Pith. sign in

REVIEW 4 major objections 6 minor 55 references

From Swath to Full-Disc: Advancing Precipitation Retrieval with Multimodal Knowledge Expansion

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper argues that infrared-only satellite imagery, after two stages of knowledge distillation from multimodal swath data, can retrieve full-disc precipitation at 4 km resolution and 5-minute cadence with accuracy comparable to the…

desk verdict Genuinely novel distillation pipeline whose full-disc vs IMERG claim needs an unfiltered evaluation before it can be believed. read the letter →

arxiv 2506.07050 v1 pith:H5N4VRUG submitted 2025-06-08 cs.CV cs.IRcs.MM

Zheng Wang , Kai Ying , Bin Xu , Chunjiao Wang , Cong Bai This is my paper
classification cs.CVcs.IRcs.MM
keywords PrecipitationRetrievalExpansioninfrared-onlyestimationknowledgedistillationmaskedimagemodelingwaveletenhancementincompletemultimodallearningfull-discsatellitenear-real-time
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper introduces Precipitation Retrieval Expansion (PRE), a framing in which infrared imagery, the only modality with full-disc coverage, is turned into precipitation estimates even where passive microwave and radar data are absent. It argues that a two-stage pipeline, first distilling knowledge from a multimodal teacher trained on swath-aligned infrared, microwave, and radar reflectivity into an infrared-only student, then adapting the student to the full disc with a guided masking fine-tune, lets a compact 48M-parameter network retrieve precipitation at 4 km and 5-minute cadence. On the authors' benchmark over East Asia, the model reports RMSE 0.8714 mm/h and CSI 0.3912 on full-disc retrieval, surpassing PERSIANN-CCS and PDIR and matching or exceeding IMERG on most reported metrics while running on infrared alone. If correct, this would mean near-real-time precipitation monitoring can be done from geostationary infrared only, at higher spatial and temporal resolution and far lower latency than current merged satellite products.

What carries the argument

The load-bearing mechanism is the Coordinated Masking and Wavelet Enhancement (CoMWE) module placed in each encoder block. It combines Re-Masked Knowledge Distillation (RMKD), which applies an initial low-ratio mask, shuffles the surviving features, splits them into $N$ parts, and re-masks each part at the high ratio $(n+\alpha-1)/n$ before reconstruction through an auto-encoder, reconciling the low-masking optimum of masked knowledge distillation with the high-masking reconstruction strength of masked autoencoders. It also combines Detail-Aware Wavelet Enhancement (DAWE), which takes a Haar discrete wavelet transform of the infrared embedding, concatenates the $HL$, $LH$, and $HH$ sub-bands, aligns them with a $1\times1$ convolution, and applies spatial attention to form high-frequency prompt features. In the second stage, Low-Rank Adaptation (LoRA) confines updates to the auto-encoder weights while convolutional layers stay frozen, and Self-MaskTune sets a threshold $ ho$ on the largest previous-epoch task losses to generate the mask $M_{AE}(y_{f,t-1})$ that decides which regions are re-learned.

What would settle it

Recompute the full-disc comparison on all 2023 patches without excluding low- or no-precipitation patches, remap IMERG to the same 4 km grid, and evaluate on all pixels; if the FAR and CSI advantage over IMERG shrinks or reverses, the exclusion was the load-bearing cause.

Watch

Extended reading notes

Core claim

The central claim is that precipitation retrieval beyond the scanning swath can be treated as an incomplete-multimodal-learning problem, and that knowledge distillation is the right bridge. PRE-Net first trains a multimodal U-Net teacher on swath-aligned infrared, passive microwave, and radar reflectivity, then trains an infrared-only student to reproduce the teacher's feature maps under coordinated masking while wavelet-enhanced high-frequency detail is injected to preserve precipitation boundaries. In the full-disc stage, the student is fine-tuned with Low-Rank Adaptation applied only to the auto-encoder modules, and Self-MaskTune builds a mask from the previous epoch's prediction errors so that parameter updates concentrate where swath-derived multimodal knowledge conflicts with full-disc infrared patterns. The reported outcome is full-disc precipitation retrieval whose regression and detection scores are close to or better than the multimodal IMERG product while operating natively at 4 km and 5-minute cadence from infrared input alone.

