{"paper_id":"41e6a136-397f-43bb-a05e-b463a5f403fa","body_text":"Bias-targeted deep learning enhances short-range heavy rainfall forecasts | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Bias-targeted deep learning enhances short-range heavy rainfall forecasts Tao Tang, Wenqiang Shen, Jiaolan Fu, Ping He, Hao Qian, Ling Luo This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8314049/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 06 Mar, 2026 Read the published version in npj Climate and Atmospheric Science → Version 1 posted 14 You are reading this latest preprint version Abstract Improving short-range forecasts of heavy rainfall remains a challenge. In artificial intelligence (AI) era, deep-learning-based methods have replaced traditional statistical post-processing for correcting precipitation forecasts from numerical weather prediction (NWP) models, but their performances are constrained by the non-negative and heavy-tailed nature of rainfall. Mainstream studies tried to solve this problem by redesigning loss functions or constructing hybrid models, yet they struggled to achieve both simplicity and transferability. Here we show that the biases between NWP forecasts and observations in heavy rainfall events follow an approximately Gaussian distribution. Accordingly, this study trains a multi-task U-Net that uses precipitation biases as the target. This bias-targeted strategy produces stable and substantial enhancements in short-range heavy rainfall forecasts, with improvements exceeding 21% across four in five regions of China. The findings highlight the critical role of target selection in deep-learning-based post-processing and provide a simple and effective pathway for advancing heavy rainfall forecasts. Earth and environmental sciences/Climate sciences Earth and environmental sciences/Environmental sciences Earth and environmental sciences/Hydrology Physical sciences/Mathematics and computing Earth and environmental sciences/Natural hazards multi-task deep learning bias correction short-range prediction heavy rainfall forecasts Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Heavy rainfall can trigger flooding, landslides, debris flows, and urban waterlogging, posing a substantial threat to socioeconomic development and to the safety of human lives and properties 1 , 2 . With rising greenhouse gas concentrations, the frequency of heavy rainfall events has increased 3 – 5 , and such events are likely to become even more frequent in the future 6 , 7 . Therefore, enhancing the forecasting and early-warning capability for heavy rainfall events are crucial for effective disaster preparedness and mitigation. Real-time observations, such as radar echoes and satellite imagery, have substantially improved heavy rainfall prediction at nowcasting lead times 8 – 10 , yet short-range forecasts from the NWP models, which solve the nonlinear, non-closed fluid dynamical equations, are better suited to meet the temporal demands of heavy rainfall disaster preparedness. Over recent decades, NWPs forecast skills have increased significantly 11 – 14 , however, coarse spatial resolution requires sub-grid parameterizations for many small-scale precipitation processes, inevitably introducing systematic biases 15 – 17 . Post-processing of NWP outputs is recognized as one of the most effective methods to reduce these biases and further enhance forecast skills 18 – 35 . Traditional statistical methods correct biases in the NWP outputs based on the long-term linear statistical relationship between model forecasts and observations 18 – 22 . A key underlying assumption of these methods is that the precipitation biases in NWP are stationary and independent of weather systems, which prevents them from effectively capturing the bias characteristics associated with different weather systems 27 , 31 , 35 , resulting in the limited post-processing precision and suboptimal performance in heavy rainfall forecasts 27 , 31 . In recent years, AI algorithms have been widely applied to the post-processing of NWP due to their ability to handle non-linear problems and capture multi-variable features, often outperforming traditional statistical methods 23 – 35 . Nevertheless, the strictly non-negative and heavy-tailed nature of precipitation still limit the effectiveness of deep learning post-processing methods in heavy rainfall forecasts 27 , 29 – 31 , 36 . Previous studies have made valuable progress by modifying loss functions 26 – 29 or by designing increasingly intricate hybrid models 30 – 33 . However, these methods often struggle to balance practical simplicity with transferability. Specifically, methods based on loss-function modification typically require fine-tuning when applied to a new study region, whereas hybrid models generally entail a rapidly increasing parameter demand. Notably, most of these studies rely on observed precipitation as the ground truth 23 – 35 , and few have exploited precipitation biases themselves 36 . Here, we train models using both precipitation biases and observed precipitation as the targets to correct the short-range rainfall forecasts derived from European Centre for Medium-Range Weather Forecast Integrated Forecasting System (ECMWF). By comparing the results with each other, this study highlights that using precipitation biases as the target more effectively enhances short-range heavy rainfall forecasts. Results Biases of ECMWF Precipitation Forecasts over the Yangtze River Delta (YRD) The YRD (116-123 \\(\\:^\\circ\\:\\) E, 27-34 \\(\\:^\\circ\\:\\) N) is one of the most densely populated and economically vital regions in China, yet its complex terrain (Fig. 1 a) and exposure to multiple weather systems, such as Mei-yu front and tropical cyclones 37 , render it highly vulnerable to heavy rainfall 2 . Enhancing short-range heavy rainfall forecast skill in the YRD is therefore of critical importance. A rainy-day sampling method (see Methods) is applied to the China Meteorological Administration Multisource Precipitation Analysis 38 (CMPA) dataset to reduce the dominance of clear-sky cases over the YRD (Supplementary Table 1). The corresponding 3-hourly ECMWF precipitation forecasts initialized at 00 UTC at 15–36 hours lead times are then evaluated against the CMPA observations. A pointwise comparison reveals that 44.5% of grids are clear-sky hits, 19.2% are false alarms, and the remaining 36.2% correspond to rainy conditions in the CMPA observations, comprising both ECMWF hits and misses (Fig. 1 b, c). For rainy condition hits, both the CMPA and the ECMWF forecasts exhibit long-tailed distributions, with decreasing frequencies toward higher rainfall intensities. In general, the 3-hourly precipitation over the YRD falls within 0–50 mm (Fig. 1 c). ECMWF underestimates the grid numbers with 3-hourly precipitation both less than 3 mm and larger than 15 mm, indicating systematic misses at both ends of the intensity spectrum. Likewise, both the positive and negative precipitation biases of ECMWF display the long-tailed distributions (Fig. 1 d). Since the negative biases occurring far more frequently than positive ones, the ECMWF precipitation forecasts tend to underestimate rainfall intensity. More importantly, the approximate Gaussian distribution of the 3-hourly precipitation biases at grid points with daily rainfall exceeding 50 mm, provides a strong rationale for adopting a bias-targeted modeling strategy to enhance heavy rainfall forecasts (Fig. 1 d). Results based on the ECMWF precipitation forecasts initialized at 12 UTC exhibit similar characteristics (Supplementary Fig. 1). Bias-targeted deep learning improves heavy rainfall forecasts over the YRD According to the characteristics of biases in the ECMWF precipitation forecasts over the YRD, grids in all samples can be classified into four types: clear-sky hits, false alarms, rainy condition hits with under-predicted intensity (including misses), and rainy condition hits with over-predicted intensity. Correspondingly, this study develops a multi-task 39 , 40 U-Net 41 with four output branches (see Methods, Fig. 2 ) that uses thirty-eight ECMWF variables and topography dataset as the inputs (Supplementary Table 2) and the ECMWF precipitation biases as the target (i.e., ground truth). The model predicts the precipitation bias at each grid point, and these predicted biases are subsequently subtracted from the raw ECMWF forecasts to yield the corrected precipitation forecasts (UnetDif). For comparison, a three output branches U-Net is trained using the CMPA observations as the ground truth to predict precipitation directly (UnetOri), a configuration aligned with previous studies 27–29,31−35 . In order to match the time steps used in the operational short-range predictions, bias corrections are applied to the 3-hourly ECMWF precipitation forecasts, while evaluated on the 24-hour accumulated precipitation over the testing period (Supplementary Table 1). In addition, due to the “double penalty” problem inherent in root mean square error 28 , 35 , threat score 42 (TS, see Methods), equitable threat score 43 (ETS, see Methods), bias score 44 (Bias, see Methods) and false alarm ratio 44 (FAR, see Methods) are used to evaluate the precipitation forecast skills. At each lead time, the ECMWF precipitation forecasts initialized at 00 and 12 UTC are merged and evaluated collectively. At a 36-hour lead time, UnetOri and UnetDif improve the clear-sky and rainy-day hit ratios (ACC; see Methods) to 0.820 and 0.817, respectively, compared with 0.774 for the raw ECMWF forecasts, primarily by reducing false alarms (Supplementary Fig. 2a, c; Supplementary Table 3). They also enhance the TS by ~ 6% and ~ 4% for daily rainfall above 10 mm, and by ~ 8% and ~ 6% for daily rainfall above 25 mm, respectively. The two deep-learning-based post-processing methods achieve these improvements through different mechanisms. Specifically, UnetOri exhibits lower Bias values (1.039 and 0.730 for daily rainfall above 10 mm and 25 mm), indicating fewer false alarms relative to the raw ECMWF forecasts. By contrast, UnetDif shows higher Bias values, suggesting fewer misses but more false alarms than the raw forecasts (Supplementary Fig. 2a, c; Supplementary Table 3). Across the YRD, the TS for heavy rainfall (daily rainfall above 50 mm) drops markedly from 0.194 in ECMWF to 0.131 (~ 32%) in UnetOri (Fig. 3 a). The reduced TS and lower Bias at most grid points (Fig. 3 b, c, e, f) indicate that UnetOri generates more misses in heavy rainfall events (Supplementary Fig. 2b, d; Supplementary Table 3). In contrast, UnetDif increases the regional TS by ~ 22% relative to ECMWF, reaching 0.237 (Fig. 3 a). TS improves at most grid points, while Bias increases where it was initially lower than 1 and decreases or remains unchanged where it was higher than 1 (Fig. 3 d, g). Furthermore, UnetDif yields a probability distribution of heavy rainfall grids that more closely matches the CMPA observations, particularly for daily rainfall above 75 mm (Supplementary Fig. 2b, d), demonstrating that UnetDif provides more accurate heavy rainfall forecasts than both ECMWF and UnetOri. The advantages of UnetDif for short-range heavy rainfall forecasts can be further illustrated by comparing TS of the top 20 events in 2025 over the YRD (Supplementary Table 4). At a 36-hour lead time, quantile mapping 18 , 19 (QM, see Methods), UnetOri, and UnetDif exceed ECMWF in 13, 4, and 15 cases, respectively. Spatial distributions of the precipitation predictions suggest that UnetDif effectively compensates for the heavy rainfall misses in the ECMWF forecasts, while UnetOri and QM do not (Supplementary Fig. 3). Furthermore, at 48- and 60-hour lead times, UnetDif enhances heavy rainfall TS relative to ECMWF by ~ 21% and ~ 30%, respectively. QM improves TS by ~ 1.1% and ~ 8%, whereas UnetOri decreases TS by ~ 14% and ~ 23%. Results evaluated by ETS are similar (Supplementary Table 3). It is worth noting that the best Bias for QM, accompanied by its nearly unchanged TS and ETS relative to ECMWF, may result from its highest FAR compared with ECMWF, UnetOri and UnetDif (Supplementary Table 3). Bias-targeted deep learning improves heavy rainfall forecasts across different regions of China. Keeping the computational procedure unchanged, the bias-targeted deep learning methods are further applied to four major precipitation regions of China (Fig. 4 a). Specifically, North China (108-112 \\(\\:^\\circ\\:\\) E, 32-45 \\(\\:^\\circ\\:\\) N), East China (112-123 \\(\\:^\\circ\\:\\) E, 22-36 \\(\\:^\\circ\\:\\) N), South China (104-118 \\(\\:^\\circ\\:\\) E, 18-28 \\(\\:^\\circ\\:\\) N) and Southwest China (89-112 \\(\\:^\\circ\\:\\) E, 20-35 \\(\\:^\\circ\\:\\) N). Furthermore, rainy-day sampling period is shortened to 2019–2023, with 2024–2025 as the testing period, to strengthen the robustness of the hindcast evaluation (Supplementary Table 1). Consistent with the results over the YRD, UnetDif improves heavy rainfall forecasts in all four regions relative to ECMWF. Specifically, TS (ETS) in UnetDif increases by ~ 4%, ~ 23%, ~ 26%, and ~ 32% in North, East, South, and Southwest China, respectively. In contrast, while UnetOri increases TS (ETS) by ~ 6.4% in Southwest China, it decreases TS (ETS) by approximately 33%, 1.4%, and 0.07% in North, East and South China, respectively (Fig. 4 b-e). A further in-depth analysis suggests that Bias increases from 0.838 to 1.314 in East China and from 0.711 to 1.027 in South China, while FAR remains nearly unchanged, indicating that UnetDif effectively reduces ECMWF misses and thereby enhances heavy rainfall forecasts (Fig. 4 c, d). In Southwest China, reductions in both Bias and FAR suggest that UnetDif enhances TS primarily by decreasing false alarms (Fig. 4 e). In North China, TS, ETS, Bias, and FAR all remain nearly unchanged (Fig. 4 b), this poor performance may be due to the insufficient training samples 45 , 46 (Supplementary Table 1). The evaluation results remain similar when using 2024 or 2025 alone as the testing period (Supplementary Fig. 4). Among the top five heavy rainfall events in each of the four regions at a 36-hour lead time, QM, UnetOri, and UnetDif exceed the raw