area to 14.6%, the average RMSE to 0.1566, the average MAE to 43.706, and the average CC to 0.906. The experiment shows that the precipitation data fusion model studied in this paper has high data fusion processing accuracy and good precipitation data prediction performance. In terms of the model's ability to capture precipitation events, the precipitation data fused by the model has an average POD of 0.818, an average FAR of 0.318, and an average MISS of 0.182 from 2016 to 2020, demonstrating its strong capability in monitoring precipitation events and good performance in reducing false alarms and missed reports. The BP-LSTM-ATTENTION model proposed in this study solves the problems of spatiotemporal correlation and time series in precipitation data of Qinghai Province, and has produced high-resolution, high spatiotemporal precision precipitation datasets, which is a of great significance for the digital ecological protection research in Qinghai Province.