Enhancing the applicability of radar-based precipitation nowcasting to hydrological applications with a machine-learning based error modelling method

Probabilistic radar-based precipitation nowcasting has become increasingly crucial for real-time hydrological applications due to its high accuracy at short lead time. However, its reliability for hydrological usage is limited by two major sources of error and uncertainty, both of which tend to exacerbate quickly with lead time. The first source lies in the limitations of nowcasting algorithms, including inaccuracies in rainfield advection and inadequate modeling of rain cell evolution. The second arises from discrepancies in precipitation measurements, referring to the differences between radar-derived estimates and rain gauge observations. Aligning these estimates with actual ground-level precipitation is vital for practical hydrological applications.

This study focuses on addressing the errors and uncertainties inherent in precipitation 'measurements', aiming to enhance the reliability of original nowcasts. Here, uncertainty refers to the range within which the true value is expected to fall at a given confidence level, while error denotes to the systematic bias between estimated and true values. The proposed methodologies utilise rain gauge data as the ground truth and employs the Short-Term Ensemble Prediction System (STEPS) to generate radar-based ensemble nowcasts. To deal with these issues, an initial attempt was conducted with the Censored Shifted Gamma Distribution (CSGD) model. However, the model faces challenges in selecting an appropriate metric as the adjusted value, limiting the potential reduction in RMSE to approximately 5–10%. To overcome this limitation, a random forest (RF) regression model is proposed, as it can avoid predefined assumptions about rainfall intensity distribution. This model incorporates variables such as nowcasted rainfall intensity, orographic features, and meteorological parameters such as wind speed, wind direction, humidity, cloud type, and cloud base height (CBH), to estimate corresponding rain gauge measurements. At each rain gauge location, the error distribution is parametrised by comparing the original and adjusted rainfall intensities and fitting them to various probability functions. These parameters are then spatially interpolated using geostatistical techniques to generate an error map. The resulting error map is applied to correct the original nowcasts across the study area, enhancing their overall accuracy and reliability.

The United Kingdom, benefiting from its comprehensive and high-quality meteorological data, was selected as the study area. The 1-km UK C-band radar composite, derived from the Met Office Nimrod System, serve as the radar rainfall dataset for generating ensemble nowcasts. Rain gauge data and additional meteorological variables are sourced from the Met Office Integrated Data Archive System (MIDAS) and the Met Office LIDARNET ceilometer network. Rainfall events from 2016 to 2022 are analysed, with events from 2016 to 2020 designated as the training period for developing random forest models and error maps. For validation, 20 events from 2021 to 2022 are selected to assess the performance of both the original and adjusted nowcasts. Preliminary results indicate that the adjusted ensemble nowcasts exhibit significantly improved alignment with rain gauge measurements compared to the original nowcasts. These findings highlight the potential of the proposed methodology to reduce both error and uncertainty in radar-based precipitation nowcasting, particularly for hydrological applications such as flood and landslide forecasting.