ECMWF forecasts in 14, 12, and 16 cases, respectively (Supplementary Table 5). For the event in each region with the largest UnetDif improvement relative to the ECMWF forecasts, the spatial distributions of precipitation forecasts indicate that UnetDif effectively compensates for ECMWF misses, particularly along the narrow heavy rainfall bands in coastal area in East and South China, whereas QM and UnetOri fail to do so (Supplementary Fig. 5f-o). Once again, the barely changed TS but improved Bias of QM suggest that it generates more false alarms rather than hits of heavy rainfall (Supplementary Fig. 5c, h, m, r). However, predicting daily rainfall exceeding 100 mm remains highly challenging. Both UnetDif and UnetOri exhibit substantial misses for these extreme precipitation grid points (Supplementary Fig. 4i, j, n, s, t), whereas QM generates numerous false alarms, even above 250 mm per day. Discussion To the best of our knowledge, this study is the first to demonstrate that training deep learning models using NWP precipitation biases as target can substantially improve short-range heavy rainfall prediction skill. A crucial starting point arises from a quiet but telling clue in the data, namely that both the positive and negative 3-hourly biases associated with daily rainfall above 50 mm show a clear Gaussian distribution. This statistical characteristic is not merely a curiosity, but also the fundamental reason why this bias-targeted deep learning method can so effectively enhance heavy rainfall forecasts. In order to improve the short-range heavy rainfall forecasts, previous studies have made valuable progress by modifying loss functions or by designing increasingly intricate hybrid models. Yet these methods often struggle to balance practical simplicity with transferability. In contrast, the method we proposed in this study remains simple, since the only preprocessing step is converting observed precipitation into NWP forecast biases. It is also practically robust, as it delivers strong performance across multiple regions of China. This combination of elegance and utility, we believe, offers a useful blueprint for future studies. It is worth noting that existing methods such as the loss function modifications or the hybrid deep learning models are not incompatible with our framework. On the contrary, they can be integrated with bias-targeted deep learning and may yield further improvements in short-range heavy rainfall forecasts. Even so, several limitations remain. UnetOri slightly outperforms UnetDif for low-threshold rainfall, highlighting the potential of combining their strengths for short-range forecasts across varying rainfall intensities; Both deep-learning-based methods, however, still struggle with daily rainfall exceeding 100 mm, likely due to the extreme scarcity of such events. This points to an unavoidable but important conclusion: the longer, high-quality observational and NWP forecast datasets are essential for further advances in deep-learning-based bias correction methods 45 , 46 . Methods Datasets The 3-hourly ECMWF forecast datasets from January 2019 to October 2025 are used in this study. ECMWF forecasts issue twice a day at 00 UTC and 12 UTC, and the forecast lead times are 0-240 h. For short-range forecasts, lead times of 15–60 h are used in this study. More specifically, thirty-eight meteorological variables provided by ECMWF are used here (Supplementary Table 2), including multiple pressure level fields of relative humidity, specific humidity, vertical velocity, temperature, zonal and meridional wind components, potential vorticity, and geopotential height, all with a spatial resolution of 0.25 \\(\\:^\\circ\\:\\) \\(\\:\\times\\:\\) 0.25 \\(\\:^\\circ\\:\\) . Moreover, single-level variables, including total column water vapor, total cloud cover, low cloud cover, large-scale precipitation, convective precipitation, and precipitation at 0.125° × 0.125° resolution are also utilized. The observed precipitation dataset used in this study is the high-quality China Meteorological Administration Multisource Precipitation Analysis System (CMPA) dataset 38 , which integrates three types of precipitation observations (i.e., gauge, satellite and radar) with spatial resolution of 0.05° × 0.05° and temporal resolution of 1 h. The 3-hourly observed precipitation is calculated by summing three 1-hourly records. The topography dataset with a resolution of 0.05° × 0.05° is downloaded from the National Earth System Science Data Center. All datasets have been interpolated to a resolution of 0.125° × 0.125° using the bilinear interpolation. Rainy day sampling method A day during the sampling period is recognized as a rainy day if at least 10% of grid points have precipitation ( \\(\\:\\ge\\:\\) 0.1 mm) in all 3-hourly intervals of that day in CMPA. The corresponding ECMWF forecasts initialized at 00 and 12 UTC with lead times of 36, 48, and 60 hours are then retrieved separately. Each identified rainy day contains eight 3-hourly samples. Among all samples, 85% are used for model training and the remaining 15% for validation. Multi-task U-Net model U-Net 41 is a classical image segmentation network composed of down-sampling, up-sampling, and skip connections, and it has been widely applied in precipitation bias correction studies 27 , 28 , 30 , 32 . In this work, we insert a batch normalization layer between each convolutional layer and its subsequent ReLU activation to facilitate convergence and improve model accuracy 28 . Furthermore, we extend the original single-task U-Net to multi-task 39 , 40 outputs with four and three output branches, respectively (Fig. 2 ). Specifically, the four output branches U-Net (UnetDif) comprises two classification branches that generate clear-sky and false-alarm masks, together with two regression branches that predict positive and negative precipitation biases. The predicted biases are first subtracted from the raw ECMWF precipitation forecasts, after which the two masks are applied to suppress spurious rainfall at non-rainy grid points. The three output branches U-Net (UnetOri) retains the two classification branches but uses a single regression branch to directly predict precipitation. The predicted precipitation is subsequently filtered with the clear-sky and false alarm masks to remove spurious rainfall over non-rainy grid points. Thirty-eight ECMWF forecast variables and topography are used as inputs for both UnetDif and UnetOri, arranged as separate channels. The ground truth for UnetDif is the ECMWF precipitation biases, while for UnetOri, it is the CMPA precipitation. Separate UnetDif and UnetOri models are constructed for each of the three lead times (36, 48, and 60-hour) over the YRD and each of the four regions (Supplementary Table 1). All of the thirty-eight ECMWF input variables are normalized using the Z-score method, while topography is scaled using min-max normalization to remove the influence of different variable magnitudes on network training. In all experiments, the batch size is set to 16 and the learning rate to 1 × 10⁻⁴. Loss function The total loss function designed for UnetDif consists of six components and is formulated as follows: \\(\\:{L}_{total}^{UnetDif}={\\stackrel{\\sim}{L}}_{focal}^{cs}+{\\stackrel{\\sim}{L}}_{focal}^{fa}+{\\stackrel{\\sim}{L}}_{MSE}^{Pos}+{\\stackrel{\\sim}{L}}_{MSE}^{neg}+{0.1*\\stackrel{\\sim}{L}}_{mae}^{pos}+{0.1*\\stackrel{\\sim}{L}}_{mae}^{neg}\\) $$\\:{\\stackrel{\\sim}{L}}_{i}=\\frac{{L}_{i}}{{L}_{i}^{\\left(1\\right)}}$$ Where \\(\\:{\\:\\stackrel{\\sim}{L}}_{i}\\) denotes the \\(\\:i\\) -th loss component normalized by its value from the first training epoch, prevents the total loss from dominating by terms with inherently larger magnitude. \\(\\:{\\stackrel{\\sim}{L}}_{focal}^{cs}\\) and \\(\\:{\\stackrel{\\sim}{L}}_{focal}^{fa}\\) are the normalized focal loss 47 calculated on clear-sky hit and false alarm grid points, respectively. \\(\\:{\\stackrel{\\sim}{L}}_{MSE}^{pos}\\) and \\(\\:{\\stackrel{\\sim}{L}}_{MSE}^{neg}\\) represent the normalized mean squared errors (MSE) computed on grid points where precipitation biases are positive and negative, respectively. \\(\\:{\\stackrel{\\sim}{L}}_{MAE}^{pos}\\) and \\(\\:{\\stackrel{\\sim}{L}}_{MAE}^{neg}\\) denote the mean absolute errors (MAE) evaluated on clear-sky hit grid points, ensuring that the predicted positive and negative biases are set to zero at those non-rainy grid points. Since precipitation is strictly non-negative, the negative bias components used in UnetDif, namely, \\(\\:{\\stackrel{\\sim}{L}}_{MSE}^{neg}\\) and \\(\\:{\\stackrel{\\sim}{L}}_{MAE}^{neg}\\) are excluded from the loss design for UnetOri. Accordingly, the total loss function for UnetOri is formulated as: \\(\\:{L}_{total}^{UnetOri}={\\stackrel{\\sim}{L}}_{focal}^{cs}+{\\stackrel{\\sim}{L}}_{focal}^{fa}+{\\stackrel{\\sim}{L}}_{MSE}^{Pos}+{0.1*\\stackrel{\\sim}{L}}_{mae}^{pos}\\) Here, \\(\\:{\\stackrel{\\sim}{L}}_{focal}^{cs}\\) and \\(\\:{\\stackrel{\\sim}{L}}_{focal}^{fa}\\) follow the same definitions as those in the UnetDif. The term \\(\\:{\\stackrel{\\sim}{L}}_{MSE}^{Pos}\\) represents the normalized MSE computed on grid points where precipitation is observed. The term \\(\\:{\\stackrel{\\sim}{L}}_{MAE}^{pos}\\) denotes the MAE evaluated on clear-sky hit grid points, ensuring that the predicted precipitation is constrained to zero at those non-rainy grid points. It is worth noting that the optimal model parameters of UnetDif and UnetOri are saved at the epoch when the cumulative TS across the 3-hourly precipitation bins (i.e., 0.1, 1, 3, 5, 7, 9 and 11 mm) reach the maximum, rather than when the total losses attain their minimum. Evaluation metrics The clear-sky and rainy-day hit ratio (ACC), TS, ETS, Bias and false alarm ratio (FAR) 44 are used to evaluate the performance of the precipitation forecasts in ECMWF and each post-processing method. They are calculated as follows: $$\\:ACC=\\frac{H+Z}{H+M+F+Z}$$ $$\\:TS=\\frac{H}{H+M+F}$$ $$\\:BS=\\frac{H+F}{H+M}$$ $$\\:FAR=\\frac{F}{H+F}$$ $$\\:ETS=\\frac{H-{H}_{r}}{H+M+F-{H}_{r}}$$ $$\\:{H}_{r}=\\frac{(H+M)(H+F)}{N}$$ Where H, M, F, and Z are the numbers of correct forecasts of occurrence (i.e., hits), incorrect forecasts of non-occurrence (i.e., misses), incorrect forecasts of occurrence (i.e., false alarms) and correct forecasts of clear-sky, respectively. \\(\\:{H}_{r}\\) is the number of hits expected by random chance over N total forecasts. Quantile mapping methods Due to the strong seasonality of precipitation in nature, a localized time-window sampling strategy is employed for quantile mapping 18 , 19 (QM) to generate the historical cumulative distribution function (CDF). Specifically, for a target date 𝐷 to be bias-corrected, QM training samples are selected from the 30 days preceding 𝐷 in the same year, the 30 days preceding 𝐷 in the previous year, the 30 days following 𝐷 in the previous year, and the 30 days following 𝐷 in the year before last. The historical CDF is constructed from these samples, and the ECMWF precipitation forecasts are subsequently bias-corrected using QM as follows: $$\\:{Pr}_{QM}={F}_{CMPA}^{-1}\\left({F}_{ECMWF}\\right({Pr}_{ECMWF})$$ Where \\(\\:{F}_{CMPA}\\) and \\(\\:{F}_{ECMWF}\\) are the CDFs of the CMPA and the ECMWF forecasts of the selected training samples, respectively. \\(\\:{F}_{CMPA}^{-1}\\) denotes the inverse CDF of the CMPA. \\(\\:{Pr}_{QM}\\) and \\(\\:{Pr}_{ECMWF}\\) denote the precipitation forecasts after QM correction and the raw ECMWF forecasts, respectively. Declarations Competing financial interests The authors declare no competing financial interests. Funding This work was supported by the Meteorological Joint Funds of National Natural Science Foundation of China (U2442204) and the Joint Research Project for Meteorological Capacity Improvement (23NLTSZ007). T.T. was supported by Open Fund Project for Heavy Rain (BYKJ2024Q24). W.S. was supported by the Special Project of Innovation and Development of China Meteorological Administration (CXFZ2023J021). H.Q. was supported by Key Innovation Team Project of China Meteorological Administration for Intelligent Forecasting Technology (CMA2022ZD04). Author Contribution T.T. and W.S. are co-first authors. T.T. conceived the central idea of the study. T.T. and W.S. designed the AI models and performed the hindcast experiments. T.T. prepared the figures and wrote the paper under supervision of H.Q. and L.L. T.T. W.S. J.F. P.H. H.Q. and L.L. contributed to interpreting results, discussion of associated mechanisms, and improvement of the writing. Data availability Data related to this paper can be downloaded from: ECMWF 3-hourly forecasts, https://www.ecmwf.int/en/forecasts/datasets; CMPA observed precipitation datasets, https://data.cma.cn/data/detail/dataCode/SURF_CMPA_RT_NC/keywords/; The global topography datasets, https://www.geodata.cn/main/face_science_detail?typeName=face_science&guid=201519481253546. Code availability Codes used in this study are available from the corresponding authors on request. References Jonkman, S. N. Global perspectives on loss of human life caused by floods. Nat. Hazards . 34, 151–175 (2005). Gong, L., Zhang, X., Liu, J. & Gui, H. Exploring the influence of urban agglomeration on extreme precipitation: evidence from the middle reaches of the Yangtze River, China. J. Hydrol. Reg. Stud. 55, 101932 (2024). Fischer, E. M. & Knutti, R. Observed heavy precipitation increase confirms theory and early models. Nat. Clim. Change. 6, 986–991 (2016). Guerreiro, S. B. et al. Detection of continental-scale intensification of hourly rainfall extremes. Nat. Clim. Change 8, 803–807 (2018). Ham, Y. G. et al. Anthropogenic fingerprints in daily precipitation revealed by deep learning. Nature 622, 301–307 (2023). O'Gorman, P. A. & Schneider, T. The physical basis for increases in precipitation extremes in simulations of 21st-century climate change. Proc. Natl. Acad. Sci. 106, 14773–14777 (2009). Zhang, W. et al. Increasing precipitation variability on daily-to-multiyear time scales in a warmer world. Sci. Adv. 7, eabf8021 (2021). Zhang, Y. et al. Skillful nowcasting of extreme precipitation with NowcastNet. Nature 619, 526–532 (2023). Du, X. & Guo, H. A multi-scale attention encoding and dynamic decoding network designed for short-term precipitation forecasting. Earth Sci. Inform. 