arxiv: 2604.15377 · v1 · submitted 2026-04-15 · 💻 cs.LG · cs.CV· cs.MM

Recognition: unknown

M3R: Localized Rainfall Nowcasting with Meteorology-Informed MultiModal Attention

Sanjeev Panta , Rhett M Morvant , Xu Yuan , Li Chen , Nian-Feng Tzeng

Authors on Pith no claims yet

Pith reviewed 2026-05-10 13:06 UTC · model grok-4.3

classification 💻 cs.LG cs.CVcs.MM

keywords rainfall nowcastingmultimodal attentionradar imageryweather stationsprecipitation predictiondeep learningmeteorological data fusion

0 comments

The pith

M3R uses weather station time series as queries to attend to radar features for improved local rainfall nowcasting.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents M3R as a new architecture that aligns NEXRAD radar imagery with personal weather station measurements and applies specialized multimodal attention to predict rainfall directly. By treating station time series as queries, the model selectively focuses on spatial precipitation patterns in the radar data. A sympathetic reader would care because accurate short-term rainfall forecasts support better disaster response and water management in specific locales. The experiments on three 100 km by 100 km areas around radar stations show gains in accuracy, efficiency, and detection over prior methods.

Core claim

M3R is a Meteorology-informed MultiModal attention-based architecture for direct Rainfall prediction that processes temporally aligned visual NEXRAD radar imagery together with numerical Personal Weather Station measurements, using the station time series as queries within the attention layers to extract focused precipitation signatures from the radar spatial features.

What carries the argument

The multimodal attention mechanism that uses weather station time series as queries to selectively attend to and extract features from aligned radar imagery.

If this is right

Higher accuracy and faster inference for rainfall nowcasts in operational settings around radar stations.
Improved ability to detect precipitation events compared with existing single-modality or less integrated models.
A reusable pipeline for aligning heterogeneous meteorological data sources in future multimodal weather models.
New benchmark numbers for multimedia rainfall prediction on the tested spatial scales.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The query-from-station design could be adapted to incorporate additional sensor streams such as satellite or ground camera data.
If the attention remains stable, the approach might support shorter forecast horizons or finer spatial grids without proportional increases in compute.
Similar attention patterns might help other environmental prediction tasks where sparse point measurements can guide dense image or grid data.

Load-bearing premise

The specialized attention can reliably pull precipitation signals from the combined radar and station data without overfitting to the three tested radar-station locations.

What would settle it

Performance measurements on radar and station data from a fourth geographic region outside the original three 100 km areas would show whether the reported accuracy and detection gains generalize or remain tied to the training regions.

Figures

Figures reproduced from arXiv: 2604.15377 by Li Chen, Nian-Feng Tzeng, Rhett M Morvant, Sanjeev Panta, Xu Yuan.

**Figure 2.** Figure 2: Illustration of Uniform Frame Sampling for Visual Embedding. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: Illustration of M3R Model with Vision Encoder, TS Encoder, [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: Comparison: (a) Zero (b) High-intensity and (c) Medium-intensity precipitation showing M3R’s consistent performance across all scenarios for LA. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

**Figure 1.** Figure 1: No Rain Events, Lake Charles (LA) Precipitation Data Treatment: Contextual interpolation identifies active precipitation periods using rolling window analysis: Pactive(t) = ( 1 if R t+τ/2 t−τ/2 1[R(t ′ ) > 0]dt′ > 0 0 otherwise (9) where τ = 2.5 hours and 1[·] is an indicator function. 2) Data Validation and Quality Assurance: Physical constraint enforcement includes: Tmax ≥ Tavg ≥ Tmin (10) RHmax ≥ RHavg… view at source ↗

**Figure 2.** Figure 2: Light Rain Events, Lake Charles (LA) where k = 15 minutes. Step 4: Sequence Validation. Apply cumulative significance criterion: Σi = X 3 j=−4 Z¯(ti + jk) (16) Step 5: Temporal Advancement. Advance search window by 4 time steps: tnext = ti + 4k (17) 2) Temporal Synchronization: Optimal PWS timestamp matching uses minimum distance criterion: t ∗ PW S = arg min tPW S |tradar − tPW S| (18) with search constr… view at source ↗

