Rainfall downscaling and adjustment methods

A crucial component to hydrological forecasting is the presence of high-quality, high-resolution quantitative precipitation estimates (QPE) and quantitative precipitation forecasts (QPF), especially for flooding situations. Current gridded precipitation products from reanalyses, forecasting products and remotely sensed products are a powerful source of information as they provide precipitation quantities in both space and time. Yet, these products often suffer from systematic biases and spatial inaccuracies due to their sometimes large grid cell sizes, which makes it challenging to capture local extremes. While conventional downscaling and adjustment methods, ranging from simple to advanced gauge-based merging in hindcasting mode and statistical methods in forecasting mode, do exist, there is no tool yet that combines all these methods for usage in operational applications. In this report, we present the open-source pyRainAdjustment Python tool, which integrates various methods for adjustment and downscaling of gridded precipitation products into a single tool compatible with Delft-FEWS. pyRainAdjustment splits the approach in three parts: (1) downscaling, for which it uses a climatology-based downscaling approach that uses a high-resolution base dataset (e.g. WorldClim2) on a monthly basis; (2) adjustment in a hindcasting mode, for which five approaches were implemented (uniform mean field bias adjustment, multiplicative error adjustment, additive error adjustment, mixed error model adjustment and kriging with external drift); and (3) adjustment in a forecasting mode, which consists of the statistical quantile mapping method. The introduced approaches in pyRainAdjustment were tested for the upper Murray River for MDBA. In hindcasting mode, the results indicate that downscaling and spatial adjustment methods in pyRainAdjustment improve the QPE quality. Although uniform adjustments provide only minimal improvement (8% for ERA5 and slightly more for radar data), spatial techniques such as mixed error modelling and kriging with external drift show better results (up to a maximum of 27% improvement), with mixed error adjustment slightly leading in performance. In forecasting mode, quantile mapping reduces the mean RMSE of ECMWF’s IFS forecasts by approximately 10%. The method effectively reduces the frequency of large forecast errors, particularly for moderate to high precipitation events (8–20 mm), while maintaining similar performance in low-error scenarios. Hence, this indicates that the methods in pyRainAdjustment can aid in reducing the errors in both QPE and QPF products. In addition to the (existing) methods that were brought together in pyRainAdjustment, we introduce the first results of an exploration of machine learning (ML) approaches for gridded precipitation downscaling and adjustment in this report. An advantage of ML methods for this purpose is that they can simultaneously downscale and adjust precipitation data without relying on real-time (gauge) observations. This report describes the current testing setup and considered model architectures and presents the first results using the U-Net architecture on ERA5 reanalysis data with MSWEP QPE as target and reference. The U-Nets show potential in hindcasting mode, with improvements over the original ERA5 QPE in several metrics (with improvement around 10%, but depending on the consulted metric). The model still tends to overpredict moderate precipitation and underpredict extremes, but largely reduces the overall bias of the QPE. These initial results suggest that while U-Net shows potential, further refinement is needed to improve its performance across all precipitation ranges. The research will continue past this project with improving the U-Nets, trying out the other architectures and then moving to the forecasting domain.