Calibration of X-Band Radar for Extreme Events in a Spatially Complex Precipitation Region in North Peru: Machine Learning vs. Empirical Approach

• Published in: Atmosphere 2021-12, Vol.12 (12), p.1561
• Main Authors: Rollenbeck, Rütger, Orellana-Alvear, Johanna, Rodriguez, Rodolfo, Macalupu, Simon, Nolasco, Pool
• Format: Article
• Language: English
• Subjects: Algorithms; Atmospheric attenuation; Atmospheric precipitations; Calibration; Clutter; Decision trees; El Nino; El Nino phenomena; extreme events; Gauges; Learning algorithms; Machine learning; Meteorological radar; Methods; Noise; Precipitation; Precipitation estimation; Precipitation patterns; quantitative precipitation estimate; Radar; Radar data; Radar reflectivity; Radar systems; Rain; Rain gauges; Rainfall; random forest; Reflectance; Slopes; Software; Spatial variability; Spatial variations; Superhigh frequencies; Time series; tropical desert; Weather radar
• ISSN: 2073-4433; 2073-4433
• DOI: 10.3390/atmos12121561

Cost-efficient single-polarized X-band radars are a feasible alternative due to their high sensitivity and resolution, which makes them well suited for complex precipitation patterns. The first horizontal scanning weather radar in Peru was installed in Piura in 2019, after the devastating impact of the 2017 coastal El Niño. To obtain a calibrated rain rate from radar reflectivity, we employ a modified empirical approach and draw a direct comparison to a well-established machine learning technique used for radar QPE. For both methods, preprocessing steps are required, such as clutter and noise elimination, atmospheric, geometric, and precipitation-induced attenuation correction, and hardware variations. For the new empirical approach, the corrected reflectivity is related to rain gauge observations, and a spatially and temporally variable parameter set is iteratively determined. The machine learning approach uses a set of features mainly derived from the radar data. The random forest (RF) algorithm employed here learns from the features and builds decision trees to obtain quantitative precipitation estimates for each bin of detected reflectivity. Both methods capture the spatial variability of rainfall quite well. Validating the empirical approach, it performed better with an overall linear regression slope of 0.65 and r of 0.82. The RF approach had limitations with the quantitative representation (slope = 0.44 and r = 0.65), but it more closely matches the reflectivity distribution, and it is independent of real-time rain-gauge data. Possibly, a weighted mean of both approaches can be used operationally on a daily basis.