**Figure 3.** Figure 3: Heavy Rain Events, Lake Charles (LA) aligned reflectivity-meteorological data: Lake Charles, Louisiana ( [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

**Figure 4.** Figure 4: Montgomery, Alabama [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗

**Figure 5.** Figure 5: Jackson, Mississippi [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗

read the original abstract

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

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

M3R presents a targeted multimodal attention design for rainfall nowcasting but its evaluation on only three regions limits confidence in the claimed gains.

read the letter

M3R's core idea is to use time series from personal weather stations as queries in a multimodal attention setup that pulls relevant features from NEXRAD radar imagery for short-term rainfall prediction. This query choice is a concrete design decision that fits the localized nature of the task and hasn't been reported in exactly this form for rainfall nowcasting. The paper does a solid job describing the full pipeline for aligning the radar images with station measurements over time, and releasing the code on GitHub is a real plus for anyone who wants to build on it or check the implementation. What the work does well is show how to handle the heterogeneous data sources in a practical way for operational use. The soft spots are around the evaluation. All experiments are limited to three 100 km by 100 km areas around specific NEXRAD stations. Without testing on additional regions or using spatial hold-out methods, it's possible the attention mechanism is capturing patterns tied to those particular sites rather than general precipitation dynamics. The abstract mentions outperformance and efficiency gains but provides no actual metrics, error bars, or ablation studies, which makes it difficult to gauge how much better it really is compared to existing methods. The central claim would be stronger with more diverse test locations and quantitative details in the main text. This paper is for applied machine learning researchers and meteorologists working on nowcasting systems. Someone looking for new ways to fuse radar and ground observations would find the architecture description useful, even if they need to run their own experiments to verify the gains. I think it deserves peer review because the multimodal attention formulation is clearly described and the code is public, giving referees something concrete to evaluate. The authors should be asked to expand the experimental section with more validation.

Referee Report

2 major / 1 minor

Summary. The paper introduces M3R, a multimodal attention architecture for rainfall nowcasting that fuses NEXRAD radar imagery with aligned Personal Weather Station (PWS) time series. Station time series are used as queries in specialized attention layers to selectively extract precipitation features from radar imagery. A temporal alignment pipeline is described for heterogeneous data. Experiments are reported on three 100 km × 100 km regions centered at NEXRAD stations, with the claim that M3R outperforms prior approaches in accuracy, efficiency, and precipitation detection. Source code is released.

Significance. If the performance gains are confirmed with rigorous metrics and broader validation, the work would provide a concrete demonstration of query-driven multimodal fusion for localized nowcasting, potentially improving operational precipitation prediction by better exploiting sparse station data alongside radar. The public code release supports reproducibility and extension.

major comments (2)

[Experimental Results] Experimental section: Evaluation is confined to three fixed 100 km × 100 km tiles centered on specific NEXRAD stations. No spatial cross-validation, hold-out on additional stations or regions, or testing for topographic/station-density biases is described, leaving open whether the attention mechanism learns transferable multimodal features or region-specific artifacts.
[Abstract] Abstract and results: The central claim of 'substantial improvements in accuracy, efficiency, and precipitation detection' is stated without any quantitative metrics, baseline names, error bars, ablation results, or statistical significance tests in the provided text, preventing verification of the outperformance assertion.

minor comments (1)

[Abstract] Abstract uses 'multimedia' where 'multimodal' would be more precise and consistent with the title and technical description.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below in detail and have revised the paper to improve clarity and rigor where appropriate.

read point-by-point responses

Referee: Experimental section: Evaluation is confined to three fixed 100 km × 100 km tiles centered on specific NEXRAD stations. No spatial cross-validation, hold-out on additional stations or regions, or testing for topographic/station-density biases is described, leaving open whether the attention mechanism learns transferable multimodal features or region-specific artifacts.