18, 41 (2025). Xiao, S., Zhang, A., Gong, J., Chen, Y., Chen, S. & Li, W. An explainable deep learning method with squeeze and excitation block for 6-hour forecast of tropical cyclone remote precipitation. J. Geophys. Res.: Mach. Learn. Comput. 2, e2025JH000675 (2025). Bauer, P., Thorpe, A. & Brunet, G. The quiet revolution of numerical weather prediction. Nature 525, 47–55 (2015). Stoelinga, M. T. et al. Improvement of microphysical parameterization through observational verification experiment. Bull. Amer. Meteor. Soc. 84, 1807–1826 (2015). Thompson, G., Field, P. R., Rasmussen, R. M. & Hall, W. D. Explicit forecasts of winter precipitation using an improved bulk micro-physics scheme. Part II: Implementation of a new snow parameterization. Mon. Wea. Rev. 136, 5095–5115 (2008). Lee, J., Min, K. & Lim, K. Comparing 3DVAR and hybrid radar data assimilation methods for heavy rain forecast. Atmos. Res. 270, 106062 (2022). Boyle, J. & Klein, S. A. Impact of horizontal resolution on climate model forecasts of tropical precipitation and diabatic heating for the twp-ice period. J. Geophys. Res.: Atmos . 115, D23113 (2010). Wilcox, E. M. & Donner, L. J. The frequency of extreme rain events in satellite rain-rate estimates and an atmospheric general circulation model. J. Clim. 20, 53–69 (2007). Zadra, A. et al. Systematic Errors in Weather and Climate Models: Nature, Origins, and Ways Forward. Bull. Amer. Meteor. Soc. 99, 67–70 (2018). Wood, A. W., Maurer, E. P., Kumar, A. & Lettenmaier, D. Long-range experimental hydrologic forecasting for the eastern United States, J. Geophys. Res.: Atmos. 107, 4429 (2002). Zhao, T. et al. How suitable is quantile mapping for postprocessing GCM precipitation forecasts? J. Clim. 30, 3185–3196 (2017). Robertson, D. E., Shrestha, D. L. & Wang, Q. J. Post-processing rainfall forecasts from numerical weather prediction models for short-term streamflow forecasting. Hydrol. Earth Syst. Sci. 17, 3587–3603 (2013). Li, L. et al. Improving categorical and continuous accuracy of precipitation forecasts by integrating empirical quantile mapping and bernoulli-gamma-gaussian distribution. Atmos. Res. 298, 107133 (2024). Javanshiri, Z. & Pakdaman, M. Bayesian joint probability approach for post-processing monthly precipitation prediction in Northwest Iran. Water Resour. Res. 60, e2023WR036846 (2024). Zhang, B. et al. A machine learning-based observational constraint correction method for seasonal precipitation prediction. Adv. Atmos. Sci. 42, 36–52 (2025). Rasp, S. & Lerch, S. Neural networks for postprocessing ensemble weather forecasts. Mon. Wea. Rev. 146, 3885–3900 (2018). Ebert-Uphoff, I. & Hilburn, K. Evaluation, tuning and interpretation of neural networks for working with images in meteorological applications. Bull. Amer. Meteor. Soc. 101, 2149–2170 (2020). Hu, Y. F., Yin, F. K. & Zhang, W. M. Deep learning-based precipitation bias correction approach for Yin-He global spectral model. Meteor. Appl. 28, e2032 (2021). Hess, P. & Boers, N. Deep learning for improving numerical weather prediction of heavy rainfall. J. Adv. Model. Earth Syst. 14, e2021MS002765 (2022). Sun, D. et al. Deep learning improves GFS wintertime precipitation forecast over southeastern China. Geophys. Res. Lett. 50, e2023GL104406 (2023). Hunt, K. M. Stop using root-mean-square error as a precipitation target! arXiv preprint arXiv:2509.08369 (2025). Lyu, Y. et al. Significant advancement in subseasonal-to-seasonal summer precipitation ensemble forecast skills in China mainland through an innovative hybrid CSG-UNET method. Environ. Res. Lett. 19, 074055 (2024). Zhao, C., Ye, A., Wu, L. & Zhan, S. A novel deep learning model for post-processing of short- and medium-term daily precipitation forecasts. Atmos. Res. 326, 108319 (2025). Cao, H. et al. UTrans-Net: a model for short-term precipitation prediction. Artif. Intell. Appl. 1, 90–97 (2022). Jiang, Q. et al. Hybrid multilayer perceptron and convolutional neural network model to predict extreme regional precipitation dominated by the large-scale atmospheric circulation. Atmos. Res. 304, 107362 (2024). Zhang, L., Zhao, Y., Cen, Y. & Lu, M. Deep learning-based precipitation simulation for tropical cyclones, mesoscale convective systems, and atmospheric rivers in East Asia. J. Geophys. Res.: Atmos. 129, e2024JD041914 (2024). Chen, G. & Wang, W.-C. Short-term precipitation prediction for contiguous United States using deep learning. Geophys. Res, Lett. 49, e2022GL097904 (2022). Ling, F., Li, Y., Luo, J.-J., Zhong, X. & Wang, Z. Two deep learning-based bias-correction pathways improve summer precipitation prediction over China. Environ. Res. Lett. 17, 124025 (2022). Tang, T. et al. Robust and explainable objective classification of extreme precipitation events using machine learning. J. Meteor. Res. 39, 1–14 (2025). Li, S., Huang, X., Wu, W., Du, B. & Jiang, Y. Evaluation of CMPAS precipitation products over Sichuan, China. Atmos. Oceanic Sci. Lett. 15, 100129 (2022). Li, Y. et al. Convective storm VIL and lightning nowcasting using satellite and weather radar measurements based on multi-task learning models, Adv. Atmos. Sci. 40, 887–899 (2023). Ling, F. et al. Multi-task machine learning improves multi-seasonal prediction of the Indian Ocean Dipole. Nat Commun 13, 7681 (2022). Ronneberger, O., Fischer, P. & Brox, T. U-Net: Convolutional networks for biomedical image segmentation. Med. Image Comput. Comput. Assist. Interv. 9351, 234–241 (2015). Donaldson, R., Dyer, R. M. & Kraus, M. J. An objective evaluator of techniques for predicting severe weather events. In Ninth conference on severe local storms 321–326 (1975). Winterrath. T. & Rosenow, W. A new module for the tracking of radar-derived precipitation with model-derived winds. Adv. Geosci. 10, 77–83 (2007). Wilks, D. S. Statistical Methods in the Atmospheric Sciences. Academic Press (2020). Srivastava, N. et al. Dropout: A simple way to prevent neural networks from overfitting. J. Mach. Learn. Res. 15, 1929–1958 (2014). Schultz, M. G. et al. Can deep learning beat numerical weather prediction? Philos. Trans. A Math. Phys. Eng. Sci. 379, 20200097 (2021). Lin, T.-Y., Goyal, P., Girshick, R., He, K. & Dollár, P. Focal loss for dense object detection. IEEE Trans. Pattern Anal. Mach. Intell. 42, 318–327 (2020). Additional Declarations No competing interests reported. Supplementary Files BiastargeteddeeplearningenhancesshortrangeheavyrainfallforecastsSupplementaryinformation.docx Cite Share Download PDF Status: Published Journal Publication published 06 Mar, 2026 Read the published version in npj Climate and Atmospheric Science → Version 1 posted Editorial decision: Revision requested 19 Jan, 2026 Reviews received at journal 11 Jan, 2026 Reviewers agreed at journal 04 Jan, 2026 Reviews received at journal 02 Jan, 2026 Reviewers agreed at journal 02 Jan, 2026 Reviewers agreed at journal 30 Dec, 2025 Reviewers agreed at journal 30 Dec, 2025 Reviewers agreed at journal 15 Dec, 2025 Reviewers agreed at journal 11 Dec, 2025 Reviewers agreed at journal 11 Dec, 2025 Reviewers invited by journal 11 Dec, 2025 Editor assigned by journal 11 Dec, 2025 Submission checks completed at journal 11 Dec, 2025 First submitted to journal 09 Dec, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {\"props\":{\"pageProps\":{\"initialData\":{\"identity\":\"rs-8314049\",\"acceptedTermsAndConditions\":true,\"allowDirectSubmit\":false,\"archivedVersions\":[],\"articleType\":\"Article\",\"associatedPublications\":[],\"authors\":[{\"id\":561289868,\"identity\":\"9683d9c4-95cc-4563-8f0c-ec08a99cf9a7\",\"order_by\":0,\"name\":\"Tao Tang\",\"email\":\"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA/ElEQVRIiWNgGAWjYBACxgYk9mMwxczcgFUpipYDQMwDVGvMwGAApBjxawEDqBY2abAWBgJamPsPP/78oeaO3X72w8eqCyr+RPO3A7X8qNiG22Ez0gwMDhx7ltzDk5Z2e8YZg9wZhxkbGHvO3MajhYch4QDb4WQehhyz27xtBrkNQC3MjG14tPSfYThw4B9QC/8bs2KQlvkEtTTkMDYcbDtsxyORY8YM0rKBoJYZacYMZ/sOJ/DceJYszXPGOHcjUMtBfH4xBIbYh4pvh+3Z+5MPfuapkMudd/7wwQc/KvBoaYDQiQ3IogdwqgcCeShtj0/RKBgFo2AUjHAAAOEgW+DbsVWRAAAAAElFTkSuQmCC\",\"orcid\":\"\",\"institution\":\"Zhejiang Meteorological Observatory\",\"correspondingAuthor\":true,\"prefix\":\"\",\"firstName\":\"Tao\",\"middleName\":\"\",\"lastName\":\"Tang\",\"suffix\":\"\"},{\"id\":561289879,\"identity\":\"887f57a9-fb89-4601-a298-e9bcc3de235f\",\"order_by\":1,\"name\":\"Wenqiang Shen\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Zhejiang Meteorological Observatory\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Wenqiang\",\"middleName\":\"\",\"lastName\":\"Shen\",\"suffix\":\"\"},{\"id\":561289883,\"identity\":\"f8b96180-9a54-4f88-bdfa-1b5264a99ea2\",\"order_by\":2,\"name\":\"Jiaolan Fu\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"National Meteorological Center\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Jiaolan\",\"middleName\":\"\",\"lastName\":\"Fu\",\"suffix\":\"\"},{\"id\":561289884,\"identity\":\"e26e0ca5-e499-406b-a199-6a67bd970302\",\"order_by\":3,\"name\":\"Ping He\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"School of Earth Science, Zhejiang University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Ping\",\"middleName\":\"\",\"lastName\":\"He\",\"suffix\":\"\"},{\"id\":561289890,\"identity\":\"89e49a42-613b-408d-98a1-29046bc34dfe\",\"order_by\":4,\"name\":\"Hao Qian\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Zhejiang Meteorological Observatory\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Hao\",\"middleName\":\"\",\"lastName\":\"Qian\",\"suffix\":\"\"},{\"id\":561289894,\"identity\":\"b41b7c84-6d54-4317-b4c9-62544d921bda\",\"order_by\":5,\"name\":\"Ling Luo\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Zhejiang Meteorological Observatory\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Ling\",\"middleName\":\"\",\"lastName\":\"Luo\",\"suffix\":\"\"}],\"badges\":[],\"createdAt\":\"2025-12-09 06:53:28\",\"currentVersionCode\":1,\"declarations\":\"\",\"doi\":\"10.21203/rs.3.rs-8314049/v1\",\"doiUrl\":\"https://doi.org/10.21203/rs.3.rs-8314049/v1\",\"draftVersion\":[],\"editorialEvents\":[{\"content\":\"https://doi.org/10.1038/s41612-026-01366-z\",\"type\":\"published\",\"date\":\"2026-03-06T15:58:42+00:00\"}],\"editorialNote\":\"\",\"failedWorkflow\":false,\"files\":[{\"id\":98451824,\"identity\":\"7011e63e-86c2-4066-9035-88cec1bf7fa9\",\"added_by\":\"auto\",\"created_at\":\"2025-12-17 17:33:37\",\"extension\":\"docx\",\"order_by\":0,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":2616150,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"Biastargeteddeeplearningenhancesshortrangeheavyrainfallforecastsmanuscript.docx\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8314049/v1/7c85b915a9433021412956bf.docx\"},{\"id\":98451831,\"identity\":\"d2101d6d-e7b6-451b-9e41-6382ebea478f\",\"added_by\":\"auto\",\"created_at\":\"2025-12-17 17:33:39\",\"extension\":\"json\",\"order_by\":1,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":7241,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"ad11a71c1d8f4f6e9ea19ae9038b5f90.json\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8314049/v1/3b19919ecb7bd2148acc329d.json\"},{\"id\":98451893,\"identity\":\"7514d788-a74c-48a4-9bfe-7c4e892fb56a\",\"added_by\":\"auto\",\"created_at\":\"2025-12-17 17:34:03\",\"extension\":\"docx\",\"order_by\":2,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":2047257,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"BiastargeteddeeplearningenhancesshortrangeheavyrainfallforecastsSupplementaryinformation.docx\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8314049/v1/ecf3a311d6f567991bd0f60b.docx\"},{\"id\":98451800,\"identity\":\"8c73f710-c93b-4d9f-8bc6-ad0dd4ba8048\",\"added_by\":\"auto\",\"created_at\":\"2025-12-17 17:33:25\",\"extension\":\"xml\",\"order_by\":3,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":98756,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"ad11a71c1d8f4f6e9ea19ae9038b5f901enriched.xml\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8314049/v1/25dbd0c17cfcbe807d84113a.xml\"},{\"id\":98451927,\"identity\":\"b776e486-322b-4752-8c12-0f1bd00e387a\",\"added_by\":\"auto\",\"created_at\":\"2025-12-17 17:34:15\",\"extension\":\"png\",\"order_by\":4,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":215299,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"floatimage1.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8314049/v1/d2c9ffd1f4d5bc47767a4e45.png\"},{\"id\":98451858,\"identity\":\"11defaf5-06fd-4cd4-9ab9-c1c35ddc6dd1\",\"added_by\":\"auto\",\"created_at\":\"2025-12-17 17:33:49\",\"extension\":\"emf\",\"order_by\":5,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":3483064,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"floatimage2.emf\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8314049/v1/ec7cf8f3d0e0db23c10dc3e2.emf\"},{\"id\":98451943,\"identity\":\"c60da42e-54a3-49a1-ab65-d5584f2c2492\",\"added_by\":\"auto\",\"created_at\":\"2025-12-17 17:34:20\",\"extension\":\"png\",\"order_by\":6,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":485228,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"floatimage3.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8314049/v1/7f07f0896901dde9c190ed4d.png\"},{\"id\":98451810,\"identity\":\"c3cbf24d-a09c-4fa5-8e46-fbfade20b3d5\",\"added_by\":\"auto\",\"created_at\":\"2025-12-17 17:33:32\",\"extension\":\"png\",\"order_by\":7,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":471756,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"floatimage4.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8314049/v1/50392d4054091d37802b8aed.png\"},{\"id\":98451861,\"identity\":\"e4531945-4842-406f-b5ac-c606fb87c1a5\",\"added_by\":\"auto\",\"created_at\":\"2025-12-17 17:33:51\",\"extension\":\"png\",\"order_by\":8,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":50216,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"Onlinefloatimage1.