Quantitative Precipitation Estimation Using Machine Learning Techniques for Weather Radar and Rain Gauge Data Fusion

Accurately estimating precipitation from weather radar data is crucial for predicting and managing various environmental and societal applications, such as agriculture and energy production. Despite technological advances, this task remains challenging. In Brazil, where agriculture and energy production are significant economic activities, improving precipitation estimation is a necessary goal. We have developed a quantitative precipitation estimation algorithm called SIPREC (System for Integrated PRECipitation) to address this issue, which has been used operationally for over 15 years. This algorithm combines data from different sources (Multi-Sourced data input), such as weather radar, rain gauges, and satellites. SIPREC utilizes an automated precipitation classification scheme based on reflectivity structures to obtain precipitation estimates. This method aggregates data from rain gauges by interpolation while maintaining the spatial distribution of the radar or satellite measurement, improving the accuracy of hydrological models compared to a model using only rain gauge data, which is capable of collecting measures only at a specific point in time. This method is an essential advantage in an operational environment since it does not require frequent processing to update the weights as in other known schemes. However, we have observed that when improving QPE (Quantitative Precipitation Estimation), problems arise from exponential approximations related to the uncertainties of the Z-R estimate, the spatial variability, and the one-hour temporal resolution. Therefore, there is a need to enhance the precision of radar-based precipitation estimates at the bin level (i.e., the volume observed by the radar, defined by the range and azimuth). Machine learning techniques have shown promise in improving the accuracy of precipitation estimates by learning from the data and capturing complex patterns and previously unknown relationships between the variables. For this Investigation, decision tree-based models within machine learning framework are used to improve radar-based precipitation estimates. Tree-based models, in particular, have gained attention in recent years due to their ability to handle large datasets and non-linear relationships. Our study evaluated the performance of various tree-based models, including Random Forest and Gradient Boosting, in predicting radar-based precipitation estimates. These models were trained and tested using polarimetric variables (Z, ZDR, KDP, RHOHV) collected from the weather radar in Parana, southern Brazil. We compared the accuracy of machine learning models with models based on predefined Z-R relations (e.g., Marshall-Palmer) by analyzing the Root Mean Square Error (RMSE) and Median Absolute Error (MAE) metrics. Our findings reveal that the machine learning models performed significantly better than those based on predefined Z-R relations, indicating that incorporating the specifics of tree-based model algorithms can improve the correlation and accuracy of precipitation estimation from radar. A performance evaluation study shows improvements in precipitation estimation, primarily when used in real-time in an operational environment. This paper presents the results of this evaluation, with applications in severe weather events with high precipitation in the area.