Authors: We appreciate this observation regarding the scope of our evaluation. The three 100 km × 100 km regions were deliberately chosen to span diverse meteorological regimes and varying station densities, as detailed in Section 4.1 of the manuscript. However, we acknowledge that the absence of explicit spatial cross-validation or hold-out testing on additional regions leaves open questions about the transferability of the learned multimodal attention features versus potential region-specific patterns. In the revised manuscript, we will expand the discussion in the experimental section to address potential topographic and station-density biases, include a limitations paragraph on generalizability, and add supplementary experiments with hold-out stations from the same NEXRAD coverage areas where aligned data is available. We believe these additions will strengthen the evidence for the attention mechanism's utility while remaining honest about the current data constraints. revision: partial
Referee: Abstract and results: The central claim of 'substantial improvements in accuracy, efficiency, and precipitation detection' is stated without any quantitative metrics, baseline names, error bars, ablation results, or statistical significance tests in the provided text, preventing verification of the outperformance assertion.

Authors: We agree that the abstract would be more informative and verifiable with concrete quantitative support for the performance claims. The main text (Sections 4.2–4.4 and Tables 1–3) already contains the full metrics, including CSI, RMSE, and F1 scores with error bars, comparisons against named baselines (e.g., PredRNN, ConvLSTM, and radar-only variants), ablation studies on the multimodal attention components, and statistical significance tests. In the revised manuscript, we will update the abstract to explicitly reference these key quantitative results (e.g., relative improvements in CSI and computational efficiency) and direct readers to the corresponding tables and figures. This change ensures the abstract stands alone while accurately summarizing the evidence presented in the body of the paper. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical model design and evaluation

full rationale

The paper introduces an empirical multimodal attention architecture (M3R) for rainfall nowcasting and reports performance gains on three fixed 100 km × 100 km NEXRAD-centered regions. No mathematical derivation, first-principles result, or prediction step is claimed; the contribution consists of model design, temporal alignment pipeline, and experimental benchmarks. No equations, fitted parameters renamed as predictions, self-citation load-bearing theorems, or ansatz smuggling appear in the provided text. The central claims rest on standard held-out test metrics rather than any reduction to inputs by construction, satisfying the default expectation of no significant circularity.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Review performed on abstract only; the model rests on standard deep-learning assumptions plus the domain claim that heterogeneous meteorological data can be temporally aligned without loss of predictive signal.

free parameters (1)

multimodal attention hyperparameters
Typical learnable weights and scaling factors in the attention layers that are fitted during training.

axioms (1)

domain assumption Heterogeneous meteorological data streams can be temporally aligned accurately enough to support joint learning
Invoked by the described pipeline for temporal alignment of radar imagery and PWS measurements.

pith-pipeline@v0.9.0 · 5496 in / 1278 out tokens · 43207 ms · 2026-05-10T13:06:57.799717+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

41 extracted references · 5 canonical work pages · 3 internal anchors

[1]

Prediff: Precipitation nowcasting with latent diffusion models,

Zhihan Gao, Xingjian Shi, Boran Han, Hao Wang, Xiaoyong Jin, Danielle Maddix, Yi Zhu, Mu Li, and Yuyang Bernie Wang, “Prediff: Precipitation nowcasting with latent diffusion models,”Advances in Neural Information Processing Systems, vol. 36, pp. 78621–78656, 2023

2023
[2]

Diffcast: A unified framework via residual diffusion for precipitation nowcasting,

Demin Yu, Xutao Li, Yunming Ye, Baoquan Zhang, Chuyao Luo, Kuai Dai, Rui Wang, and Xunlai Chen, “Diffcast: A unified framework via residual diffusion for precipitation nowcasting,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 27758–27767

2024
[3]

Comprehensive transformer-based model architecture for real-world storm prediction,

Fudong Lin, Xu Yuan, Yihe Zhang, Purushottam Sigdel, Li Chen, Lu Peng, and Nian-Feng Tzeng, “Comprehensive transformer-based model architecture for real-world storm prediction,” inProceedings of Joint European Conference on Machine Learning and Knowledge Discovery in Databases, 2023, pp. 54–71

2023
[4]

Convolutional lstm network: A machine learning approach for precipitation nowcasting,

Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo, “Convolutional lstm network: A machine learning approach for precipitation nowcasting,”Advances in neural information processing systems, vol. 28, 2015

2015
[5]

Deep learning for precipitation nowcasting: A benchmark and a new model,

Xingjian Shi, Zhihan Gao, Leonard Lausen, Hao Wang, Dit-Yan Yeung, Wai-kin Wong, and Wang-chun Woo, “Deep learning for precipitation nowcasting: A benchmark and a new model,”Advances in neural information processing systems, vol. 30, 2017