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8314049/v1/b4d153d58ee29f8936c85e1a.png\"},{\"id\":98451874,\"identity\":\"44ba9a1c-7c86-4983-bec0-efc4528ac0a1\",\"added_by\":\"auto\",\"created_at\":\"2025-12-17 17:33:57\",\"extension\":\"png\",\"order_by\":9,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":32150,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"Onlinefloatimage2.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8314049/v1/074a33c40036b16aadfefde7.png\"},{\"id\":98451929,\"identity\":\"f23a81d0-d666-4fc9-9f68-5822c22af151\",\"added_by\":\"auto\",\"created_at\":\"2025-12-17 17:34:15\",\"extension\":\"png\",\"order_by\":10,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":124628,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"Onlinefloatimage3.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8314049/v1/9b94745017a6825621128328.png\"},{\"id\":98451839,\"identity\":\"2b67f12f-a7f7-4910-b44a-7dd7c6a25a22\",\"added_by\":\"auto\",\"created_at\":\"2025-12-17 17:33:41\",\"extension\":\"png\",\"order_by\":11,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":91510,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"Onlinefloatimage4.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8314049/v1/6f988e3e40cab7e84caff43a.png\"},{\"id\":98451862,\"identity\":\"5ef0e904-e229-4cec-922b-6338c70ac027\",\"added_by\":\"auto\",\"created_at\":\"2025-12-17 17:33:52\",\"extension\":\"xml\",\"order_by\":12,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":98599,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"ad11a71c1d8f4f6e9ea19ae9038b5f901structuring.xml\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8314049/v1/d45866387c07e23b422df951.xml\"},{\"id\":98451821,\"identity\":\"9ea6a741-b19b-40aa-8838-4c4f1a9ea38b\",\"added_by\":\"auto\",\"created_at\":\"2025-12-17 17:33:36\",\"extension\":\"html\",\"order_by\":13,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"acdc-reference\",\"size\":111347,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"earlyproof.html\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8314049/v1/c5c9ccd5d5fd024dcfb78a0d.html\"},{\"id\":98451851,\"identity\":\"ad435aab-4276-4de8-85bf-5c0cffe2ac56\",\"added_by\":\"auto\",\"created_at\":\"2025-12-17 17:33:46\",\"extension\":\"jpg\",\"order_by\":1,\"title\":\"Figure 1\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":115135,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003e\\u003cstrong\\u003eBiases in the ECMWF precipitation forecasts over the Yangtze River Delta (YRD). \\u003c/strong\\u003e(a) Geographic and topographic features of the YRD, with the red solid box indicating the study domain. (b) Grid point groups within the study domain after rainy-day sampling (see Methods) at a 36-hour lead time initialized at 00 UTC from ECMWF. Blue, orange, and green denote the proportions of clear-sky hit, false alarm, and rainy hit and miss grid points, respectively. (c) Count and cumulative distribution function (CDF) of 3-hourly precipitation at grid points belonging to the rainy hit and miss group. The CMPA precipitation and the ECMWF forecasts are shown in coral and sky-blue bars, respectively. (d) Cyan bar shows the number of grid points belonging to the rainy hit and miss group in each bin of 3-hourly precipitation bias (ECMWF minus CMPA), and blue bar highlights those with daily rainfall exceeding 50 mm.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"1.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8314049/v1/37c09f8f6520f25600b324f8.jpg\"},{\"id\":98451755,\"identity\":\"a131aff6-408e-4305-b5ce-c6f45f6e221b\",\"added_by\":\"auto\",\"created_at\":\"2025-12-17 17:33:16\",\"extension\":\"jpg\",\"order_by\":2,\"title\":\"Figure 2\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":104857,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003e\\u003cstrong\\u003eThe Multi-task U-Net model architecture.\\u003c/strong\\u003e UnetDif (red arrow) and UnetOri (cyan arrow) share the same thirty-nine channels and an identical U-Net architecture, while they use the ECWMF precipitation biases and CMPA precipitation as the ground truth, respectively. Using the YRD region as an example, each colored cube represents a feature map with their width, height and the number of channels attached. The blue arrow denotes a convolution block consisting of a convolution, batch normalization, and a ReLU activation function. The orange, green, and dashed blue arrows denote the max pooling, up convolution, and skip connection operations, respectively. UnetDif outputs four branches, two for positive and negative biases and two for clear-sky and false-alarm masks. The predicted biases are removed from the raw ECMWF forecasts, and the masks suppress spurious rainfall. UnetOri outputs three branches, one for precipitation and two masks, and applies the same mask-based filtering.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"2.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8314049/v1/229a6482f706c42f8f3be6b5.jpg\"},{\"id\":98451910,\"identity\":\"e9c44c3a-5ada-44be-880c-81172ea836ff\",\"added_by\":\"auto\",\"created_at\":\"2025-12-17 17:34:07\",\"extension\":\"jpg\",\"order_by\":3,\"title\":\"Figure 3\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":229211,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003e\\u003cstrong\\u003ePrediction skill of heavy rainfall over YRD at a 36-hour lead time. \\u003c/strong\\u003e(a) Forecast skills of heavy rainfall over the YRD for ECMWF, UnetOri, and UnetDif from 1 January to 31 October 2025. TS, ETS, and Bias are represented by blue, red, and hatched bars, respectively. (b) TS for the ECMWF forecasts at each grid point. (c) Differences in TS between UnetOri and ECMWF at each grid point. (d) Differences in TS between UnetDif and ECMWF at each grid point. (e)-(g) Same as (b)-(d), but for Bias.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"3.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8314049/v1/8f5f144aaf1df240706054c7.jpg\"},{\"id\":98451887,\"identity\":\"e67c8b11-d3e3-4796-b699-908be795ffb2\",\"added_by\":\"auto\",\"created_at\":\"2025-12-17 17:34:01\",\"extension\":\"jpg\",\"order_by\":4,\"title\":\"Figure 4\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":147497,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003e\\u003cstrong\\u003ePrediction skill of heavy rainfall across four regions of China at a 36-hour lead time. \\u003c/strong\\u003e(a) Topography over China, with North, East, South, and Southwest China outlined by purple, red, blue, and green boxes, respectively. (b) Prediction skills of heavy rainfall in North China from 1 January 2024 to 31 October 2025 for ECMWF, UnetOri, and UnetDif. TS, ETS, FAR, and Bias are shown in blue, red, cyan, and hatched bars, respectively. (c)-(e) Same as (b), but for East, South, and Southwest China, respectively.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"4.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8314049/v1/7364235754200227e5f7d7ce.jpg\"},{\"id\":104251954,\"identity\":\"6a6bbd41-027f-4562-a8fd-6bda52915c33\",\"added_by\":\"auto\",\"created_at\":\"2026-03-09 16:16:20\",\"extension\":\"pdf\",\"order_by\":0,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"manuscript-pdf\",\"size\":1315757,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"manuscript.pdf\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8314049/v1/14aa79cc-59b1-4dd9-99b1-f812c3ec8b75.pdf\"},{\"id\":98451834,\"identity\":\"0cb606a9-71e0-4be4-b487-4e3895d42253\",\"added_by\":\"auto\",\"created_at\":\"2025-12-17 17:33:40\",\"extension\":\"docx\",\"order_by\":0,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"supplement\",\"size\":2047257,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"BiastargeteddeeplearningenhancesshortrangeheavyrainfallforecastsSupplementaryinformation.docx\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8314049/v1/52af87c642c3ffb1af911a3f.docx\"}],\"financialInterests\":\"No competing interests reported.\",\"formattedTitle\":\"Bias-targeted deep learning enhances short-range heavy rainfall forecasts\",\"fulltext\":[{\"header\":\"Introduction\",\"content\":\"\\u003cp\\u003eHeavy rainfall can trigger flooding, landslides, debris flows, and urban waterlogging, posing a substantial threat to socioeconomic development and to the safety of human lives and properties\\u003csup\\u003e\\u003cspan citationid=\\\"CR1\\\" class=\\\"CitationRef\\\"\\u003e1\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR2\\\" class=\\\"CitationRef\\\"\\u003e2\\u003c/span\\u003e\\u003c/sup\\u003e. With rising greenhouse gas concentrations, the frequency of heavy rainfall events has increased\\u003csup\\u003e\\u003cspan additionalcitationids=\\\"CR4\\\" citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR5\\\" class=\\\"CitationRef\\\"\\u003e5\\u003c/span\\u003e\\u003c/sup\\u003e, and such events are likely to become even more frequent in the future\\u003csup\\u003e\\u003cspan citationid=\\\"CR6\\\" class=\\\"CitationRef\\\"\\u003e6\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR7\\\" class=\\\"CitationRef\\\"\\u003e7\\u003c/span\\u003e\\u003c/sup\\u003e. Therefore, enhancing the forecasting and early-warning capability for heavy rainfall events are crucial for effective disaster preparedness and mitigation.\\u003c/p\\u003e \\u003cp\\u003eReal-time observations, such as radar echoes and satellite imagery, have substantially improved heavy rainfall prediction at nowcasting lead times\\u003csup\\u003e\\u003cspan additionalcitationids=\\\"CR9\\\" citationid=\\\"CR8\\\" class=\\\"CitationRef\\\"\\u003e8\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR10\\\" class=\\\"CitationRef\\\"\\u003e10\\u003c/span\\u003e\\u003c/sup\\u003e, yet short-range forecasts from the NWP models, which solve the nonlinear, non-closed fluid dynamical equations, are better suited to meet the temporal demands of heavy rainfall disaster preparedness. Over recent decades, NWPs forecast skills have increased significantly\\u003csup\\u003e\\u003cspan additionalcitationids=\\\"CR12 CR13\\\" citationid=\\\"CR11\\\" class=\\\"CitationRef\\\"\\u003e11\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR14\\\" class=\\\"CitationRef\\\"\\u003e14\\u003c/span\\u003e\\u003c/sup\\u003e, however, coarse spatial resolution requires sub-grid parameterizations for many small-scale precipitation processes, inevitably introducing systematic biases\\u003csup\\u003e\\u003cspan additionalcitationids=\\\"CR16\\\" citationid=\\\"CR15\\\" class=\\\"CitationRef\\\"\\u003e15\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR17\\\" class=\\\"CitationRef\\\"\\u003e17\\u003c/span\\u003e\\u003c/sup\\u003e. Post-processing of NWP outputs is recognized as one of the most effective methods to reduce these biases and further enhance forecast skills\\u003csup\\u003e\\u003cspan additionalcitationids=\\\"CR19 CR20 CR21 CR22 CR23 CR24 CR25 CR26 CR27 CR28 CR29 CR30 CR31 CR32 CR33 CR34\\\" citationid=\\\"CR18\\\" class=\\\"CitationRef\\\"\\u003e18\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR35\\\" class=\\\"CitationRef\\\"\\u003e35\\u003c/span\\u003e\\u003c/sup\\u003e.\\u003c/p\\u003e \\u003cp\\u003eTraditional statistical methods correct biases in the NWP outputs based on the long-term linear statistical relationship between model forecasts and observations\\u003csup\\u003e\\u003cspan additionalcitationids=\\\"CR19 CR20 CR21\\\" citationid=\\\"CR18\\\" class=\\\"CitationRef\\\"\\u003e18\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR22\\\" class=\\\"CitationRef\\\"\\u003e22\\u003c/span\\u003e\\u003c/sup\\u003e. A key underlying assumption of these methods is that the precipitation biases in NWP are stationary and independent of weather systems, which prevents them from effectively capturing the bias characteristics associated with different weather systems\\u003csup\\u003e\\u003cspan citationid=\\\"CR27\\\" class=\\\"CitationRef\\\"\\u003e27\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR31\\\" class=\\\"CitationRef\\\"\\u003e31\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR35\\\" class=\\\"CitationRef\\\"\\u003e35\\u003c/span\\u003e\\u003c/sup\\u003e, resulting in the limited post-processing precision and suboptimal performance in heavy rainfall forecasts\\u003csup\\u003e\\u003cspan citationid=\\\"CR27\\\" class=\\\"CitationRef\\\"\\u003e27\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR31\\\" class=\\\"CitationRef\\\"\\u003e31\\u003c/span\\u003e\\u003c/sup\\u003e.\\u003c/p\\u003e \\u003cp\\u003eIn recent years, AI algorithms have been widely applied to the post-processing of NWP due to their ability to handle non-linear problems and capture multi-variable features, often outperforming traditional statistical methods\\u003csup\\u003e\\u003cspan additionalcitationids=\\\"CR24 CR25 CR26 CR27 CR28 CR29 CR30 CR31 CR32 CR33 CR34\\\" citationid=\\\"CR23\\\" class=\\\"CitationRef\\\"\\u003e23\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR35\\\" class=\\\"CitationRef\\\"\\u003e35\\u003c/span\\u003e\\u003c/sup\\u003e. Nevertheless, the strictly non-negative and heavy-tailed nature of precipitation still limit the effectiveness of deep learning post-processing methods in heavy rainfall forecasts\\u003csup\\u003e\\u003cspan citationid=\\\"CR27\\\" class=\\\"CitationRef\\\"\\u003e27\\u003c/span\\u003e,\\u003cspan additionalcitationids=\\\"CR30\\\" citationid=\\\"CR29\\\" class=\\\"CitationRef\\\"\\u003e29\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR31\\\" class=\\\"CitationRef\\\"\\u003e31\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR36\\\" class=\\\"CitationRef\\\"\\u003e36\\u003c/span\\u003e\\u003c/sup\\u003e.\\u003c/p\\u003e \\u003cp\\u003ePrevious studies have made valuable progress by modifying loss functions\\u003csup\\u003e\\u003cspan additionalcitationids=\\\"CR27 CR28\\\" citationid=\\\"CR26\\\" class=\\\"CitationRef\\\"\\u003e26\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR29\\\" class=\\\"CitationRef\\\"\\u003e29\\u003c/span\\u003e\\u003c/sup\\u003e or by designing increasingly intricate hybrid models\\u003csup\\u003e\\u003cspan additionalcitationids=\\\"CR31 CR32\\\" citationid=\\\"CR30\\\" class=\\\"CitationRef\\\"\\u003e30\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR33\\\" class=\\\"CitationRef\\\"\\u003e33\\u003c/span\\u003e\\u003c/sup\\u003e. However, these methods often struggle to balance practical simplicity with transferability. Specifically, methods based on loss-function modification typically require fine-tuning when applied to a new study region, whereas hybrid models generally entail a rapidly increasing parameter demand.