Organizations: Sanya Science and Education Innovation Park, Wuhan University of Technology, Sanya, 572000, China · School of Computer Science and Artificial Intelligence, Wuhan University of Technology, Wuhan 430070, China · School of Earth and Space Science and Technology, Wuhan University, Wuhan 430072, China · School of Water Resources and Hydropower Engineering, Wuhan University, Wuhan 430062, China · GNSS Research Center, Wuhan University, Wuhan 430062, China
Accurate station-level precipitation nowcasting is critical for agriculture, water resource management, and disaster prevention, which typically is formulated as a time series forecasting problem. However, conventional time-series modeling techniques face two major challenges in addressing station-level precipitation nowcasting: (1) Lack of Physics-Guided Modeling}, where meteorological variables are treated as a homogeneous set without accounting for their distinct roles in precipitation formation, leads to predictions that deviate from the physical processes governing precipitation. (2) Severe zero inflation in precipitation, where dry intervals dominate the dataset, obscuring meaningful precipitation patterns and complicating the predictive modeling. To address these challenges, we propose MZ-Rain, a moisture-budget-guided zero-inflated sLSTM framework for station-level precipitation nowcasting. Guided by the moisture budget equation, MZ-Rain decomposes the precipitation formation process into process-specific pathways corresponding to moisture storage, moisture transport, surface evaporation, and precipitation persistence, and captures their temporal evolution through dedicated sLSTM branches. To account for the zero-inflated nature of precipitation, MZ-Rain introduces an adaptive Tweedie modeling strategy that adaptively modulates the rainfall mean while jointly learning precipitation occurrence as an auxiliary task, enabling the model to better balance dry-wet discrimination and quantitative precipitation estimation. Extensive experiments across diverse geographical and climatic regimes demonstrate that MZ-Rain consistently outperforms strong baselines on multiple evaluation metrics, including CSI, FAR, MSE, and MAE. In particular, the model exhibits superior skill in forecasting heavy precipitation events, while benefiting from physically grounded process modeling.
Figures & tables
Fig. 1: Motivation of the proposed MZ-Rain, highlighting the two key challenges and their corresponding solutions.
Variable
Description
Source
Pathway
Wt
Precipitable water vapor
GNSS
ms
ΔWt
Temporal difference of PWV
GNSS-derived
ms
Ptsfc
Surface pressure
ERA5 Reanalysis
ms
Ct
Moisture-flux convergence
ERA5-derived
tr
Et
Surface evaporation
ERA5 Reanalysis
ev
yt
Historical precipitation
IMERG
pr
TABLE I: Input variables used by MZ-Rain and their corresponding data sources and process pathways.
Fig. 2: Geographical distribution of the six GNSS stations used in this study. The selected stations span diverse geographic locations and climatic conditions, providing broad spatial coverage for precipitation forecasting evaluation.
Fig. 3: Architecture of MZ-Rain. The left side shows four process-specific pathways for moisture storage, moisture transport and convergence, surface evaporation, and precipitation persistence. The right side shows adaptive Tweedie modeling for quantitative precipitation estimation together with auxiliary rainfall-occurrence supervision.
Model
2 h
4 h
6 h
Avg. Rank
MAE ↓
MSE ↓
CSI ↑
FAR ↓
MAE ↓
MSE ↓
CSI ↑
FAR ↓
MAE ↓
MSE ↓
CSI ↑
FAR ↓
sLSTM
0.2909
1.3102
0.2999
0.6748
0.3311
1.5199
0.2567
0.7186
0.3547
1.6203
0.2370
0.7438
6.67
FilterTS
0.2732
1.3985
0.3428
0.5839
0.3033
1.6486
0.2904
0.6373
0.3127
1.7699
0.2603
0.6671
5.58
TimeFilter
0.2510
1.3726
0.3577
0.5658
0.2794
1.6178
0.3002
0.6210
0.2941
1.7262
0.2641
0.6686
3.75
TimeKAN
0.2559
1.3848
0.3625
0.5535
0.2796
1.6200
0.3060
0.6098
0.2945
1.7386
0.2690
0.6511
3.67
xPatch
0.2517
1.4010
0.3664
0.5493
0.2780
1.6162
0.2984
0.6228
0.2928
1.7369
0.2657
0.6593
3.50
TABLE II: Cross-station macro-average performance at 2/4/6 h horizons. CSI ↑ , FAR ↓ , MSE ↓ , and MAE ↓ . Avg. Rank denotes the mean rank over all twelve horizon–metric combinations, where a lower value is better.