2017
[6]

Machine learning for precipitation nowcast- ing from radar images,

Shreya Agrawal, Luke Barrington, Carla Bromberg, John Burge, Cenk Gazen, and Jason Hickey, “Machine learning for precipitation nowcast- ing from radar images,”arXiv preprint arXiv:1912.12132, 2019

work page arXiv 1912
[7]

Sevir: A storm event imagery dataset for deep learning applications in radar and satellite meteorology,

Mark Veillette, Siddharth Samsi, and Chris Mattioli, “Sevir: A storm event imagery dataset for deep learning applications in radar and satellite meteorology,”Advances in Neural Information Processing Systems, vol. 33, pp. 22009–22019, 2020

2020
[8]

Are transformers effective for time series forecasting?,

Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu, “Are transformers effective for time series forecasting?,” inProceedings of the AAAI conference on artificial intelligence, 2023, vol. 37, pp. 11121–11128

2023
[9]

A Time Series is Worth 64 Words: Long-term Forecasting with Transformers

Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam, “A time series is worth 64 words: Long-term forecasting with transformers,”arXiv preprint arXiv:2211.14730, 2022

work page internal anchor Pith review arXiv 2022
[10]

Revisiting the seasonal trend decomposition for enhanced time series forecasting,

Sanjeev Panta, Xu Yuan, Li Chen, and Nian-Feng Tzeng, “Revisiting the seasonal trend decomposition for enhanced time series forecasting,” arXiv preprint arXiv:2602.18465, 2026

work page arXiv 2026
[11]

iTransformer: Inverted Transformers Are Effective for Time Series Forecasting

Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long, “itransformer: Inverted transformers are effective for time series forecasting,”arXiv preprint arXiv:2310.06625, 2023

work page internal anchor Pith review arXiv 2023
[12]

Regional weather variable predictions by machine learning with near-surface observational and atmospheric numerical data,

Yihe Zhang, Bryce Turney, Purushottam Sigdel, Xu Yuan, Eric Rap- pin, Adrian L Lago, et al., “Regional weather variable predictions by machine learning with near-surface observational and atmospheric numerical data,”IEEE Transactions on Geoscience and Remote Sensing, 2025

2025
[13]

Precise weather parameter predictions for target regions via neural networks,

Yihe Zhang, Xu Yuan, Sytske K Kimball, Eric Rappin, Li Chen, Paul Darby, Tom Johnsten, Lu Peng, Boisy Pitre, David Bourrie, and Nian- Feng Tzeng, “Precise weather parameter predictions for target regions via neural networks,” inProceedings of European Conference on Machine Learning and Principles and Practices of Knowledge Discovery in Databases (ECML-PKD...

2021
[14]

Earthformer: Exploring space-time trans- formers for earth system forecasting,

Zhihan Gao, Xingjian Shi, Hao Wang, Yi Zhu, Yuyang Bernie Wang, Mu Li, and Dit-Yan Yeung, “Earthformer: Exploring space-time trans- formers for earth system forecasting,”Advances in Neural Information Processing Systems, vol. 35, pp. 25390–25403, 2022

2022
[15]

Alphapre: Amplitude- phase disentanglement model for precipitation nowcasting,

Kenghong Lin, Baoquan Zhang, Demin Yu, Wenzhi Feng, Shidong Chen, Feifan Gao, Xutao Li, and Yunming Ye, “Alphapre: Amplitude- phase disentanglement model for precipitation nowcasting,” inProceed- ings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 17841–17850

2025
[16]

Mmst-vit: Climate change-aware crop yield prediction via multi-modal spatial-temporal vision transformer,

Fudong Lin, Summer Crawford, Kaleb Guillot, Yihe Zhang, Yan Chen, Xu Yuan, Li Chen, et al., “Mmst-vit: Climate change-aware crop yield prediction via multi-modal spatial-temporal vision transformer,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 5774–5784

2023
[17]

An open and large-scale dataset for multi-modal climate change-aware crop yield predictions,

Fudong Lin, Kaleb Guillot, Summer Crawford, Yihe Zhang, Xu Yuan, and Nian-Feng Tzeng, “An open and large-scale dataset for multi-modal climate change-aware crop yield predictions,” inProceedings of the ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), 2024, pp. 5375–5386

