DPSformer: A long-tail-aware model for improving heavy rainfall prediction

TL;DR

DPSformer reframes heavy rainfall forecasting as a long-tailed learning problem and introduces a dual-path Transformer architecture with a high-resolution spatial branch to boost representation of rare, high-impact events. By adaptively fusing global backbone features with localized high-resolution cues and employing a dual-loss training scheme, the model significantly improves heavy rainfall prediction, elevating CSI from $0.012$ to $0.067$ for events $\geq$ $50$ mm/6 h and achieving top-tail FSS above $0.4$ for the most extreme cases. The approach is validated with a tropical cyclone case (e.g., Tropical Storm Doksuri) and attribution analyses showing physically consistent feature usage (e.g., emphasis on upper-level moisture and deep lifting), while uncertainty is quantified via conformal prediction. This work offers a practical, long-tailed paradigm for operational heavy rainfall forecasting, with potential applicability to other rare, high-impact meteorological phenomena and integration with multi-source data for broader generalization.

Abstract

Accurate and timely forecasting of heavy rainfall remains a critical challenge for modern society. Precipitation exhibits a highly imbalanced distribution: most observations record no or light rain, while heavy rainfall events are rare. Such an imbalanced distribution obstructs deep learning models from effectively predicting heavy rainfall events. To address this challenge, we treat rainfall forecasting explicitly as a long-tailed learning problem, identifying the insufficient representation of heavy rainfall events as the primary barrier to forecasting accuracy. Therefore, we introduce DPSformer, a long-tail-aware model that enriches representation of heavy rainfall events through a high-resolution branch. For heavy rainfall events $ \geq $ 50 mm/6 h, DPSformer lifts the Critical Success Index (CSI) of a baseline Numerical Weather Prediction (NWP) model from 0.012 to 0.067. For the top 1% coverage of heavy rainfall events, its Fraction Skill Score (FSS) exceeds 0.45, surpassing existing methods. Our work establishes an effective long-tailed paradigm for heavy rainfall prediction, offering a practical tool to enhance early warning systems and mitigate the societal impacts of extreme weather events.

Figures (8)

• Figure 1: Rainfall distribution in the training dataset. (a) Probability density distribution of rainfall (mm/6 h) in the training dataset, with a fit obtained using a gamma distribution. (b) Percentage distribution of rainfall classified into six levels based on predefined thresholds (0.1, 3, 10, 20, and 50 mm/6 h).
• Figure 2: Performance comparison of baseline methods under long-tailed rainfall conditions. CSI scores of DPSformer and five baselines evaluated at five increasing 6-h accumulated rainfall thresholds. Error bars denote 95% confidence intervals, estimated via bootstrap (n=1000).
• Figure 3: Performance comparison of baseline methods for high-coverage heavy rainfall events. (a) Cumulative distribution of the occurrence proportion of heavy rainfall ($\geq$50 mm / 6 h) in the test dataset, with lines marking the top 5% and top 1% percentiles. (b) Probability distribution of the high-coverage heavy rainfall events across all test dataset. (c) CSI and FSS performance of different methods for high-coverage heavy rainfall events. Evaluations are based on rainfall maps ranked by descending order of heavy rainfall coverage percentage (from top 25% to top 1%). Bar heights represent CSI values (left y-axis), while the overlaid lines indicate FSS scores (right y-axis) for the respective models. FSS is calculated using a neighborhood size of 5 grid points.
• Figure 4: Warm-sector mesoscale convective system (MCS) heavy rainfall events. This figure compares the ability of TIGGE and three deep learning methods (SegFormer, OBDice and DPSformer) to localize heavy rainfall events. The comparison is shown for two heavy rainfall categories: Level 4 ([20, 50) mm/6 h) and Level 5 ($\geq$50 mm/6 h). While the ground truth and the TIGGE output are represented as binary occurrence maps (presence/absence), the deep learning methods generate probability fields. (a) shows the case for 2012.03.04 18:00-24:00 UTC; (b) shows the case for 2012.05.11 18:00-24:00 UTC.
• Figure 5: Tropical cyclone heavy rainfall (Tropical Storm Doksuri). Event date: 2012-06-29 (UTC). (a) Study region used for evaluation, outlined in red on the rainfall map, together with the 6-hour accumulated rainfall from 18--24 h UTC on the previous day, representing the initial rainfall state before the case day. (b) Performance comparison of DPSformer and five baselines in terms of CSI and FSS at heavy rainfall Level 4 ([20, 50) mm/6 h) and Level 5 ($\geq$50 mm/6 h), evaluated for four consecutive 6-hour periods within the case day. FSS is calculated using a neighborhood size of 5 grid points. (c) Spatial distribution of 6-hourly accumulated rainfall over four consecutive periods within that day (00--06 h, 06--12 h, 12--18 h, and 18--24 h UTC) from ground truth, DPSformer, and the five baselines.