Accurate Precipitation Forecast by Efficiently Learning from Massive Atmospheric Variables and Unbalanced Distribution

Shuangliang Li; Siwei Li; Li Li; Weijie Zou; Jie Yang; Maolin Zhang

cs.LGMar 27, 2026

Accurate Precipitation Forecast by Efficiently Learning from Massive Atmospheric Variables and Unbalanced Distribution

Authors: Shuangliang Li, Siwei Li, Li Li, Weijie Zou, Jie Yang, Maolin Zhang

Organizations: Hubei Key Laboratory of Quantitative Remote Sensing of Land and Atmosphere, School of Remote Sensing and Information Engineering, Wuhan University, Wuhan, 430079, China · Perception and Effectiveness Assessment for Carbon-neutrality Efforts, Engineering Research Center of Ministry of Education, Institute for Carbon Neutrality, Wuhan University, Wuhan, 430072, China · Hubei Meteorological Information and Technology Support Center, Wuhan, 430074, China

Abstract

Short-term (0-24 hours) precipitation forecasting is highly valuable to socioeconomic activities and public safety. However, the highly complex evolution patterns of precipitation events, the extreme imbalance between precipitation and non-precipitation samples, and the inability of existing models to efficiently and effectively utilize large volumes of multi-source atmospheric observation data hinder improvements in precipitation forecasting accuracy and computational efficiency. To address the above challenges, this study developed a novel forecasting model capable of effectively and efficiently utilizing massive atmospheric observations by automatically extracting and iteratively predicting the latent features strongly associated with precipitation evolution. Furthermore, this study introduces a 'WMCE' loss function, designed to accurately discriminate extremely scarce precipitation events while precisely predicting their intensity values. Extensive experiments on two datasets demonstrate that our proposed model substantially and consistently outperforms all prevalent baselines in both accuracy and efficiency. Moreover, the proposed forecasting model substantially lowers the computational cost required to obtain valuable predictions compared to existing approaches, thereby positioning it as a milestone for efficient and practical precipitation forecasting.

Figures & tables

	Cumulative Precipitation Rates (QPE, mm/h)
Measure Type	$\leq$ 0.2	0.2-1	1-2	2-4	4-8	$\geq$ 8
‘USA’ Dataset	93.31%	3.45%	1.42%	0.97%	0.52%	0.33%
	Radar Reflectivity (dBZ)
Measure Type	$\leq$ 20	20-25	25-30	30-35	35-40	$\geq$ 40
‘Hubei’ Dataset	98.28%	0.82%	0.51%	0.26%	0.10%	0.04%

TABLE I: Hourly cumulative Precipitation Rates (mm/h) and Radar Reflectivity (dBZ) Value Distribution by Bucket. Note that the radar reflectivity exceeding 20 dBZ generally corresponds to the presence of precipitation events with the precipitation rate approximately great than 0.2 mm/h [ 1 ] [ 2 ] [ 3 ] .

Fig. 1: Comparison between the precipitation prediction framework. ConvLSTM [ 4 ] and PredRNN [ 5 ] primarily focus on ultra-short-term severe convection and thunderstorms nowcasting. Metnet models [ 6 ] [ 7 ] [ 8 ] employs a dense interval-partitioned cross-entropy loss to train its single-lead-time network. In contrast, our novelly designed latent space iterative prediction framework integrates a ‘WMCE’ loss to train the network for short-term precipitation forecasting.

Fig. 2: Network structure and loss function of our proposed forecasting model. The ‘labels’ include the MRMS hourly cumulative QPE from ‘USA’ dataset or near-surface radar reflectivity from ‘Hubei’ dataset. And the ‘MAE’ and ‘CE’ losses for precipitation samples employ a weighting scheme, as presented in Eq. 17 and Eq. 18 . ‘LPM’ represents the latent feature iterative prediction model. It is worth noting that the encoder simultaneously encodes the latent features from two time steps.

Fig. 3: Network structure of the designed encoder. Note that the input for the ‘USA’ dataset does not include satellite imagery and the input for the ‘Hubei’ dataset does not include MRMS QPE.

Fig. 4: The network structure of ViT-based latent feature iterative prediction model ( $LPM$ ). It takes the latent atmospheric features of the previous two time steps as input and outputs the predicted latent feature for the next time step. The input also includes the corresponding time embeddings and constant embeddings. After passing through the multi-scale residual blocks, 3D convolutional layer, ViT, and the multi-scale residual blocks again, the predicted results are output. Note that the ‘multi-scale residual blocks’ are identical to those in the encoder, as depicted in Fig. 3 .

Fig. 5: A schematic diagram of the inference process for the improved HTA algorithm, which takes two time steps as input for iterative prediction. Arrows of different colors represent predictions with different time step intervals.