\\u003c/p\\u003e \\u003cp\\u003eNotably, most of these studies rely on observed precipitation as the ground truth\\u003csup\\u003e\\u003cspan additionalcitationids=\\\"CR24 CR25 CR26 CR27 CR28 CR29 CR30 CR31 CR32 CR33 CR34\\\" citationid=\\\"CR23\\\" class=\\\"CitationRef\\\"\\u003e23\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR35\\\" class=\\\"CitationRef\\\"\\u003e35\\u003c/span\\u003e\\u003c/sup\\u003e, and few have exploited precipitation biases themselves\\u003csup\\u003e\\u003cspan citationid=\\\"CR36\\\" class=\\\"CitationRef\\\"\\u003e36\\u003c/span\\u003e\\u003c/sup\\u003e. Here, we train models using both precipitation biases and observed precipitation as the targets to correct the short-range rainfall forecasts derived from European Centre for Medium-Range Weather Forecast Integrated Forecasting System (ECMWF). By comparing the results with each other, this study highlights that using precipitation biases as the target more effectively enhances short-range heavy rainfall forecasts.\\u003c/p\\u003e\"},{\"header\":\"Results\",\"content\":\"\\u003cdiv id=\\\"Sec3\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eBiases of ECMWF Precipitation Forecasts over the Yangtze River Delta (YRD)\\u003c/h2\\u003e \\u003cp\\u003eThe YRD (116-123\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:^\\\\circ\\\\:\\\\)\\u003c/span\\u003e\\u003c/span\\u003eE, 27-34\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:^\\\\circ\\\\:\\\\)\\u003c/span\\u003e\\u003c/span\\u003eN) is one of the most densely populated and economically vital regions in China, yet its complex terrain (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig1\\\" class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003ea) and exposure to multiple weather systems, such as Mei-yu front and tropical cyclones\\u003csup\\u003e\\u003cspan citationid=\\\"CR37\\\" class=\\\"CitationRef\\\"\\u003e37\\u003c/span\\u003e\\u003c/sup\\u003e, render it highly vulnerable to heavy rainfall\\u003csup\\u003e\\u003cspan citationid=\\\"CR2\\\" class=\\\"CitationRef\\\"\\u003e2\\u003c/span\\u003e\\u003c/sup\\u003e. Enhancing short-range heavy rainfall forecast skill in the YRD is therefore of critical importance.\\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e \\u003cp\\u003eA rainy-day sampling method (see Methods) is applied to the China Meteorological Administration Multisource Precipitation Analysis\\u003csup\\u003e\\u003cspan citationid=\\\"CR38\\\" class=\\\"CitationRef\\\"\\u003e38\\u003c/span\\u003e\\u003c/sup\\u003e (CMPA) dataset to reduce the dominance of clear-sky cases over the YRD (Supplementary Table\\u0026nbsp;1). The corresponding 3-hourly ECMWF precipitation forecasts initialized at 00 UTC at 15\\u0026ndash;36 hours lead times are then evaluated against the CMPA observations. A pointwise comparison reveals that 44.5% of grids are clear-sky hits, 19.2% are false alarms, and the remaining 36.2% correspond to rainy conditions in the CMPA observations, comprising both ECMWF hits and misses (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig1\\\" class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003eb, c).\\u003c/p\\u003e \\u003cp\\u003eFor rainy condition hits, both the CMPA and the ECMWF forecasts exhibit long-tailed distributions, with decreasing frequencies toward higher rainfall intensities. In general, the 3-hourly precipitation over the YRD falls within 0\\u0026ndash;50 mm (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig1\\\" class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003ec). ECMWF underestimates the grid numbers with 3-hourly precipitation both less than 3 mm and larger than 15 mm, indicating systematic misses at both ends of the intensity spectrum. Likewise, both the positive and negative precipitation biases of ECMWF display the long-tailed distributions (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig1\\\" class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003ed). Since the negative biases occurring far more frequently than positive ones, the ECMWF precipitation forecasts tend to underestimate rainfall intensity.\\u003c/p\\u003e \\u003cp\\u003eMore importantly, the approximate Gaussian distribution of the 3-hourly precipitation biases at grid points with daily rainfall exceeding 50 mm, provides a strong rationale for adopting a bias-targeted modeling strategy to enhance heavy rainfall forecasts (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig1\\\" class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003ed). Results based on the ECMWF precipitation forecasts initialized at 12 UTC exhibit similar characteristics (Supplementary Fig.\\u0026nbsp;1).\\u003c/p\\u003e \\u003c/div\\u003e\\n\\u003ch3\\u003eBias-targeted deep learning improves heavy rainfall forecasts over the YRD\\u003c/h3\\u003e\\n\\u003cp\\u003eAccording to the characteristics of biases in the ECMWF precipitation forecasts over the YRD, grids in all samples can be classified into four types: clear-sky hits, false alarms, rainy condition hits with under-predicted intensity (including misses), and rainy condition hits with over-predicted intensity. Correspondingly, this study develops a multi-task\\u003csup\\u003e\\u003cspan citationid=\\\"CR39\\\" class=\\\"CitationRef\\\"\\u003e39\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR40\\\" class=\\\"CitationRef\\\"\\u003e40\\u003c/span\\u003e\\u003c/sup\\u003e U-Net\\u003csup\\u003e\\u003cspan citationid=\\\"CR41\\\" class=\\\"CitationRef\\\"\\u003e41\\u003c/span\\u003e\\u003c/sup\\u003e with four output branches (see Methods, Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig2\\\" class=\\\"InternalRef\\\"\\u003e2\\u003c/span\\u003e) that uses thirty-eight ECMWF variables and topography dataset as the inputs (Supplementary Table\\u0026nbsp;2) and the ECMWF precipitation biases as the target (i.e., ground truth). The model predicts the precipitation bias at each grid point, and these predicted biases are subsequently subtracted from the raw ECMWF forecasts to yield the corrected precipitation forecasts (UnetDif). For comparison, a three output branches U-Net is trained using the CMPA observations as the ground truth to predict precipitation directly (UnetOri), a configuration aligned with previous studies\\u003csup\\u003e27\\u0026ndash;29,31\\u0026minus;35\\u003c/sup\\u003e.\\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e \\u003cp\\u003eIn order to match the time steps used in the operational short-range predictions, bias corrections are applied to the 3-hourly ECMWF precipitation forecasts, while evaluated on the 24-hour accumulated precipitation over the testing period (Supplementary Table\\u0026nbsp;1). In addition, due to the \\u0026ldquo;double penalty\\u0026rdquo; problem inherent in root mean square error\\u003csup\\u003e\\u003cspan citationid=\\\"CR28\\\" class=\\\"CitationRef\\\"\\u003e28\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR35\\\" class=\\\"CitationRef\\\"\\u003e35\\u003c/span\\u003e\\u003c/sup\\u003e, threat score\\u003csup\\u003e\\u003cspan citationid=\\\"CR42\\\" class=\\\"CitationRef\\\"\\u003e42\\u003c/span\\u003e\\u003c/sup\\u003e (TS, see Methods), equitable threat score\\u003csup\\u003e\\u003cspan citationid=\\\"CR43\\\" class=\\\"CitationRef\\\"\\u003e43\\u003c/span\\u003e\\u003c/sup\\u003e (ETS, see Methods), bias score\\u003csup\\u003e\\u003cspan citationid=\\\"CR44\\\" class=\\\"CitationRef\\\"\\u003e44\\u003c/span\\u003e\\u003c/sup\\u003e (Bias, see Methods) and false alarm ratio\\u003csup\\u003e\\u003cspan citationid=\\\"CR44\\\" class=\\\"CitationRef\\\"\\u003e44\\u003c/span\\u003e\\u003c/sup\\u003e (FAR, see Methods) are used to evaluate the precipitation forecast skills. At each lead time, the ECMWF precipitation forecasts initialized at 00 and 12 UTC are merged and evaluated collectively.\\u003c/p\\u003e \\u003cp\\u003eAt a 36-hour lead time, UnetOri and UnetDif improve the clear-sky and rainy-day hit ratios (ACC; see Methods) to 0.820 and 0.817, respectively, compared with 0.774 for the raw ECMWF forecasts, primarily by reducing false alarms (Supplementary Fig.\\u0026nbsp;2a, c; Supplementary Table\\u0026nbsp;3). They also enhance the TS by ~\\u0026thinsp;6% and ~\\u0026thinsp;4% for daily rainfall above 10 mm, and by ~\\u0026thinsp;8% and ~\\u0026thinsp;6% for daily rainfall above 25 mm, respectively.\\u003c/p\\u003e \\u003cp\\u003eThe two deep-learning-based post-processing methods achieve these improvements through different mechanisms. Specifically, UnetOri exhibits lower Bias values (1.039 and 0.730 for daily rainfall above 10 mm and 25 mm), indicating fewer false alarms relative to the raw ECMWF forecasts. By contrast, UnetDif shows higher Bias values, suggesting fewer misses but more false alarms than the raw forecasts (Supplementary Fig.\\u0026nbsp;2a, c; Supplementary Table\\u0026nbsp;3).\\u003c/p\\u003e \\u003cp\\u003eAcross the YRD, the TS for heavy rainfall (daily rainfall above 50 mm) drops markedly from 0.194 in ECMWF to 0.131 (~\\u0026thinsp;32%) in UnetOri (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig3\\\" class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003ea). The reduced TS and lower Bias at most grid points (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig3\\\" class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003eb, c, e, f) indicate that UnetOri generates more misses in heavy rainfall events (Supplementary Fig.\\u0026nbsp;2b, d; Supplementary Table\\u0026nbsp;3). In contrast, UnetDif increases the regional TS by ~\\u0026thinsp;22% relative to ECMWF, reaching 0.237 (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig3\\\" class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003ea). TS improves at most grid points, while Bias increases where it was initially lower than 1 and decreases or remains unchanged where it was higher than 1 (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig3\\\" class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003ed, g). Furthermore, UnetDif yields a probability distribution of heavy rainfall grids that more closely matches the CMPA observations, particularly for daily rainfall above 75 mm (Supplementary Fig.\\u0026nbsp;2b, d), demonstrating that UnetDif provides more accurate heavy rainfall forecasts than both ECMWF and UnetOri.\\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e \\u003cp\\u003eThe advantages of UnetDif for short-range heavy rainfall forecasts can be further illustrated by comparing TS of the top 20 events in 2025 over the YRD (Supplementary Table\\u0026nbsp;4). At a 36-hour lead time, quantile mapping\\u003csup\\u003e\\u003cspan citationid=\\\"CR18\\\" class=\\\"CitationRef\\\"\\u003e18\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR19\\\" class=\\\"CitationRef\\\"\\u003e19\\u003c/span\\u003e\\u003c/sup\\u003e (QM, see Methods), UnetOri, and UnetDif exceed ECMWF in 13, 4, and 15 cases, respectively. Spatial distributions of the precipitation predictions suggest that UnetDif effectively compensates for the heavy rainfall misses in the ECMWF forecasts, while UnetOri and QM do not (Supplementary Fig.\\u0026nbsp;3).\\u003c/p\\u003e \\u003cp\\u003eFurthermore, at 48- and 60-hour lead times, UnetDif enhances heavy rainfall TS relative to ECMWF by ~\\u0026thinsp;21% and ~\\u0026thinsp;30%, respectively. QM improves TS by ~\\u0026thinsp;1.1% and ~\\u0026thinsp;8%, whereas UnetOri decreases TS by ~\\u0026thinsp;14% and ~\\u0026thinsp;23%. Results evaluated by ETS are similar (Supplementary Table\\u0026nbsp;3). It is worth noting that the best Bias for QM, accompanied by its nearly unchanged TS and ETS relative to ECMWF, may result from its highest FAR compared with ECMWF, UnetOri and UnetDif (Supplementary Table\\u0026nbsp;3).\\u003c/p\\u003e \\u003cp\\u003e \\u003cb\\u003eBias-targeted deep learning improves heavy rainfall forecasts across different regions of China.\\u003c/b\\u003e \\u003c/p\\u003e \\u003cp\\u003eKeeping the computational procedure unchanged, the bias-targeted deep learning methods are further applied to four major precipitation regions of China (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig4\\\" class=\\\"InternalRef\\\"\\u003e4\\u003c/span\\u003ea). Specifically, North China (108-112\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:^\\\\circ\\\\:\\\\)\\u003c/span\\u003e\\u003c/span\\u003eE, 32-45\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:^\\\\circ\\\\:\\\\)\\u003c/span\\u003e\\u003c/span\\u003eN), East China (112-123\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:^\\\\circ\\\\:\\\\)\\u003c/span\\u003e\\u003c/span\\u003eE, 22-36\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:^\\\\circ\\\\:\\\\)\\u003c/span\\u003e\\u003c/span\\u003eN), South China (104-118\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:^\\\\circ\\\\:\\\\)\\u003c/span\\u003e\\u003c/span\\u003eE, 18-28\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:^\\\\circ\\\\:\\\\)\\u003c/span\\u003e\\u003c/span\\u003eN) and Southwest China (89-112\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:^\\\\circ\\\\:\\\\)\\u003c/span\\u003e\\u003c/span\\u003eE, 20-35\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:^\\\\circ\\\\:\\\\)\\u003c/span\\u003e\\u003c/span\\u003eN). Furthermore, rainy-day sampling period is shortened to 2019\\u0026ndash;2023, with 2024\\u0026ndash;2025 as the testing period, to strengthen the robustness of the hindcast evaluation (Supplementary Table\\u0026nbsp;1).\\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e \\u003cp\\u003eConsistent with the results over the YRD, UnetDif improves heavy rainfall forecasts in all four regions relative to ECMWF. Specifically, TS (ETS) in UnetDif increases by ~\\u0026thinsp;4%, ~\\u0026thinsp;23%, ~\\u0026thinsp;26%, and ~\\u0026thinsp;32% in North, East, South, and Southwest China, respectively. In contrast, while UnetOri increases TS (ETS) by ~\\u0026thinsp;6.4% in Southwest China, it decreases TS (ETS) by approximately 33%, 1.4%, and 0.07% in North, East and South China, respectively (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig4\\\" class=\\\"InternalRef\\\"\\u003e4\\u003c/span\\u003eb-e).