Fig. 4: Per-station detection performance at 2/4/6 h horizons. Each panel shows all ten models across the six GNSS stations (J340, ZIMM, P095, MTLA, ARTA, BFTA); the red solid line is MZ-Rain (Ours) . Top: CSI ( ↑ , higher is better). Bottom: FAR ( ↓ , lower is better). MZ-Rain attains the highest CSI at every station and every horizon and the lowest FAR in nearly all cases; at P095, the lower FAR of ZIP is accompanied by an extremely low CSI, indicating severe under-detection of precipitation events(cf. Figure 4 and Figure 5 ).
Fig. 5: Per-station intensity error at 2/4/6 h horizons. Top: MSE ((mm/h) 2 , ↓ ). Bottom: MAE (mm/h, ↓ ). MZ-Rain (red) achieves the lowest MAE at all six stations and all horizons, and the lowest MSE in nearly all cases.
−∂W/∂t
C=−∇⋅Q
E
H (h)
Spear.
AP
Spear.
AP
Spear.
AP
1
0.286
0.387
0.354
0.509
0.172
0.257
2
0.309
0.401
0.368
0.506
0.186
0.271
4
0.321
0.405
0.368
0.465
0.201
0.282
6
0.307
0.382
0.345
0.410
0.194
0.274
TABLE III: Lead–lag relationship between moisture-budget components at time t and accumulated future precipitation Pt+1:t+H at J340. Spear. denotes Spearman correlation and AP denotes wet-event average precision. The wet-event base rate is approximately 0.154.
Variant
ΔW
C
E
Persist.
MAE ↓
MSE ↓
CSI ↑
FAR ↓
Full
✓
✓
✓
✓
0.2841
1.5929
0.3243
0.5895
w/o storage
×
✓
✓
✓
0.2887
1.6154
0.3132
0.6028
w/o convergence
✓
×
✓
✓
0.2948
1.6537
0.2961
0.6254
w/o evaporation
✓
✓
×
✓
0.2869
1.6068
0.3176
0.5978
w/o persistence
✓
✓
✓
×
0.3015
1.6842
0.3040
0.6189
TABLE IV: Ablation of moisture-budget-related inputs and precipitation persistence at the 6 h horizon. Results are six-station macro averages.
Variant
MAE ↓
MSE ↓
CSI ↑
FAR ↓
Shared Encoder
0.3076
1.6812
0.3004
0.6018
Generic Multi-Branch
0.2918
1.6350
0.3110
0.6072
Process-Specific Organization
0.2841
1.5929
0.3243
0.5895
TABLE V: Comparison of different representation organizations at the 6 h forecasting horizon.
Variant
MAE ↓
MSE ↓
CSI ↑
FAR ↓
w/o PWV gate
0.2896
1.6217
0.3154
0.6018
w/o transport gate
0.2915
1.6359
0.3097
0.6108
w/o persistence gate
0.2924
1.6463
0.3168
0.6006
w/o all contextual gates
0.2988
1.6942
0.2989
0.6245
Full
0.2841
1.5929
0.3243
0.5895
TABLE VI: Ablation study of the contextual gating mechanism at the 6 h forecasting horizon. Results are macro-averaged over the six stations.
Variant
Tweedie
Mod.
Occ.