2024
[18]

Mm-rnn: A multimodal rnn for precipitation nowcasting,

Zhifeng Ma, Hao Zhang, and Jie Liu, “Mm-rnn: A multimodal rnn for precipitation nowcasting,”IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1–14, 2023

2023
[19]

Fsrgan: A satellite and radar-based fusion prediction network for precipitation nowcasting,

Dan Niu, Yinghao Li, Hongbin Wang, Zengliang Zang, Mingbo Jiang, Xunlai Chen, and Qunbo Huang, “Fsrgan: A satellite and radar-based fusion prediction network for precipitation nowcasting,”IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 17, pp. 7002–7013, 2024

2024
[20]

Meteo-dlnet: quantitative precipitation nowcasting net based on meteorological features and deep learning,

Jianping Hu, Bo Yin, and Chaoqun Guo, “Meteo-dlnet: quantitative precipitation nowcasting net based on meteorological features and deep learning,”Remote Sensing, vol. 16, no. 6, pp. 1063, 2024

2024
[21]

A review of nexrad level ii: Data, distribution, and applications,

Matthew Huber and Jeff Trapp, “A review of nexrad level ii: Data, distribution, and applications,”Journal of Terrestrial Observation, vol. 1, no. 2, pp. 4, 2009

2009
[22]

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

Dosovitskiy Alexey, “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv preprint arXiv: 2010.11929, 2020

work page internal anchor Pith review Pith/arXiv arXiv 2010
[23]

Vivit: A video vision transformer,

Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Luˇci´c, and Cordelia Schmid, “Vivit: A video vision transformer,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 6836–6846

2021
[24]

Attention is all you need,

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin, “Attention is all you need,”Advances in neural information processing systems, vol. 30, 2017. M3R: Localized Rainfall Nowcasting with Meteorology-Informed MultiModal Attention Supplementary Material Sanjeev Panta⋆, Rhett M Morvant ⋆...

2017
[25]

Files containing maintenance data mes- sages (identified by ”MDM” suffixes) are excluded to ensure data quality

Data Acquisition and Format Conversion:NEXRAD Level-2 radar data are systematically downloaded from the NOAA Big Data Program’s Amazon S3 repository using automated retrieval protocols targeting the KLCH, KMXX and KDGX radar stations. Files containing maintenance data mes- sages (identified by ”MDM” suffixes) are excluded to ensure data quality. Raw files...
[26]

Coordinate Transformation and Spatial Extraction: Polar coordinate radar data are transformed to Cartesian grids using LROSE Radx2Grid with Lambert Conformal Conic projection at 1 km × 1 km resolution. The region of interest extraction algorithm identifies the optimal grid point using Euclidean distance minimization: dmin = min i,j q (loni,j −lon target)2...
[27]

Composite Reflectivity Generation:Column-maximum composite reflectivity fields are computed using the four low- est elevation angles to minimize beam blockage and ground clutter: Zc(x, y) = 4 max k=1 Zk(x, y)(2) whereZ k represents reflectivity at elevation anglek
[28]

Corresponding author: Dr

Temporal Interpolation:Irregular radar observation times are interpolated to regular 15-minute intervals using piecewise linear interpolation: The research is supported in part by the NSF under grants OIA- 2327452, OIA-2019511, and 2425812, in part by the Louisiana BoR under LEQSF(2024-27)-RD-B-03. Corresponding author: Dr. Li Chen (li.chen@louisiana.edu)...

2024
[29]

The cubic spline S(t)for variableVsatisfies: S(ti) =V i andS ′(t− i ) =S ′(t+ i ), S′′(t− i ) =S ′′(t+ i )(4) ensuring continuity in first and second derivatives

Advanced Gap Filling Methodology:Continuous Me- teorological Variables:Temperature, humidity, dewpoint, and pressure utilize cubic spline interpolation. The cubic spline S(t)for variableVsatisfies: S(ti) =V i andS ′(t− i ) =S ′(t+ i ), S′′(t− i ) =S ′′(t+ i )(4) ensuring continuity in first and second derivatives. Wind Vector Processing:Wind direction and...
[30]

