Downscaling and Bias Correction of Extreme Tropical Cyclone Rainfall in ERA5 Using Deep Learning

Accurately modeling tropical cyclone (TC)-induced rainfall is an exceptionally challenging task due to the underlying complex interplay of dynamic and thermodynamic processes. Previous studies have discovered that the ERA5 reanalysis, which is commonly used as input to hydrological models that analyze TC-induced floods, underestimates heavy precipitation events. However, there are only few existing studies that perform bias correction of ERA5 precipitation data, and to our knowledge, none of them correct ERA5 biases for TC-associated precipitation. Additionally, most of the existing works focus on adjusting pixel-level or image-level metrics of bias by computing the average per-pixel or per-image difference in rainfall amounts compared to some reference. However, it is equally important to ensure that the peaks of rainfall are correctly located within the rainfall maps, especially if these maps are then used as inputs to hydrological models. Therefore, our research aims to address two literature gaps: bias correction of ERA5 rainfall specifically for TC-induced rainfall, and the use of spatial metrics of bias—not just per-pixel and per-image ones—to assess the goodness of the bias adjustment. As our reference against which to bias adjust ERA5 we used the MSWEP dataset, a high-fidelity, multi-source rainfall product. The analysis is based on four variables: total precipitation, to be bias adjusted; temperature (at 850 hPa), relative humidity (at 850 hPa), and vertically integrated water content to supply additional information to our model. We extracted from the IBTrACS archive the location of all TCs between 2001 and 2019, eliminating the records of storms that never achieved TC status (i.e., wind speed > 34 knots). For our model, we used an Attention Residual U-Net (ARU-Net), a modern variant of the popular convolutional neural network U-Net. As our goal was to bias-correct not just the amount of total rainfall in an image but also the location of the peaks of rainfall, we designed a loss function that would capture both aspects. Specifically, we used a modified version of a spatial verification metric frequently used in atmospheric science and hydrology: the Fractions Skill Score (FSS). In short, our implementation of the FSS quantifies the similarity of patterns of rainfall of two images in 15 x 15 pixels patches for the 10, 5, and 1% most intense rainfall in the input grids. The loss function we used to train ARU-Net was a combination of the FSS thus defined and of the Mean Squared Error (MSE), both calculated between the output of ARU-Net and the corresponding MSWEP rainfall map; we refer to this loss as “compound loss”. To validate our results, we also considered two other loss functions: a plain MSE, and a loss function proposed by Hess and Boers (2023), who also developed a U-Net model for the bias adjustment of rainfall maps. We trained three identical models using the three different loss functions (ARU-NetMSE, ARU-Netcompound, ARU-NetHess), and evaluated their performance on the test set using the following metrics: FSS (calculated for the 80th, 95th, and 99th percentiles), Mean Absolute Error (MAE), MSE, spatial correlation, the absolute value of the difference between the total rainfall of the model and of MSWEP (“absolute delta bias”), and the distance between the peak (i.e., the pixel with highest intensity) in the model’s output and the peak in the MSWEP map. The results show how ARU-Netcompound, compared to ERA5, improves all metrics by 3-25% (statistically significant at p<0.05), whereas the other models fail to improve—and indeed, they make worse—the absolute delta bias. In conclusion, our work presents a novel approach to the bias correction of ERA5 rainfall data, specifically targeting TC-induced rainfall, by incorporating spatial metrics and a modern deep learning architecture guided by a novel loss function, which we show outperforms other loss functions present in the literature. Our work holds potential as groundwork for improving flood prediction models, thereby contributing to more effective disaster management and mitigation strategies.