Output
MAE ↓
MSE ↓
CSI ↑
FAR ↓
Point regression
×
×
×
μ~
0.3048
1.7246
0.2891
0.6328
Tweedie only
✓
×
×
μ~
0.2962
1.6657
0.3028
0.6164
Tweedie + occurrence
✓
×
✓
μ~
0.2908
1.6294
0.3165
0.6008
Tweedie + modulation
✓
✓
×
μ
0.2889
1.6172
0.3116
0.6052
Full
✓
✓
✓
μ
0.2841
1.5929
0.3243
0.5895
TABLE VII: Component-wise ablation of the adaptive Tweedie-based multi-task formulation at the 6h forecasting horizon. Results are macro-averaged over the six stations. μ~ denotes the unmodulated mean prediction, while μ=gμμ~+ϵ denotes the adaptively modulated mean. The occurrence prediction q is used only for auxiliary supervision and does not enter the final precipitation amount.
Bias (mm/h)
MAE (mm/h)
Peak ratio
Regime
raw μ~
gated μ
raw
gated
raw
gated
Dry
+0.071
+0.034
0.079
0.050
N/A
N/A
Light
+0.083
+0.032
0.196
0.174
1.18
1.08
Moderate
+0.061
+0.018
0.514
0.482
1.07
1.01
Heavy
-0.182
-0.205
1.218
1.226
0.94
0.92
TABLE VIII: Behavior of adaptive mean modulation across observed precipitation regimes on the six-station test set. Bias is defined as prediction minus observation. A peak ratio closer to one indicates better preservation of high-intensity precipitation.
Model
ARTA
BFTA
J340
MTLA
P095
ZIMM
MAE ↓
MSE ↓
MAE ↓
MSE ↓
MAE ↓
MSE ↓
MAE ↓
MSE ↓
MAE ↓
MSE ↓
MAE ↓
MSE ↓
sLSTM
12.2241
185.1727
13.3234
203.5121
9.9672
121.2693
11.8161
163.0870
19.2073
593.6420
16.1272
332.9984
TimeFilter
12.7852
202.5680
13.4437
208.0682
10.0554
128.3993
12.2085
174.8581
19.1827
598.4038
15.7476
319.3812
TimeKAN
12.7985
202.5675
13.4795
209.8473
10.0872
129.2633
11.9936
170.6457
19.1857
598.1708
15.8169
321.3663
xPatch
12.8199
202.7773
13.4888
209.8547
10.1312
130.0660
12.1957
174.6668
19.1528
597.0271
15.8126
321.6467
FilterTS
13.1206
209.9214
13.3106
204.9712
10.0436
128.8823
12.0950
173.4394
19.2252
599.3501
15.6878
317.1874
TABLE IX: Conditional intensity accuracy for observed heavy precipitation (10–50mm/h) under the 24-to-6 prediction setting.
Model
Brier ↓
AUCPR ↑
ECE ↓
MAE ↓
MSE ↓
CSI ↑
FAR ↓
Full ( q + BCE)
0.086
0.612
0.028
0.2841
1.5929
0.3243
0.5895
w/o occurrence task
N/A
N/A
N/A
0.2889
1.6172
0.3116
0.6052
TABLE X: Quality of the auxiliary occurrence prediction and its contribution to precipitation forecasting on the six-station test set. Brier score, AUCPR, and ECE evaluate the supervised occurrence probability q . N/A indicates that the metric is not applicable because the occurrence head is absent.
Model
MAE
MSE
sLSTM
69.9135
5475.8630
TimeFilter
69.5046
5452.9663
TimeKAN
69.9554
5499.4810
xPatch
69.9045
5491.0386
FilterTS
70.9552
5607.1426
BFPF
69.9475
5472.3105
TABLE XI: Conditional intensity accuracy for extreme precipitation ( >50 mm/h) at ARTA under the 24-to-6 prediction setting. This regime is reported as an exploratory evaluation because valid samples are available only at ARTA.
Fig. 6: Qualitative forecasting comparisons under three representative precipitation regimes. (a) Dry event at Station J340 (2023-12-27 01:00–06:00): MZ-Rain maintains near-zero predictions and effectively suppresses false positives, while baselines produce spurious rainfall signals. (b) Normal precipitation event at Station ZIMM (2023-05-05 16:00–21:00): MZ-Rain tracks moderate temporal variations more faithfully, whereas baseline methods tend to over-smooth the signal and underestimate fluctuations. (c) Heavy precipitation event at Station ARTA (2024-05-24 12:00–17:00): MZ-Rain preserves sharp peak intensity, while competing models underestimate the event magnitude due to regression-to-the-mean effects.
