24AI Advanced Precipitation Estimation Using Weather Radar and Machine Learning: Applications and Results in Southern Brazil.

In addressing the challenges of quantitative precipitation estimation (QPE) using weather radar, it is paramount to enhance rainfall estimates for applications such as flash flood forecasting and hydropower generation management. This study employs dual-polarization weather radar data to refine the traditional Z-R relationship, often inaccurate in areas exhibiting complex meteorological phenomena. By integrating advanced machine learning (ML) techniques, specifically, tree-based algorithms such as random forest (RF) and gradient boosting (GB), the research aims to capture intricate patterns within the Z-R relationship using polarimetric variables. The innovative aspect of this study lies in the hybrid approach, which combines classification and regression methodologies to enhance precipitation estimation. The Z-R relationship, fundamental to radar meteorology, has various limitations. These include variability in raindrop size distribution, mixed-phase precipitation, radar signal attenuation, and physical obstructions. These challenges are particularly pronounced in Western Paraná, Brazil, a region of significant hydrological and meteorological complexity. The study addresses these issues by incorporating polarimetric variables, including differential reflectivity (ZDR), specific differential phase (KDP), and co-polar correlation coefficient (RhoHV) machine learning (ML) models, to enhance the accuracy of rainfall estimation. The data were collected from a dual-polarization weather radar system installed in Cascavel, Paraná, Brazil, and a network of 36 rain gauges. The radar system, which has been operational since 2014, provides high-granularity atmospheric data, which, along with the rain gauge data, formed the basis of the dataset used in this study. Data preprocessing involved distance filtering, handling missing data, selecting crucial polarimetric variables, and ensuring consistency between radar reflectivity and rain gauge data. Feature engineering was a critical step, where variables such as horizontal reflectivity (DBZH), ZDR, KDP, and RhoHV were selected for their relevance and transformed to improve the precision of predictions. The models were trained and evaluated using data from 2018 to 2022, focusing on precipitation events exceeding 0.2 mm per 15 minutes, aligning with the calibration settings of the rain gauges. The model development involved two primary stages: classification and regression. The classification stage used RF and GB techniques to distinguish between "rain" and "no rain" events, while the regression stage estimated the precipitation intensity. The hybrid approach effectively combined the strengths of both techniques, ensuring comprehensive and accurate QPE. The model evaluation employed statistical metrics such as root mean square error (RMSE), mean absolute error (MAE), and Kling-Gupta efficiency (KGE). The results demonstrated the superiority of the ML models over traditional Z-R relations, with RF and GB models achieving high accuracy, recall, and precision. The RMSE and MAE values were significantly lower for the ML models than traditional methods, indicating better performance in quantitative estimation. A comparative analysis of different ML models and Z-R relationships highlighted the importance of algorithm selection in accurately modeling intense rain events. The GB model showed robustness under extreme precipitation conditions, making it a valuable tool for practical applications where accurate identification of heavy rainfall is crucial. This study's innovative hybrid approach, integrating advanced ML techniques with polarimetric radar data, significantly improves the accuracy of QPE. The findings have tangible implications for academic research and practical applications, such as weather forecasting, climate analysis, water resource management, and urban planning. Additionally, this methodology will be employed in more dual-polarization radars across Brazil due to the positive outcomes.