\\u003c/p\\u003e \\u003cp\\u003eA further in-depth analysis suggests that Bias increases from 0.838 to 1.314 in East China and from 0.711 to 1.027 in South China, while FAR remains nearly unchanged, indicating that UnetDif effectively reduces ECMWF misses and thereby enhances heavy rainfall forecasts (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig4\\\" class=\\\"InternalRef\\\"\\u003e4\\u003c/span\\u003ec, d). In Southwest China, reductions in both Bias and FAR suggest that UnetDif enhances TS primarily by decreasing false alarms (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig4\\\" class=\\\"InternalRef\\\"\\u003e4\\u003c/span\\u003ee). In North China, TS, ETS, Bias, and FAR all remain nearly unchanged (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig4\\\" class=\\\"InternalRef\\\"\\u003e4\\u003c/span\\u003eb), this poor performance may be due to the insufficient training samples\\u003csup\\u003e\\u003cspan citationid=\\\"CR45\\\" class=\\\"CitationRef\\\"\\u003e45\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR46\\\" class=\\\"CitationRef\\\"\\u003e46\\u003c/span\\u003e\\u003c/sup\\u003e (Supplementary Table\\u0026nbsp;1). The evaluation results remain similar when using 2024 or 2025 alone as the testing period (Supplementary Fig.\\u0026nbsp;4).\\u003c/p\\u003e \\u003cp\\u003eAmong the top five heavy rainfall events in each of the four regions at a 36-hour lead time, QM, UnetOri, and UnetDif exceed the raw ECMWF forecasts in 14, 12, and 16 cases, respectively (Supplementary Table\\u0026nbsp;5). For the event in each region with the largest UnetDif improvement relative to the ECMWF forecasts, the spatial distributions of precipitation forecasts indicate that UnetDif effectively compensates for ECMWF misses, particularly along the narrow heavy rainfall bands in coastal area in East and South China, whereas QM and UnetOri fail to do so (Supplementary Fig.\\u0026nbsp;5f-o). Once again, the barely changed TS but improved Bias of QM suggest that it generates more false alarms rather than hits of heavy rainfall (Supplementary Fig.\\u0026nbsp;5c, h, m, r).\\u003c/p\\u003e \\u003cp\\u003eHowever, predicting daily rainfall exceeding 100 mm remains highly challenging. Both UnetDif and UnetOri exhibit substantial misses for these extreme precipitation grid points (Supplementary Fig.\\u0026nbsp;4i, j, n, s, t), whereas QM generates numerous false alarms, even above 250 mm per day.\\u003c/p\\u003e\"},{\"header\":\"Discussion\",\"content\":\"\\u003cp\\u003eTo the best of our knowledge, this study is the first to demonstrate that training deep learning models using NWP precipitation biases as target can substantially improve short-range heavy rainfall prediction skill. A crucial starting point arises from a quiet but telling clue in the data, namely that both the positive and negative 3-hourly biases associated with daily rainfall above 50 mm show a clear Gaussian distribution. This statistical characteristic is not merely a curiosity, but also the fundamental reason why this bias-targeted deep learning method can so effectively enhance heavy rainfall forecasts.\\u003c/p\\u003e \\u003cp\\u003eIn order to improve the short-range heavy rainfall forecasts, previous studies have made valuable progress by modifying loss functions or by designing increasingly intricate hybrid models. Yet these methods often struggle to balance practical simplicity with transferability. In contrast, the method we proposed in this study remains simple, since the only preprocessing step is converting observed precipitation into NWP forecast biases. It is also practically robust, as it delivers strong performance across multiple regions of China. This combination of elegance and utility, we believe, offers a useful blueprint for future studies.\\u003c/p\\u003e \\u003cp\\u003eIt is worth noting that existing methods such as the loss function modifications or the hybrid deep learning models are not incompatible with our framework. On the contrary, they can be integrated with bias-targeted deep learning and may yield further improvements in short-range heavy rainfall forecasts.\\u003c/p\\u003e \\u003cp\\u003eEven so, several limitations remain. UnetOri slightly outperforms UnetDif for low-threshold rainfall, highlighting the potential of combining their strengths for short-range forecasts across varying rainfall intensities; Both deep-learning-based methods, however, still struggle with daily rainfall exceeding 100 mm, likely due to the extreme scarcity of such events. This points to an unavoidable but important conclusion: the longer, high-quality observational and NWP forecast datasets are essential for further advances in deep-learning-based bias correction methods\\u003csup\\u003e\\u003cspan citationid=\\\"CR45\\\" class=\\\"CitationRef\\\"\\u003e45\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR46\\\" class=\\\"CitationRef\\\"\\u003e46\\u003c/span\\u003e\\u003c/sup\\u003e.\\u003c/p\\u003e\"},{\"header\":\"Methods\",\"content\":\"\\u003cdiv id=\\\"Sec7\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eDatasets\\u003c/h2\\u003e \\u003cp\\u003eThe 3-hourly ECMWF forecast datasets from January 2019 to October 2025 are used in this study. ECMWF forecasts issue twice a day at 00 UTC and 12 UTC, and the forecast lead times are 0-240 h. For short-range forecasts, lead times of 15\\u0026ndash;60 h are used in this study. More specifically, thirty-eight meteorological variables provided by ECMWF are used here (Supplementary Table\\u0026nbsp;2), including multiple pressure level fields of relative humidity, specific humidity, vertical velocity, temperature, zonal and meridional wind components, potential vorticity, and geopotential height, all with a spatial resolution of 0.25\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:^\\\\circ\\\\:\\\\)\\u003c/span\\u003e\\u003c/span\\u003e \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:\\\\times\\\\:\\\\)\\u003c/span\\u003e\\u003c/span\\u003e 0.25\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:^\\\\circ\\\\:\\\\)\\u003c/span\\u003e\\u003c/span\\u003e. Moreover, single-level variables, including total column water vapor, total cloud cover, low cloud cover, large-scale precipitation, convective precipitation, and precipitation at 0.125\\u0026deg; \\u0026times; 0.125\\u0026deg; resolution are also utilized.\\u003c/p\\u003e \\u003cp\\u003eThe observed precipitation dataset used in this study is the high-quality China Meteorological Administration Multisource Precipitation Analysis System (CMPA) dataset\\u003csup\\u003e\\u003cspan citationid=\\\"CR38\\\" class=\\\"CitationRef\\\"\\u003e38\\u003c/span\\u003e\\u003c/sup\\u003e, which integrates three types of precipitation observations (i.e., gauge, satellite and radar) with spatial resolution of 0.05\\u0026deg; \\u0026times; 0.05\\u0026deg; and temporal resolution of 1 h. The 3-hourly observed precipitation is calculated by summing three 1-hourly records. The topography dataset with a resolution of 0.05\\u0026deg; \\u0026times; 0.05\\u0026deg; is downloaded from the National Earth System Science Data Center. All datasets have been interpolated to a resolution of 0.125\\u0026deg; \\u0026times; 0.125\\u0026deg; using the bilinear interpolation.\\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec8\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eRainy day sampling method\\u003c/h2\\u003e \\u003cp\\u003eA day during the sampling period is recognized as a rainy day if at least 10% of grid points have precipitation (\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:\\\\ge\\\\:\\\\)\\u003c/span\\u003e\\u003c/span\\u003e 0.1 mm) in all 3-hourly intervals of that day in CMPA. The corresponding ECMWF forecasts initialized at 00 and 12 UTC with lead times of 36, 48, and 60 hours are then retrieved separately. Each identified rainy day contains eight 3-hourly samples. Among all samples, 85% are used for model training and the remaining 15% for validation.\\u003c/p\\u003e \\u003c/div\\u003e\\n\\u003ch3\\u003eMulti-task U-Net model\\u003c/h3\\u003e\\n\\u003cp\\u003eU-Net\\u003csup\\u003e\\u003cspan citationid=\\\"CR41\\\" class=\\\"CitationRef\\\"\\u003e41\\u003c/span\\u003e\\u003c/sup\\u003e is a classical image segmentation network composed of down-sampling, up-sampling, and skip connections, and it has been widely applied in precipitation bias correction studies\\u003csup\\u003e\\u003cspan citationid=\\\"CR27\\\" class=\\\"CitationRef\\\"\\u003e27\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR28\\\" class=\\\"CitationRef\\\"\\u003e28\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR30\\\" class=\\\"CitationRef\\\"\\u003e30\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR32\\\" class=\\\"CitationRef\\\"\\u003e32\\u003c/span\\u003e\\u003c/sup\\u003e. In this work, we insert a batch normalization layer between each convolutional layer and its subsequent ReLU activation to facilitate convergence and improve model accuracy\\u003csup\\u003e\\u003cspan citationid=\\\"CR28\\\" class=\\\"CitationRef\\\"\\u003e28\\u003c/span\\u003e\\u003c/sup\\u003e. Furthermore, we extend the original single-task U-Net to multi-task\\u003csup\\u003e\\u003cspan citationid=\\\"CR39\\\" class=\\\"CitationRef\\\"\\u003e39\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR40\\\" class=\\\"CitationRef\\\"\\u003e40\\u003c/span\\u003e\\u003c/sup\\u003e outputs with four and three output branches, respectively (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig2\\\" class=\\\"InternalRef\\\"\\u003e2\\u003c/span\\u003e).\\u003c/p\\u003e \\u003cp\\u003eSpecifically, the four output branches U-Net (UnetDif) comprises two classification branches that generate clear-sky and false-alarm masks, together with two regression branches that predict positive and negative precipitation biases. The predicted biases are first subtracted from the raw ECMWF precipitation forecasts, after which the two masks are applied to suppress spurious rainfall at non-rainy grid points.\\u003c/p\\u003e \\u003cp\\u003eThe three output branches U-Net (UnetOri) retains the two classification branches but uses a single regression branch to directly predict precipitation. The predicted precipitation is subsequently filtered with the clear-sky and false alarm masks to remove spurious rainfall over non-rainy grid points.\\u003c/p\\u003e \\u003cp\\u003eThirty-eight ECMWF forecast variables and topography are used as inputs for both UnetDif and UnetOri, arranged as separate channels. The ground truth for UnetDif is the ECMWF precipitation biases, while for UnetOri, it is the CMPA precipitation. Separate UnetDif and UnetOri models are constructed for each of the three lead times (36, 48, and 60-hour) over the YRD and each of the four regions (Supplementary Table\\u0026nbsp;1). All of the thirty-eight ECMWF input variables are normalized using the Z-score method, while topography is scaled using min-max normalization to remove the influence of different variable magnitudes on network training. In all experiments, the batch size is set to 16 and the learning rate to 1 \\u0026times; 10⁻⁴.\\u003c/p\\u003e\\n\\u003ch3\\u003eLoss function\\u003c/h3\\u003e\\n\\u003cp\\u003eThe total loss function designed for UnetDif consists of six components and is formulated as follows:\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{L}_{total}^{UnetDif}={\\\\stackrel{\\\\sim}{L}}_{focal}^{cs}+{\\\\stackrel{\\\\sim}{L}}_{focal}^{fa}+{\\\\stackrel{\\\\sim}{L}}_{MSE}^{Pos}+{\\\\stackrel{\\\\sim}{L}}_{MSE}^{neg}+{0.1*\\\\stackrel{\\\\sim}{L}}_{mae}^{pos}+{0.1*\\\\stackrel{\\\\sim}{L}}_{mae}^{neg}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003cdiv id=\\\"Equa\\\" class=\\\"Equation\\\"\\u003e\\u003cdiv format=\\\"TEX\\\" class=\\\"mathdisplay\\\" id=\\\"FileID_Equa\\\" name=\\\"EquationSource\\\"\\u003e\\n$$\\\\:{\\\\stackrel{\\\\sim}{L}}_{i}=\\\\frac{{L}_{i}}{{L}_{i}^{\\\\left(1\\\\right)}}$$\\u003c/div\\u003e\\u003c/div\\u003e\\u003c/p\\u003e \\u003cp\\u003eWhere\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{\\\\:\\\\stackrel{\\\\sim}{L}}_{i}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e denotes the \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:i\\\\)\\u003c/span\\u003e\\u003c/span\\u003e-th loss component normalized by its value from the first training epoch, prevents the total loss from dominating by terms with inherently larger magnitude. \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{\\\\stackrel{\\\\sim}{L}}_{focal}^{cs}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e and \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{\\\\stackrel{\\\\sim}{L}}_{focal}^{fa}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e are the normalized focal loss\\u003csup\\u003e\\u003cspan citationid=\\\"CR47\\\" class=\\\"CitationRef\\\"\\u003e47\\u003c/span\\u003e\\u003c/sup\\u003e calculated on clear-sky hit and false alarm grid points, respectively. \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{\\\\stackrel{\\\\sim}{L}}_{MSE}^{pos}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e and \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{\\\\stackrel{\\\\sim}{L}}_{MSE}^{neg}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e represent the normalized mean squared errors (MSE) computed on grid points where precipitation biases are positive and negative, respectively. \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{\\\\stackrel{\\\\sim}{L}}_{MAE}^{pos}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e and \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{\\\\stackrel{\\\\sim}{L}}_{MAE}^{neg}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e denote the mean absolute errors (MAE) evaluated on clear-sky hit grid points, ensuring that the predicted positive and negative biases are set to zero at those non-rainy grid points.