Uncertainty analysis of radar rainfall estimates induced by atmospheric conditions using long short-term memory networks

The unstable performance of radar rainfall measurement hinders its application in hydrology. Radar rainfall uncertainties induced by atmospheric conditions as raindrops fall from the radar sampling altitude to the ground are critical to radar-based quantitative precipitation estimation. However, these have rarely been considered in previous radar rainfall correction procedures. This study first demonstrates that the correlations between the radar–gauge rainfall discrepancy (RGD) and atmospheric fields are strong. A systematic radar-rainfall adjustment method is then proposed to decrease the discrepancies originating from changes in atmospheric conditions using a long short-term memory (LSTM) network. Three RGD adjustment models were established using a mass-variation scheme, a wind-drift scheme, and a combined mass-variation and wind-drift scheme, based on long-term (2013–2017) data covering most of the United Kingdom. The evaluation results demonstrate that all of the designed models performed well from overall, single-site, and event perspectives. Overall, the combined model, which exhibited the best performance, decreased the root-mean-square error between the rainfall levels measured by radar and gauges by 23.84%, increased Pearson's correlation coefficient from 0.23 to 0.53, and improved the critical success index of the radar rainfall estimation from 0.56 to 0.92. The results indicate that the performances of the proposed models improved with an increase in average relative humidity or wind speed, which demonstrates that they can correct rainfall under high levels of relative humidity and wind speed. This study establishes a highly capable and comprehensive adjustment framework for radar rainfall estimation uncertainty induced by atmospheric conditions, which is essential for achieving high-quality radar products.