RainBalance: Alleviating Dual Imbalance in GNSS-based Precipitation Nowcasting via Continuous Probability Modeling

TL;DR

RainBalance Tackles dual imbalance in GNSS-based precipitation nowcasting by transforming imbalanced targets into continuous probabilistic labels through temporal sample clustering and a VAE. The method is plug-and-play, integrating with diverse backbones via a fusion module that adds a probabilistic latent to backbone embeddings. Key contributions include a novel Temporal Sample Cluster Assignment, a continuous probability modeling stage, and a balanced fusion mechanism, validated across multiple stations and resolutions with notable improvements in both standard and extreme precipitation forecasting. The approach demonstrates strong generalization and practical value for disaster risk reduction and real-time hydrometeorological decision support.

Abstract

Global navigation satellite systems (GNSS) station-based Precipitation Nowcasting aims to predict rainfall within the next 0-6 hours by leveraging a GNSS station's historical observations of precipitation, GNSS-PWV, and related meteorological variables, which is crucial for disaster mitigation and real-time decision-making. In recent years, time-series forecasting approaches have been extensively applied to GNSS station-based precipitation nowcasting. However, the highly imbalanced temporal distribution of precipitation, marked not only by the dominance of non-rainfall events but also by the scarcity of extreme precipitation samples, significantly limits model performance in practical applications. To address the dual imbalance problem in precipitation nowcasting, we propose a continuous probability modeling-based framework, RainBalance. This plug-and-play module performs clustering for each input sample to obtain its cluster probability distribution, which is further mapped into a continuous latent space via a variational autoencoder (VAE). By learning in this continuous probabilistic space, the task is reformulated from fitting single and imbalance-prone precipitation labels to modeling continuous probabilistic label distributions, thereby alleviating the imbalance issue. We integrate this module into multiple state-of-the-art models and observe consistent performance gains. Comprehensive statistical analysis and ablation studies further validate the effectiveness of our approach.

Figures (6)

• Figure 1: Distribution of precipitation values at the J340 GNSS station. The left panel shows the proportion of zero and non-zero precipitation samples, revealing a strong zero-inflation (80.80% of all samples record no rainfall). The right panel further decomposes the non-zero samples, indicating that light rainfall events (0–8 mm/h) dominate the dataset (93.38%), while extreme precipitation events (>=8 mm/h) are relatively rare (6.62%).
• Figure 2: An Overview of RainBalance for Addressing the Dual Imbalance Problem in Precipitation Nowcasting. This diagram illustrates a staged methodology for tackling both zero-dominance and extreme scarcity in precipitation data. It begins by clustering ambiguous "no-rain" states into multi weather states, and then proceeds with continuous probability modeling to effectively capture and forecast precipitation events.
• Figure 3: An overview of the proposed RainBalance framework. The input time series is processed by the encoder of a backbone model to obtain hidden embeddings. The Temporal Sample Cluster Assignment Module (lower left) computes sample similarities and assigns samples to clusters, generating a cluster probability distribution and prototype embeddings. The Continuous Probability Modeling Module (upper left), structured as a Variational Autoencoder (VAE), encodes this distribution into a latent space z and decodes it to obtain new cluster probabilities. Finally, the Fusion Prediction Module (upper right) integrates the backbone embeddings with the weighted cluster prototypes to produce the final, balanced precipitation forecast.
• Figure 4: Global distribution of 140 selected GNSS stations from the RainfallBench dataset across seven continents, ensuring balanced spatial coverage for evaluating precipitation forecasting models zhang2025effectivetimeseriesmodelsprecipitation. Four representative stations with different spatial distributions and temporal resolutions (J340, ZIMM, P095, and JFNG) were selected as case-study sites for detailed analysis.
• Figure 5: Comparison of the Mean Absolute Error (MAE) improvement rates achieved by different configurations at various sites.