Dataset	Region	Input Variables	Spatial	Predicted variable	Time
name		(2 time step)	resolution	(24 time steps)	range
USA	100.9°W-83.1°W ‌	ERA5 variables(243)	0.25°(77×73)	Cumulative	2015.01.01-
dataset	27.6°N-46.4°N	3D Nexrad reflectivity(24)	6km(256×256)	MRMS QPE	2018.01.01
		Cumulative MRMS QPE (1)	6km(256×256)	(1×256×256)
Hubei	111.9°E-116.8°E	ERA5 variables(243)	0.25°(37×37)	Near ground radar	2022.6.1-10.1
dataset	28.3°N-32.7°N	FY4B satellite L1 data(15)	4km(256×256)	reflectivity	2023.5.1-10.1

TABLE II: The details of two datasets. ERA5 variables include 6 upper-air variables in 37 levels and 21 surface variables. 6 upper-air variables include ‘geopotential’, ‘temperature’, ‘u component of wind’, ‘v component of wind’, ‘specific humidity’ and ‘vertical velocity’. Please refer to the supplementary materials for more details. The time resolution of these two datasets are 1 hour and 15 minutes, respectively.

Fig. 6: The 24 lead times’ quantitative comparison results on ‘USA’ dataset. 0.2, 1, 2, 4 and 8mm/hour represent different Cumulative precipitation rate thresholds.

Method	POD			CSI			HSS
	0.2	2	8	0.2	2	8	0.2	2	8
ConvGRU	0.007	0.000	0.000	0.006	0.000	0.000	-0.001	0.000	0.000
SimVP	0.218	0.153	0.035	0.181	0.119	0.029	0.254	0.184	0.052
SimVP_HTA	0.099	0.047	0.015	0.077	0.032	0.009	0.110	0.048	0.015
Metnet2	0.183	0.149	0.039	0.162	0.114	0.029	0.237	0.180	0.051
EarthFarseer	0.369	0.046	0.000	0.073	0.017	0.000	0.032	0.013	0.001

TABLE III: The 24 lead times’ averaged quantitative comparison results on ‘USA’ dataset. The values 0.2, 2, and 8mm/h specify the precipitation rate thresholds used for calculating the metric indices, with the thresholds of 1 and 4 omitted from presentation here.

Fig. 7: The qualitative comparison on ‘USA’ dataset. The input data covers the period: 2016-10-26 03:00 - 2016-10-26-07:00 (UTC). The results of EarthFarseer are not shown here due to its poor qualitative performance.

Fig. 8: The qualitative comparison on ‘USA’ dataset. The input data covers the period: 2016-11-28 12:00 - 16:00 (UTC).

Fig. 9: The qualitative comparison on ‘USA’ dataset. The input data covers the period: 2016-11-29 16:00 - 20:00 (UTC).

Method	POD			CSI			HSS
	20	25	30	20	25	30	20	25	30
ConvGRU	0.274	0.175	0.017	0.210	0.142	0.016	0.320	0.226	0.025
SimVP	0.323	0.252	0.106	0.195	0.152	0.074	0.294	0.240	0.126
SimVP_HTA	0.254	0.188	0.050	0.191	0.142	0.042	0.292	0.227	0.071
Metnet2	0.080	0.075	0.046	0.077	0.070	0.042	0.130	0.121	0.077
Ours	0.410	0.387	0.281	0.216	0.176	0.114	0.325	0.276	0.191

TABLE IV: The 24 lead times’ averaged quantitative comparison results on ‘Hubei’ dataset. 20, 25, 30 dBZ represent the calculation of metric indices under different reflectivity thresholds. The results of EarthFarseer are not shown here due to its poor performance in this dataset.

Fig. 10: The 24 lead times’ quantitative comparison results on ‘Hubei’ dataset. The values 20, 25, and 30 correspond to different radar reflectivity thresholds.

Fig. 11: The qualitative comparison on ‘Hubei’ dataset. The input data covers the period: 2022-07-27 15:00 - 16:00 (UTC).

Fig. 12: The qualitative comparison on ‘Hubei’ dataset. The input data covers the period: 2022-06-29 05:30 - 06:30 (UTC).

Fig. 13: The qualitative comparison on ‘Hubei’ dataset. The input data covers the period: 2023-08-27 08:45 - 09:45 (UTC).

Fig. 14: The quantitative ablation comparison results of loss functions on ‘USA’ dataset. Note that ‘Weighted_MAE_Loss’ refers to the MAE loss weighted by the parameters defined in Eq. 17 (The weighting schemes are identical across different ablation studies). The term ‘Weighted_MAE+Original_CE_Loss’ denotes a combination of the weighted MAE loss (with non-precipitation samples masked out, as in Eq. 17 ) and the unweighted cross-entropy loss. The focal loss is adopted from [ 40 ] .