Precipitation nowcasting is a vital spatio-temporal prediction task for meteorological applications but faces challenges due to the chaotic property of precipitation systems. Existing methods predominantly rely on single-source radar data to build either deterministic or probabilistic models for extrapolation. However, the single deterministic model suffers from blurring due to MSE convergence. The single probabilistic model, typically represented by diffusion models, can generate fine details but suffers from spurious artifacts that compromise accuracy and computational inefficiency. To address these challenges, this paper proposes a novel coarse-to-fine Vision Mamba Unet and residual Diffusion (VMU-Diff) based precipitation nowcasting framework. It realizes precipitation nowcasting through a two-stage process, i.e., a deterministic model-based coarse stage to predict global motion trends and a probabilistic model-based fine stage to generate fine prediction details. In the coarse prediction stage, rather than single-source radar data, both radar and multi-band satellite data are taken as input. A spatial-temporal attention block and several Vision mamba state-space blocks realize multi-source data fusion, and predict the future echo global dynamics. The fine-grained stage is realized by a spatio-temporal refine generator based on residual conditional diffusion models. It first obtains spatio-temporal residual features based on coarse prediction and ground truth, and further reconstructs the residual via conditional Mamba state-space module. Experiments on Jiangsu SWAN datasets demonstrate the improvements of our method over state-of-the-art methods, particularly in short-term forecasts.
Chunlei Shi, Hao Li, Yufeng Zhu +6
Department of Automation, Southeast University, Nanjing 210096, China
Accurate and timely rainfall nowcasting is crucial for disaster mitigation and water resource management. Despite recent advances in deep learning, precipitation prediction remains challenging due to limitations in effectively leveraging diverse multimedia data sources. We introduce M3R, a Meteorology-informed MultiModal attention-based architecture for direct Rainfall prediction that synergistically combines visual NEXRAD radar imagery with numerical Personal Weather Station (PWS) measurements, using a comprehensive pipeline for temporal alignment of heterogeneous meteorological data. With specialized multimodal attention mechanisms, M3R novelly leverages weather station time series as queries to selectively attend to spatial radar features, enabling focused extraction of precipitation signatures. Experimental results for three spatial areas of 100 km * 100 km centered at NEXRAD radar stations demonstrate that M3R outperforms existing approaches, achieving substantial improvements in accuracy, efficiency, and precipitation detection capabilities. Our work establishes new benchmarks for multimedia-based precipitation nowcasting and provides practical tools for operational weather prediction systems. The source code is available at https://github.com/Sanjeev97/M3Rain
Sanjeev Panta, Rhett M Morvant, Xu Yuan +2
University of Louisiana at Lafayette, Lafayette, LA, USA · University of Delaware, Newark, DE, USA
Proper short-term forecasting of precipitation is crucial in disaster management and preparedness. Nonetheless, the variability and nonlinearity of precipitation make short-term forecasting challenging for meteorologists. Moreover, capturing temporal dependencies in spatiotemporal data is a challenge in precipitation nowcasting. In this article, we introduce a lightweight deep learning model for half-hourly precipitation nowcasting. This model has been designed by incorporating the DenseNet architecture, residual connections, and transformer encoders for effective precipitation nowcasting with reduced model parameters. The North-Eastern region of India has been selected as the area of interest for our study. The region receives the highest precipitation during the months of June-September due to the monsoon season. The proposed model takes the previous five time-steps of half-hourly precipitation as inputs and predicts the precipitation in the next two half-hours. The GPM IMERG precipitation dataset with a 30-minute cadence has been used in this study for training and testing the model. The proposed architecture achieves best MAE of 0.235 millimetres, RMSE of 0.735 millimetres, and KGE score of 0.816 at an interval of 30 minutes.
Gourav Jyoti Kalita, Hidam Kumarjit Singh
Department of Electronics and Communication Technology, Gauhati University, Guwahati-781014, Assam, India