Customized deep learning for precipitation bias correction and downscaling

Abstract

Abstract. Systematic biases and coarse resolutions are major limitations of current precipitation datasets. Many deep learning (DL)-based studies have been conducted for precipitation bias correction and downscaling. However, it is still challenging for the current approaches to handle complex features of hourly precipitation, resulting in the incapability of reproducing small-scale features, such as extreme events. This study developed a customized DL model by incorporating customized loss functions, multitask learning and physically relevant covariates to bias correct and downscale hourly precipitation data. We designed six scenarios to systematically evaluate the added values of weighted loss functions, multitask learning, and atmospheric covariates compared to the regular DL and statistical approaches. The models were trained and tested using the Modern-era Retrospective Analysis for Research and Applications version 2 (MERRA2) reanalysis and the Stage IV radar observations over the northern coastal region of the Gulf of Mexico on an hourly time scale. We found that all the scenarios with weighted loss functions performed notably better than the other scenarios with conventional loss functions and a quantile mapping-based approach at hourly, daily, and monthly time scales as well as extremes. Multitask learning showed improved performance on capturing fine features of extreme events and accounting for atmospheric covariates highly improved model performance at hourly and aggregated time scales, while the improvement is not as large as from weighted loss functions. We show that the customized DL model can better downscale and bias correct hourly precipitation datasets and provide improved precipitation estimates at fine spatial and temporal resolutions where regular DL and statistical methods experience challenges.

Intercomparison of Deep Learning Architectures for the Prediction of Precipitation Fields With a Focus on Extremes

“…Then, we calculated the correlation patterns between the model outputs and the truth to quantify the ability of the networks to capture the spatial distribution of precipitation amounts. Additionally, the metric intersection over union (IoU) to evaluate the model's performance to classify heavy events (Prabhat et al., 2021; Wang et al., 2023) is calculated. The IoU was also quantified by considering an extreme event to be occurring for any probability higher than 0.5.…”

A precipitation downscaling method using a super-resolution deconvolution neural network with step orography

“…However, it is not clearly shown that adding orography improves the fine-scale spatial variability of downscaled precipitation products from coarse to a fine grid. Recently, Wang et al (2023) developed super-resolution deep residual network models with customised loss functions to downscale the hourly precipitation data (for a downscale factor of 12) over the northern region of the Gulf of Mexico and showed their model performs better than the quantile mapping based empirical downscaling technique. In these previous studies, super-resolution deep learning models are mostly based on CNNs.…”

“…For climate projections, this will be useful for developing region-specific climate adaptation strategies. For these purposes, recent deep learning methods, especially, the super-resolution approaches, are being widely used for empirical precipitation downscaling (Vandal et al, 2017; Harilal et al, 2021; Kumar et al, 2021; Chandra et al, 2022; Wang et al, 2023). In precipitation downscaling, super-resolution models are trained to learn the relationship between the coarse-scale precipitation data and the fine-scale precipitation data, and then apply this relationship to generate high-resolution precipitation maps.…”

“…Previous studies have developed super-resolution-based deep learning models for precipitation downscaling using a convolutional neural network (CNN) (Vandal et al, 2017; Chandra et al, 2022; Wang et al, 2023). For example, Vandal et al (2017) used the super-resolution CNN (SRCNN) model architecture (Dong et al, 2014) and applied it to the precipitation downscaling by introducing orography as an auxiliary input to the model and named it DeepSD.…”

Joint Generalized Neural Models and Censored Spatial Copulas for Probabilistic Rainfall Forecasting

“…To provide local-scale precipitation forecasting, it is common practice to downscale coarse CM forecasts to a finer grid with the use of post-processing methods. Similar to the works by Woo and Wong [2017], Vandal et al [2018], Adewoyin et al [2021], Xiang et al [2022] and Wang et al [2023], one of our main aims is to perform statistical downscaling of rainfall, that is predicting high-resolution precipitation from low-resolution weather variables. By conditioning on forecasts of future weather variables, our model can gain in reach due to the accuracy of CM weather forecasts on longer-time frames compared to statistical models directly predicting the weather.…”