Fig. 15: The quantitative ablation comparison results between ‘Weighted MAE+Focal’ loss and ‘WMCE’ loss on ‘USA’ dataset across different epochs.

Fig. 16: The ablation results of latent space iterative prediction and inputting more variables (ERA5) on ‘USA’ dataset. ‘Physical_Space_Iter_WO_ERA5’ denotes the iterative prediction conducted in the physical variable space, employing only MRMS QPE and radar data while excluding ERA5 variables.‘Latent_Space_Iter_With_ERA5’ means perform the iterative prediction in latent variable space with inputting more variables (including MRMS QPE, radar and ERA5 variables).

Fig. 17: The ablation results of ‘Importance Sampling’ on ‘USA’ dataset. ‘W_Importance_Sampling’ refers to the incorporation of importance sampling during the training phase to preprocess the training data by filtering out a portion of non-precipitation sequences. ‘WO_Importance_Sampling’, on the other hand, does not employ this technique and uses all available data for training.

Fig. 18: Attribution magnitude of input variables on ‘USA’ dataset. The dots and lines in different colors represent different prediction intervals. Please refer to the supplementary materials for more information on ERA5 surface variables.

Methods	Para	T-time	I-time
	(M)	(h)	(s)
ConvGRU	0.045	30.556	2.245
SimVP	5.827	10.167	1.286
SimVP_HTA	1.485	18.389	2.6515
Metnet2	13.246	20.389	5.115
EarthFarseer	103.576	86.667	5.318

TABLE V: Average computational complexity on ‘USA’ dataset. Para

\rightarrow

Parameters. T-time

\rightarrow

Training time. I-time

\rightarrow

Inference time.

Explore similar work

Improving global precipitation forecasts with an AI weather model trained on satellite observations
Sep 2, 2026Julian F. Schmitt, Bertrand Delorme, Robert C. King +4Precipitation Forecasting Probabilistic Forecasting
X, The Moonshot Factory, Google LLC, Mountain View, CA, USA · Google, Google LLC, Mountain View, CA, USA · Environmental Science and Engineering, California Institute of Technology, Pasadena, CA, USA +1
Diffusion Precipitation Generation with Medium-range Weather Forecasting
Oct 9, 2026Guowen Li, Yang Liu, Mengxuan Chen +11Precipitation Forecasting Weather Forecasting
Tsinghua University · The Chinese University of Hong Kong · The Hong Kong University of S & T +5
Dense-Cast: A lightweight ensemble of deep learning architectures for precipitation nowcasting
Aug 6, 2026Gourav Jyoti Kalita, Hidam Kumarjit SinghPrecipitation Forecasting Deep Ensembles
Department of Electronics and Communication Technology, Gauhati University, Guwahati-781014, Assam, India

Sep 2, 2026J/K move · Enter open · S save

Improving global precipitation forecasts with an AI weather model trained on satellite observations

Julian F. Schmitt, Bertrand Delorme, Robert C. King, Yashica Patodia+3

X, The Moonshot Factory, Google LLC, Mountain View, CA, USA · Google, Google LLC, Mountain View, CA, USA · Environmental Science and Engineering, California Institute of Technology, Pasadena, CA, USA+1

Precipitation Forecasting Probabilistic Forecasting Weather Forecasting

Open paper AskPDF arXiv

Precipitation forecasts shape decision-making across the global economy, particularly in sectors such as agriculture. However, unlike variables such as temperature, precipitation is highly intermittent and localized, making it difficult to forecast. While recent advances in AI weather prediction systems have enabled them to surpass physical models on globally averaged metrics, improvements in mean error rarely translate to actionable forecasts of severe flooding or dry crop fields. Furthermore, most of these models are trained and evaluated against a reanalysis data product, ERA5, which has well-known biases. Here we retrain AIFS, ECMWF's widely-used, open-source operational 0.25° probabilistic graph-transformer weather model, on satellite-based precipitation observations. Our model, Laxmi, improves global probabilistic accuracy by 19% and resolves systematic distributional biases in ERA5. Specifically, Laxmi reduces drizzle overprediction by 33% for amounts less than 3 mm per day. It also mitigates extreme rainfall underprediction, improving the global 95th percentile Brier skill score by 57%. Across a case study of 10 Indian tropical storms, Laxmi delivered the most accurate forecast of 150 mm event-total precipitation in 7 events, compared to 1 for AIFS and 2 for the leading physical model, IFS. Our results demonstrate that incorporating observation-based precipitation data directly into training can substantially improve forecasts.

Abstract

Explore similar work

Improving global precipitation forecasts with an AI weather model trained on satellite observations

Diffusion Precipitation Generation with Medium-range Weather Forecasting

Dense-Cast: A lightweight ensemble of deep learning architectures for precipitation nowcasting