- 1University of Birmingham, UK
- 2National Centre for Medium Range Weather Forecasting, India
- 3British Antarctic Survey, UK
Extreme precipitation events in India are becoming more frequent and intense, increasing the need for reliable ensemble precipitation forecasts to support early warning systems and disaster preparedness. However, current Numerical Weather Prediction models often underestimate extreme precipitation, and their forecast skill is constrained by errors in initial conditions, numerical approximations, inadequate representation of sub-grid convective processes, and coarse spatial resolution. Additionally, systematic biases in ensemble forecast distributions, such as deviations in central tendency and under- or over-dispersion, further limit the accuracy of probabilistic forecasts. Ensemble Model Output Statistics (EMOS) can reduce some of these limitations by correcting systematic biases in the ensemble mean and spread, and by partly adjusting the predicted overall distribution. Classical EMOS relies on linear transformations, limiting the ability to capture non-linear relationships between the original forecast and the corrected ensemble, and to correct asymmetric distribution errors. Moreover, it derives the corrected distribution at a given location only from the original forecast ensemble for this location. Deep learning-based distributional regression methods, such as U-Net architectures, can generalise classical EMOS by linking the original full spatial field of ensemble forecasts in a complex way to the corrected ensemble forecasts.
This study presents a U-Net based distributional regression (DRU) for daily rainfall forecasts over India, that predicts parametric marginal distributions at each forecast grid cell from the statistics of the original ensemble forecast at all grid cells. It minimises the area mean Continuous Ranked Probability Score (CRPS), while classical EMOS minimises the CRPS individually for each location. DRU is applied to postprocess forecasts with one day lead time for daily precipitation at 12-km resolution from 11 members of the National Centre for Medium Range Weather Forecasting (NCMRWF) global ensemble prediction system for the period 2018-2024. The observations for U-Net training are gridded precipitation data at 0.25° resolution from the Indian Meteorological Department and the NCMRWF forecasts were regridded to this resolution prior to DRU training. Over most of India, DRU improves local precipitation distributions, including for higher quantiles, and corrects under- or overdispersion. The forecast skill in terms of Continuous Ranked Probability Skill Score increases over large areas, particularly northern and western India, while in central and northeastern regions, there are locations where the skill decreases. For some of these, the marginal distributions are also not improved. Additionally, DRU improves the reliability for predicting the exceedance probability of various precipitation thresholds. Future endeavours will focus on optimizing DRUs for postprocessing heavy precipitation events and evaluating the forecast skill using the Brier Score.
How to cite: Thakur, C., Widmann, M., Ashrit, R., Orr, A., C. Leckebusch, G., and Geen, R.: Improving precipitation ensemble forecasts over India using a convolutional distributional regression framework, EGU General Assembly 2026, Vienna, Austria, 3–8 May 2026, EGU26-9569, https://doi.org/10.5194/egusphere-egu26-9569, 2026.