\\u003c/p\\u003e \\u003cp\\u003eSince precipitation is strictly non-negative, the negative bias components used in UnetDif, namely, \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{\\\\stackrel{\\\\sim}{L}}_{MSE}^{neg}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e and \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{\\\\stackrel{\\\\sim}{L}}_{MAE}^{neg}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e are excluded from the loss design for UnetOri. Accordingly, the total loss function for UnetOri is formulated as:\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{L}_{total}^{UnetOri}={\\\\stackrel{\\\\sim}{L}}_{focal}^{cs}+{\\\\stackrel{\\\\sim}{L}}_{focal}^{fa}+{\\\\stackrel{\\\\sim}{L}}_{MSE}^{Pos}+{0.1*\\\\stackrel{\\\\sim}{L}}_{mae}^{pos}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/p\\u003e \\u003cp\\u003eHere, \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{\\\\stackrel{\\\\sim}{L}}_{focal}^{cs}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e and \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{\\\\stackrel{\\\\sim}{L}}_{focal}^{fa}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e follow the same definitions as those in the UnetDif. The term \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{\\\\stackrel{\\\\sim}{L}}_{MSE}^{Pos}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e represents the normalized MSE computed on grid points where precipitation is observed. The term \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{\\\\stackrel{\\\\sim}{L}}_{MAE}^{pos}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e denotes the MAE evaluated on clear-sky hit grid points, ensuring that the predicted precipitation is constrained to zero at those non-rainy grid points.\\u003c/p\\u003e \\u003cp\\u003eIt is worth noting that the optimal model parameters of UnetDif and UnetOri are saved at the epoch when the cumulative TS across the 3-hourly precipitation bins (i.e., 0.1, 1, 3, 5, 7, 9 and 11 mm) reach the maximum, rather than when the total losses attain their minimum.\\u003c/p\\u003e \\u003cdiv id=\\\"Sec11\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eEvaluation metrics\\u003c/h2\\u003e \\u003cp\\u003eThe clear-sky and rainy-day hit ratio (ACC), TS, ETS, Bias and false alarm ratio (FAR)\\u003csup\\u003e\\u003cspan citationid=\\\"CR44\\\" class=\\\"CitationRef\\\"\\u003e44\\u003c/span\\u003e\\u003c/sup\\u003e are used to evaluate the performance of the precipitation forecasts in ECMWF and each post-processing method. They are calculated as follows:\\u003cdiv id=\\\"Equb\\\" class=\\\"Equation\\\"\\u003e\\u003cdiv format=\\\"TEX\\\" class=\\\"mathdisplay\\\" id=\\\"FileID_Equb\\\" name=\\\"EquationSource\\\"\\u003e\\n$$\\\\:ACC=\\\\frac{H+Z}{H+M+F+Z}$$\\u003c/div\\u003e\\u003c/div\\u003e\\u003cdiv id=\\\"Equc\\\" class=\\\"Equation\\\"\\u003e\\u003cdiv format=\\\"TEX\\\" class=\\\"mathdisplay\\\" id=\\\"FileID_Equc\\\" name=\\\"EquationSource\\\"\\u003e\\n$$\\\\:TS=\\\\frac{H}{H+M+F}$$\\u003c/div\\u003e\\u003c/div\\u003e\\u003cdiv id=\\\"Equd\\\" class=\\\"Equation\\\"\\u003e\\u003cdiv format=\\\"TEX\\\" class=\\\"mathdisplay\\\" id=\\\"FileID_Equd\\\" name=\\\"EquationSource\\\"\\u003e\\n$$\\\\:BS=\\\\frac{H+F}{H+M}$$\\u003c/div\\u003e\\u003c/div\\u003e\\u003cdiv id=\\\"Eque\\\" class=\\\"Equation\\\"\\u003e\\u003cdiv format=\\\"TEX\\\" class=\\\"mathdisplay\\\" id=\\\"FileID_Eque\\\" name=\\\"EquationSource\\\"\\u003e\\n$$\\\\:FAR=\\\\frac{F}{H+F}$$\\u003c/div\\u003e\\u003c/div\\u003e\\u003cdiv id=\\\"Equf\\\" class=\\\"Equation\\\"\\u003e\\u003cdiv format=\\\"TEX\\\" class=\\\"mathdisplay\\\" id=\\\"FileID_Equf\\\" name=\\\"EquationSource\\\"\\u003e\\n$$\\\\:ETS=\\\\frac{H-{H}_{r}}{H+M+F-{H}_{r}}$$\\u003c/div\\u003e\\u003c/div\\u003e\\u003cdiv id=\\\"Equg\\\" class=\\\"Equation\\\"\\u003e\\u003cdiv format=\\\"TEX\\\" class=\\\"mathdisplay\\\" id=\\\"FileID_Equg\\\" name=\\\"EquationSource\\\"\\u003e\\n$$\\\\:{H}_{r}=\\\\frac{(H+M)(H+F)}{N}$$\\u003c/div\\u003e\\u003c/div\\u003e\\u003c/p\\u003e \\u003cp\\u003eWhere H, M, F, and Z are the numbers of correct forecasts of occurrence (i.e., hits), incorrect forecasts of non-occurrence (i.e., misses), incorrect forecasts of occurrence (i.e., false alarms) and correct forecasts of clear-sky, respectively. \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{H}_{r}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e is the number of hits expected by random chance over N total forecasts.\\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec12\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003eQuantile mapping methods\\u003c/h2\\u003e \\u003cp\\u003eDue to the strong seasonality of precipitation in nature, a localized time-window sampling strategy is employed for quantile mapping\\u003csup\\u003e\\u003cspan citationid=\\\"CR18\\\" class=\\\"CitationRef\\\"\\u003e18\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR19\\\" class=\\\"CitationRef\\\"\\u003e19\\u003c/span\\u003e\\u003c/sup\\u003e (QM) to generate the historical cumulative distribution function (CDF). Specifically, for a target date \\u0026#119863; to be bias-corrected, QM training samples are selected from the 30 days preceding \\u0026#119863; in the same year, the 30 days preceding \\u0026#119863; in the previous year, the 30 days following \\u0026#119863; in the previous year, and the 30 days following \\u0026#119863; in the year before last. The historical CDF is constructed from these samples, and the ECMWF precipitation forecasts are subsequently bias-corrected using QM as follows:\\u003cdiv id=\\\"Equh\\\" class=\\\"Equation\\\"\\u003e\\u003cdiv format=\\\"TEX\\\" class=\\\"mathdisplay\\\" id=\\\"FileID_Equh\\\" name=\\\"EquationSource\\\"\\u003e\\n$$\\\\:{Pr}_{QM}={F}_{CMPA}^{-1}\\\\left({F}_{ECMWF}\\\\right({Pr}_{ECMWF})$$\\u003c/div\\u003e\\u003c/div\\u003e\\u003c/p\\u003e \\u003cp\\u003eWhere \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{F}_{CMPA}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e and \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{F}_{ECMWF}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e are the CDFs of the CMPA and the ECMWF forecasts of the selected training samples, respectively. \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{F}_{CMPA}^{-1}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e denotes the inverse CDF of the CMPA. \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{Pr}_{QM}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e and \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{Pr}_{ECMWF}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e denote the precipitation forecasts after QM correction and the raw ECMWF forecasts, respectively.\\u003c/p\\u003e \\u003c/div\\u003e \"},{\"header\":\"Declarations\",\"content\":\"\\u003cp\\u003e \\u003ch2\\u003eCompeting financial interests\\u003c/h2\\u003e \\u003cp\\u003eThe authors declare no competing financial interests.\\u003c/p\\u003e \\u003c/p\\u003e\\u003ch2\\u003eFunding\\u003c/h2\\u003e \\u003cp\\u003eThis work was supported by the Meteorological Joint Funds of National Natural Science Foundation of China (U2442204) and the Joint Research Project for Meteorological Capacity Improvement (23NLTSZ007). T.T. was supported by Open Fund Project for Heavy Rain (BYKJ2024Q24). W.S. was supported by the Special Project of Innovation and Development of China Meteorological Administration (CXFZ2023J021). H.Q. was supported by Key Innovation Team Project of China Meteorological Administration for Intelligent Forecasting Technology (CMA2022ZD04).\\u003c/p\\u003e\\u003ch2\\u003eAuthor Contribution\\u003c/h2\\u003e\\u003cp\\u003eT.T. and W.S. are co-first authors. T.T. conceived the central idea of the study. T.T. and W.S. designed the AI models and performed the hindcast experiments. T.T. prepared the figures and wrote the paper under supervision of H.Q. and L.L. T.T. W.S. J.F. P.H. H.Q. and L.L. contributed to interpreting results, discussion of associated mechanisms, and improvement of the writing.\\u003c/p\\u003e\\u003cp\\u003e\\u003cstrong\\u003eData availability\\u003c/strong\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eData related to this paper can be downloaded from:\\u003c/p\\u003e\\n\\u003cp\\u003eECMWF 3-hourly forecasts,\\u003c/p\\u003e\\n\\u003cp\\u003ehttps://www.ecmwf.int/en/forecasts/datasets;\\u0026nbsp;\\u003c/p\\u003e\\n\\u003cp\\u003eCMPA observed precipitation datasets,\\u003c/p\\u003e\\n\\u003cp\\u003ehttps://data.cma.cn/data/detail/dataCode/SURF_CMPA_RT_NC/keywords/;\\u003c/p\\u003e\\n\\u003cp\\u003eThe global topography datasets,\\u003c/p\\u003e\\n\\u003cp\\u003ehttps://www.geodata.cn/main/face_science_detail?typeName=face_science\\u0026amp;guid=201519481253546.\\u003c/p\\u003e\\n\\u003cp\\u003e\\u003cstrong\\u003eCode availability\\u003c/strong\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eCodes used in this study are available from the corresponding authors on request.\\u003c/p\\u003e\"},{\"header\":\"References\",\"content\":\"\\u003col\\u003e\\u003cli\\u003e\\u003cspan\\u003eJonkman, S. N. Global perspectives on loss of human life caused by floods. \\u003cem\\u003eNat. Hazards\\u003c/em\\u003e. 34, 151\\u0026ndash;175 (2005).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eGong, L., Zhang, X., Liu, J. \\u0026amp; Gui, H. Exploring the influence of urban agglomeration on extreme precipitation: evidence from the middle reaches of the Yangtze River, China. \\u003cem\\u003eJ. Hydrol. Reg. Stud.\\u003c/em\\u003e 55, 101932 (2024).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eFischer, E. M. \\u0026amp; Knutti, R. Observed heavy precipitation increase confirms theory and early models. \\u003cem\\u003eNat. Clim. Change.\\u003c/em\\u003e 6, 986\\u0026ndash;991 (2016).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eGuerreiro, S. B. et al. Detection of continental-scale intensification of hourly rainfall extremes. \\u003cem\\u003eNat. Clim. Change\\u003c/em\\u003e 8, 803\\u0026ndash;807 (2018).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eHam, Y. G. et al. Anthropogenic fingerprints in daily precipitation revealed by deep learning. \\u003cem\\u003eNature\\u003c/em\\u003e 622, 301\\u0026ndash;307 (2023).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eO'Gorman, P. A. \\u0026amp; Schneider, T. The physical basis for increases in precipitation extremes in simulations of 21st-century climate change. \\u003cem\\u003eProc. Natl. Acad. Sci.\\u003c/em\\u003e 106, 14773\\u0026ndash;14777 (2009).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eZhang, W. et al. Increasing precipitation variability on daily-to-multiyear time scales in a warmer world. \\u003cem\\u003eSci. Adv.\\u003c/em\\u003e 7, eabf8021 (2021).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eZhang, Y. et al. Skillful nowcasting of extreme precipitation with NowcastNet. \\u003cem\\u003eNature\\u003c/em\\u003e 619, 526\\u0026ndash;532 (2023).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eDu, X. \\u0026amp; Guo, H. A multi-scale attention encoding and dynamic decoding network designed for short-term precipitation forecasting. \\u003cem\\u003eEarth Sci. Inform.\\u003c/em\\u003e 18, 41 (2025).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eXiao, S., Zhang, A., Gong, J., Chen, Y., Chen, S. \\u0026amp; Li, W. An explainable deep learning method with squeeze and excitation block for 6-hour forecast of tropical cyclone remote precipitation. \\u003cem\\u003eJ. Geophys. Res.: Mach. Learn. Comput.\\u003c/em\\u003e 2, e2025JH000675 (2025).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eBauer, P., Thorpe, A. \\u0026amp; Brunet, G. The quiet revolution of numerical weather prediction. \\u003cem\\u003eNature\\u003c/em\\u003e 525, 47\\u0026ndash;55 (2015).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eStoelinga, M. T. et al. Improvement of microphysical parameterization through observational verification experiment. \\u003cem\\u003eBull. Amer. Meteor. Soc.\\u003c/em\\u003e 84, 1807\\u0026ndash;1826 (2015).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eThompson, G., Field, P. R., Rasmussen, R. M. \\u0026amp; Hall, W. D. Explicit forecasts of winter precipitation using an improved bulk micro-physics scheme. Part II: Implementation of a new snow parameterization. \\u003cem\\u003eMon. Wea. Rev.\\u003c/em\\u003e 136, 5095\\u0026ndash;5115 (2008).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eLee, J., Min, K. \\u0026amp; Lim, K. Comparing 3DVAR and hybrid radar data assimilation methods for heavy rain forecast. \\u003cem\\u003eAtmos. Res.\\u003c/em\\u003e 270, 106062 (2022).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eBoyle, J. \\u0026amp; Klein, S. A. Impact of horizontal resolution on climate model forecasts of tropical precipitation and diabatic heating for the twp-ice period. \\u003cem\\u003eJ. Geophys. Res.: Atmos\\u003c/em\\u003e. 115, D23113 (2010).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eWilcox, E. M. \\u0026amp; Donner, L. J. The frequency of extreme rain events in satellite rain-rate estimates and an atmospheric general circulation model. \\u003cem\\u003eJ. Clim.\\u003c/em\\u003e 20, 53\\u0026ndash;69 (2007).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eZadra, A. et al. Systematic Errors in Weather and Climate Models: Nature, Origins, and Ways Forward. \\u003cem\\u003eBull. Amer. Meteor. Soc.\\u003c/em\\u003e 99, 67\\u0026ndash;70 (2018).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eWood, A. W., Maurer, E. P., Kumar, A. \\u0026amp; Lettenmaier, D. Long-range experimental hydrologic forecasting for the eastern United States, \\u003cem\\u003eJ. Geophys. Res.: Atmos.\\u003c/em\\u003e 107, 4429 (2002).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eZhao, T. et al. How suitable is quantile mapping for postprocessing GCM precipitation forecasts? \\u003cem\\u003eJ. Clim.\\u003c/em\\u003e 30, 3185\\u0026ndash;3196 (2017).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eRobertson, D. E., Shrestha, D. L. \\u0026amp; Wang, Q. J. Post-processing rainfall forecasts from numerical weather prediction models for short-term streamflow forecasting. \\u003cem\\u003eHydrol. Earth Syst. Sci.\\u003c/em\\u003e 17, 3587\\u0026ndash;3603 (2013).