Multi-Modal Alignment - Detailed Algorithm

Data Validation and Quality Assurance:Physical con- straint enforcement includes: Tmax ≥T avg ≥T min (10) RHmax ≥RH avg ≥RH min (11) Vgust ≥V wind (12) C. Multi-Modal Alignment - Detailed Algorithm
[31]

Weather Event Selection Algorithm:Step 1: Temporal Aggregation.For each radar observation timet i, compute spatial mean reflectivity: ¯Z(t i) = 1 Nx ·N y NxX x=1 NyX y=1 Z(x, y, ti)(13) whereN x =N y = 100. Step 2: Significance Classification.Apply meteorological significance thresholdZ threshold = 3.0dBZ: S(ti) = ( 1if ¯Z(t i)> Z threshold 0otherwise (14...
[32]

Temporal Synchronization:Optimal PWS timestamp matching uses minimum distance criterion: t∗ P W S = arg min tP W S |tradar −t P W S|(18) with search constrained to ±7.5 minutes
[33]

Reflectivity Quantization Scheme:Meteorologically- informed quantization function: Q(Z) =    0ifZ <8 8if8≤Z <16 16if16≤Z <20 ⌊Z⌋if20≤Z <70 70ifZ≥70 255ifZis missing (19)
[34]

Dataset Partitioning:Chronological partitioning func- tion: P(i) = ( Train ifi <⌊0.85·N total⌋ Test otherwise (20) whereN total represents total valid sequences. D. Final Dataset Statistics The complete processing pipeline is used to generate datasets for three different locations with 96,359 instances of Fig. 3. Heavy Rain Events, Lake Charles (LA) align...
[35]

At LA, M3R dominates with best performance across most metrics (RMSE: 2.87, R²: 0.29, CC: 0.54)

Consistent Excellence Across Stations:M3R demon- stratesremarkable consistency, achieving first or second- best RMSE and best MAE at all three stations. At LA, M3R dominates with best performance across most metrics (RMSE: 2.87, R²: 0.29, CC: 0.54). At AL and MS, M3R maintains competitive RMSE while achieving best MAE (0.36) at both stations
[36]

Station-Specific Strengths: •LA station: M3R excels across all metrics with 7% RMSE improvement over AlphaPre, 3.6× R² improve- ment over iTransformer, and exceptional detection (CSI 0.1: 0.410, CSI 10: 0.236) •AL station: M3R achieves best MAE (0.36) and strongest light precipitation detection among M3R stations (CSI 0.1: 0.300), with competitive RMSE (3...
[37]

Geographical Robustness:The consistent performance improvements across three geographically diverse stations validate M3R’s ability to generalize across different meteoro- logical conditions and precipitation patterns, unlike baselines that show high variability (e.g., AlphaPre’s RMSE ranges from 2.94 to 3.35). E. Efficiency & Deployment Analysis Training...
[38]

Multi-Modal vs. Single-Modal Advantage:The sub- stantial performance gap between our M3R model and both time series baselines and precipitation-specific methods vali- dates thatmulti-modal learning captures complementary informationunavailable to single-modal approaches. The 21-39% RMSE improvements over Diffcast-SimVP across stations demonstrate clear su...
[39]

However, our M3R model’s 1.8-3.6× improvement in R² across stations indicatessignificantly enhanced pattern recognitioncapability through effective spatial-temporal inte- gration

Pattern Recognition Superiority:The relatively low R² values across baseline methods (highest: 0.16 for AlphaPre at AL) reflect theinherent difficulty of precipitation pre- diction. However, our M3R model’s 1.8-3.6× improvement in R² across stations indicatessignificantly enhanced pattern recognitioncapability through effective spatial-temporal inte- gration
[40]

Early Warning System Effectiveness:Our model excels in light precipitation detection across all stations (CSI 0.1: 0.300-0.414), showing3.3-5.5× improvement over Diffcast- SimVPand21-173% improvement over AlphaPre, which is critical for operational early warning systems and emergency response applications
[41]

Multi-Scale Precipitation Handling:M3R maintains superior or competitive performance across different precipita- tion intensities at all stations, demonstratingrobust handling of the complete precipitation spectrumfrom light drizzle (average CSI 0.1: 0.375) to heavy rainfall events (average CSI 10: 0.210). G. Methodological Contributions Direct Precipitat...