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eLi, L. et al. Improving categorical and continuous accuracy of precipitation forecasts by integrating empirical quantile mapping and bernoulli-gamma-gaussian distribution. \\u003cem\\u003eAtmos. Res.\\u003c/em\\u003e 298, 107133 (2024).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eJavanshiri, Z. \\u0026amp; Pakdaman, M. Bayesian joint probability approach for post-processing monthly precipitation prediction in Northwest Iran. \\u003cem\\u003eWater Resour. Res.\\u003c/em\\u003e 60, e2023WR036846 (2024).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eZhang, B. et al. A machine learning-based observational constraint correction method for seasonal precipitation prediction. \\u003cem\\u003eAdv. Atmos. Sci.\\u003c/em\\u003e 42, 36\\u0026ndash;52 (2025).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eRasp, S. \\u0026amp; Lerch, S. Neural networks for postprocessing ensemble weather forecasts. \\u003cem\\u003eMon. Wea. Rev.\\u003c/em\\u003e 146, 3885\\u0026ndash;3900 (2018).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eEbert-Uphoff, I. \\u0026amp; Hilburn, K. Evaluation, tuning and interpretation of neural networks for working with images in meteorological applications. \\u003cem\\u003eBull. Amer. Meteor. Soc.\\u003c/em\\u003e 101, 2149\\u0026ndash;2170 (2020).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eHu, Y. F., Yin, F. K. \\u0026amp; Zhang, W. M. Deep learning-based precipitation bias correction approach for Yin-He global spectral model. \\u003cem\\u003eMeteor. Appl.\\u003c/em\\u003e 28, e2032 (2021).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eHess, P. \\u0026amp; Boers, N. Deep learning for improving numerical weather prediction of heavy rainfall. \\u003cem\\u003eJ. Adv. Model. Earth Syst.\\u003c/em\\u003e 14, e2021MS002765 (2022).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eSun, D. et al. Deep learning improves GFS wintertime precipitation forecast over southeastern China. \\u003cem\\u003eGeophys. Res. Lett.\\u003c/em\\u003e 50, e2023GL104406 (2023).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eHunt, K. M. Stop using root-mean-square error as a precipitation target! \\u003cem\\u003earXiv preprint\\u003c/em\\u003e arXiv:2509.08369 (2025).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eLyu, Y. et al. Significant advancement in subseasonal-to-seasonal summer precipitation ensemble forecast skills in China mainland through an innovative hybrid CSG-UNET method. \\u003cem\\u003eEnviron. Res. Lett.\\u003c/em\\u003e 19, 074055 (2024).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eZhao, C., Ye, A., Wu, L. \\u0026amp; Zhan, S. A novel deep learning model for post-processing of short- and medium-term daily precipitation forecasts. \\u003cem\\u003eAtmos. Res.\\u003c/em\\u003e 326, 108319 (2025).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eCao, H. et al. UTrans-Net: a model for short-term precipitation prediction. \\u003cem\\u003eArtif. Intell. Appl.\\u003c/em\\u003e 1, 90\\u0026ndash;97 (2022).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eJiang, Q. et al. Hybrid multilayer perceptron and convolutional neural network model to predict extreme regional precipitation dominated by the large-scale atmospheric circulation. \\u003cem\\u003eAtmos. Res.\\u003c/em\\u003e 304, 107362 (2024).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eZhang, L., Zhao, Y., Cen, Y. \\u0026amp; Lu, M. Deep learning-based precipitation simulation for tropical cyclones, mesoscale convective systems, and atmospheric rivers in East Asia. \\u003cem\\u003eJ. Geophys. Res.: Atmos.\\u003c/em\\u003e 129, e2024JD041914 (2024).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eChen, G. \\u0026amp; Wang, W.-C. Short-term precipitation prediction for contiguous United States using deep learning. \\u003cem\\u003eGeophys. Res, Lett.\\u003c/em\\u003e 49, e2022GL097904 (2022).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eLing, F., Li, Y., Luo, J.-J., Zhong, X. \\u0026amp; Wang, Z. Two deep learning-based bias-correction pathways improve summer precipitation prediction over China. \\u003cem\\u003eEnviron. Res. Lett.\\u003c/em\\u003e 17, 124025 (2022).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eTang, T. et al. Robust and explainable objective classification of extreme precipitation events using machine learning. \\u003cem\\u003eJ. Meteor. Res.\\u003c/em\\u003e 39, 1\\u0026ndash;14 (2025).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eLi, S., Huang, X., Wu, W., Du, B. \\u0026amp; Jiang, Y. Evaluation of CMPAS precipitation products over Sichuan, China. \\u003cem\\u003eAtmos. Oceanic Sci. Lett.\\u003c/em\\u003e 15, 100129 (2022).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eLi, Y. et al. Convective storm VIL and lightning nowcasting using satellite and weather radar measurements based on multi-task learning models, \\u003cem\\u003eAdv. Atmos. Sci.\\u003c/em\\u003e 40, 887\\u0026ndash;899 (2023).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eLing, F. et al. Multi-task machine learning improves multi-seasonal prediction of the Indian Ocean Dipole. \\u003cem\\u003eNat Commun\\u003c/em\\u003e 13, 7681 (2022).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eRonneberger, O., Fischer, P. \\u0026amp; Brox, T. U-Net: Convolutional networks for biomedical image segmentation. \\u003cem\\u003eMed. Image Comput. Comput. Assist. Interv.\\u003c/em\\u003e 9351, 234\\u0026ndash;241 (2015).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eDonaldson, R., Dyer, R. M. \\u0026amp; Kraus, M. J. An objective evaluator of techniques for predicting severe weather events. \\u003cem\\u003eIn Ninth conference on severe local storms\\u003c/em\\u003e 321\\u0026ndash;326 (1975).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eWinterrath. T. \\u0026amp; Rosenow, W. A new module for the tracking of radar-derived precipitation with model-derived winds. \\u003cem\\u003eAdv. Geosci.\\u003c/em\\u003e 10, 77\\u0026ndash;83 (2007).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eWilks, D. S. Statistical Methods in the Atmospheric Sciences. \\u003cem\\u003eAcademic Press\\u003c/em\\u003e (2020).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eSrivastava, N. et al. Dropout: A simple way to prevent neural networks from overfitting. \\u003cem\\u003eJ. Mach. Learn. Res.\\u003c/em\\u003e 15, 1929\\u0026ndash;1958 (2014).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eSchultz, M. G. et al. Can deep learning beat numerical weather prediction? \\u003cem\\u003ePhilos. Trans. A Math. Phys. Eng. Sci.\\u003c/em\\u003e 379, 20200097 (2021).\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eLin, T.-Y., Goyal, P., Girshick, R., He, K. \\u0026amp; Doll\\u0026aacute;r, P. Focal loss for dense object detection. \\u003cem\\u003eIEEE Trans. Pattern Anal. Mach. Intell.\\u003c/em\\u003e 42, 318\\u0026ndash;327 (2020).\\u003c/span\\u003e\\u003c/li\\u003e\\u003c/ol\\u003e\"}],\"fulltextSource\":\"\",\"fullText\":\"\",\"funders\":[],\"hasAdminPriorityOnWorkflow\":false,\"hasManuscriptDocX\":true,\"hasOptedInToPreprint\":true,\"hasPassedJournalQc\":\"\",\"hasAnyPriority\":false,\"hideJournal\":false,\"highlight\":\"\",\"institution\":\"\",\"isAcceptedByJournal\":true,\"isAuthorSuppliedPdf\":false,\"isDeskRejected\":\"\",\"isHiddenFromSearch\":false,\"isInQc\":false,\"isInWorkflow\":false,\"isPdf\":false,\"isPdfUpToDate\":true,\"isWithdrawnOrRetracted\":false,\"journal\":{\"display\":true,\"email\":\"info@researchsquare.com\",\"identity\":\"npj-climate-and-atmospheric-science\",\"isNatureJournal\":false,\"hasQc\":true,\"allowDirectSubmit\":false,\"externalIdentity\":\"npjclimatsci\",\"sideBox\":\"Learn more about [npj Climate and Atmospheric Science](http://www.nature.com/npjclimatsci/)\",\"snPcode\":\"41612\",\"submissionUrl\":\"https://submission.springernature.com/new-submission/41612/3\",\"title\":\"npj Climate and Atmospheric Science\",\"twitterHandle\":\"\",\"acdcEnabled\":true,\"dfaEnabled\":true,\"editorialSystem\":\"stoa\",\"reportingPortfolio\":\"NPJ\",\"inReviewEnabled\":true,\"inReviewRevisionsEnabled\":true},\"keywords\":\"multi-task deep learning, bias correction, short-range prediction, heavy rainfall forecasts\",\"lastPublishedDoi\":\"10.21203/rs.3.rs-8314049/v1\",\"lastPublishedDoiUrl\":\"https://doi.org/10.21203/rs.3.rs-8314049/v1\",\"license\":{\"name\":\"CC BY 4.0\",\"url\":\"https://creativecommons.org/licenses/by/4.0/\"},\"manuscriptAbstract\":\"\\u003cp\\u003eImproving short-range forecasts of heavy rainfall remains a challenge. In artificial intelligence (AI) era, deep-learning-based methods have replaced traditional statistical post-processing for correcting precipitation forecasts from numerical weather prediction (NWP) models, but their performances are constrained by the non-negative and heavy-tailed nature of rainfall. Mainstream studies tried to solve this problem by redesigning loss functions or constructing hybrid models, yet they struggled to achieve both simplicity and transferability. Here we show that the biases between NWP forecasts and observations in heavy rainfall events follow an approximately Gaussian distribution. Accordingly, this study trains a multi-task U-Net that uses precipitation biases as the target. This bias-targeted strategy produces stable and substantial enhancements in short-range heavy rainfall forecasts, with improvements exceeding 21% across four in five regions of China. The findings highlight the critical role of target selection in deep-learning-based post-processing and provide a simple and effective pathway for advancing heavy rainfall forecasts.\\u003c/p\\u003e\",\"manuscriptTitle\":\"Bias-targeted deep learning enhances short-range heavy rainfall forecasts\",\"msid\":\"\",\"msnumber\":\"\",\"nonDraftVersions\":[{\"code\":1,\"date\":\"2025-12-17 17:27:36\",\"doi\":\"10.21203/rs.3.rs-8314049/v1\",\"editorialEvents\":[{\"type\":\"communityComments\",\"content\":0},{\"type\":\"decision\",\"content\":\"Revision requested\",\"date\":\"2026-01-19T17:05:05+00:00\",\"index\":\"\",\"fulltext\":\"\"},{\"type\":\"editorInvitedReview\",\"content\":\"\",\"date\":\"2026-01-11T07:44:36+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewerAgreed\",\"content\":\"249353598912848953070078882528887870242\",\"date\":\"2026-01-04T15:47:10+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"editorInvitedReview\",\"content\":\"\",\"date\":\"2026-01-02T20:22:18+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewerAgreed\",\"content\":\"339523652442673136508604453222750644189\",\"date\":\"2026-01-02T12:52:55+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewerAgreed\",\"content\":\"85677070074630624829420312803282342520\",\"date\":\"2025-12-31T03:31:33+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewerAgreed\",\"content\":\"50502948013335844183131279290175374767\",\"date\":\"2025-12-30T09:04:09+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewerAgreed\",\"content\":\"110193762477689221314804438456213831096\",\"date\":\"2025-12-15T14:42:22+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewerAgreed\",\"content\":\"114403343457300498549485232083667689768\",\"date\":\"2025-12-11T14:42:40+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewerAgreed\",\"content\":\"82175937923735454050020419947645340995\",\"date\":\"2025-12-11T14:36:12+00:00\",\"index\":\"hide\",\"fulltext\":\"\"},{\"type\":\"reviewersInvited\",\"content\":\"\",\"date\":\"2025-12-11T13:26:56+00:00\",\"index\":\"\",\"fulltext\":\"\"},{\"type\":\"editorAssigned\",\"content\":\"\",\"date\":\"2025-12-11T05:22:03+00:00\",\"index\":\"\",\"fulltext\":\"\"},{\"type\":\"checksComplete\",\"content\":\"\",\"date\":\"2025-12-11T05:17:48+00:00\",\"index\":\"\",\"fulltext\":\"\"},{\"type\":\"submitted\",\"content\":\"npj Climate and Atmospheric Science\",\"date\":\"2025-12-09T06:40:11+00:00\",\"index\":\"\",\"fulltext\":\"\"}],\"status\":\"published\",\"journal\":{\"display\":true,\"email\":\"info@researchsquare.com\",\"identity\":\"npj-climate-and-atmospheric-science\",\"isNatureJournal\":false,\"hasQc\":true,\"allowDirectSubmit\":false,\"externalIdentity\":\"npjclimatsci\",\"sideBox\":\"Learn more about [npj Climate and Atmospheric Science](http://www.nature.com/npjclimatsci/)\",\"snPcode\":\"41612\",\"submissionUrl\":\"https://submission.springernature.com/new-submission/41612/3\",\"title\":\"npj Climate and Atmospheric Science\",\"twitterHandle\":\"\",\"acdcEnabled\":true,\"dfaEnabled\":true,\"editorialSystem\":\"stoa\",\"reportingPortfolio\":\"NPJ\",\"inReviewEnabled\":true,\"inReviewRevisionsEnabled\":true}}],\"origin\":\"\",\"ownerIdentity\":\"24f9fb54-a073-45b3-8b9f-95ccc6c259bb\",\"owner\":[],\"postedDate\":\"December 17th, 2025\",\"published\":true,\"recentEditorialEvents\":[],\"rejectedJournal\":[],\"revision\":\"\",\"amendment\":\"\",\"status\":\"published-in-journal\",\"subjectAreas\":[{\"id\":59763578,\"name\":\"Earth and environmental sciences/Climate sciences\"},{\"id\":59763579,\"name\":\"Earth and environmental sciences/Environmental sciences\"},{\"id\":59763580,\"name\":\"Earth and environmental sciences/Hydrology\"},{\"id\":59763581,\"name\":\"Physical sciences/Mathematics and computing\"},{\"id\":59763582,\"name\":\"Earth and environmental sciences/Natural hazards\"}],\"tags\":[],\"updatedAt\":\"2026-03-09T16:12:07+00:00\",\"versionOfRecord\":{\"articleIdentity\":\"rs-8314049\",\"link\":\"https://doi.org/10.1038/s41612-026-01366-z\",\"journal\":{\"identity\":\"npj-climate-and-atmospheric-science\",\"isVorOnly\":false,\"title\":\"npj Climate and Atmospheric Science\"},\"publishedOn\":\"2026-03-06 15:58:42\",\"publishedOnDateReadable\":\"March 6th, 2026\"},\"versionCreatedAt\":\"2025-12-17 17:27:36\",\"video\":\"\",\"vorDoi\":\"10.1038/s41612-026-01366-z\",\"vorDoiUrl\":\"https://doi.org/10.1038/s41612-026-01366-z\",\"workflowStages\":[]},\"version\":\"v1\",\"identity\":\"rs-8314049\",\"journalConfig\":\"researchsquare\"},\"__N_SSP\":true},\"page\":\"/article/[identity]/[[...version]]\",\"query\":{\"redirect\":\"/article/rs-8314049\",\"identity\":\"rs-8314049\",\"version\":[\"v1\"]},\"buildId\":\"8U1c8b4HqxoKbykW_rLl7\",\"isFallback\":false,\"isExperimentalCompile\":false,\"dynamicIds\":[84888],\"gssp\":true,\"scriptLoader\":[]}","source_license":"CC-BY-4.0","license_restricted":false}