Load-bearing premise

The evaluation drops swaths and patches with little or no precipitation before computing scores, so the reported hit and false-alarm statistics describe a rain-enriched sample rather than the full disc the method claims to retrieve.

Editorial extensions

If this is right

  • Operational near-real-time precipitation monitoring could run from geostationary infrared alone, at 4 km and 5-minute cadence, avoiding the multi-hour latency of merged products like IMERG.
  • If the reported numbers hold, PRE-Net would surpass the infrared-based operational products PERSIANN-CCS and PDIR by substantial margins in detection and critical success metrics on the study region.
  • The multimodal teacher used for distillation outperforms GPM 2B-CMB on swath retrieval, so the same method also offers an improved swath-scale precipitation product.
  • A regional wet-season evaluation over Australia shows the model generalizes beyond its training region, suggesting full-disc or global deployment is within reach.
  • Because PRE-Net needs only infrared input at inference, it can be updated across the full disc every 5 minutes, a cadence no passive-microwave- or radar-based product can match.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the reported metrics come from a test sample that excludes low- and no-precipitation patches, an operational evaluation over every pixel would likely show higher false-alarm rates and lower critical success scores, and the ranking against IMERG could change.
  • The same two-stage recipe should transfer to other geostationary imagers since only infrared input is required, but the paper demonstrates only one sensor and one regional transfer test, so cross-sensor generalization remains untested.
  • A stricter test than the paper's neighbor-metric comparison would be a single-grid, pixel-aligned comparison over all full-disc pixels, with IMERG remapped to the same 4 km grid and no precipitation-based filtering.
  • Self-MaskTune is a generic way to balance old and new knowledge during domain adaptation, so the mechanism could be reused for other swath-to-full-disc problems or for adapting precipitation models across seasons and regions.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The manuscript introduces the Precipitation Retrieval Expansion (PRE) task, in which an infrared-only model is trained to retrieve full-disc precipitation after distilling knowledge from a multimodal teacher that sees PMW/PR/IR within the GPM swath. The proposed PRE-Net uses a two-stage pipeline: Swath-Distilling with Coordinated Masking and Wavelet Enhancement (CoMWE), comprising Re-Masked Knowledge Distillation and Detail-Aware Wavelet Enhancement, followed by Full-Disc Adaptation with LoRA and Self-MaskTune. The authors introduce the Swath-MPR and Full-IPR datasets built from FY-4A/AGRI, GPM/GMI, GPM/DPR, and CLDAS-V2.0, and compare PRE-Net with PERSIANN-CCS, PDIR, and IMERG, claiming superior or comparable full-disc retrieval accuracy.

Significance. The task is practically motivated: accurate IR-only full-disc precipitation at 4-km/5-min resolution would address the latency and coverage limitations of current operational products. The proposed CoMWE and Self-MaskTune modules are well motivated, and the internal ablations in Tables 2-4 show consistent, sometimes large, improvements over the baselines. The release of the benchmark and code would be a useful community resource. However, the headline comparison against operational products rests on an evaluation protocol (Section 4.1.2) that filters out no-precipitation patches and uses CLDAS-V2.0 both as training label and evaluation ground truth; until these issues are addressed, the claim of parity with or superiority to IMERG is not established.

major comments (4)
  1. [§4.1.2, §5.2, Table 5] The headline claim of full-disc retrieval performance comparable to or better than IMERG is based on the Full-IPR test set, from which 'patches with minimal or no precipitation were excluded' (Section 4.1.2). This filtering is load-bearing: POD, FAR, and CSI depend on the prevalence of rain/no-rain pixels, so removing dry patches eliminates a large number of easy true negatives over the full disc and can inflate POD/CSI and deflate FAR relative to an unfiltered evaluation. Since the PRE task defined in Section 3.1 is explicitly full-disc retrieval, the authors should report results on the unfiltered Full-IPR set (or on all pixels of the full disc) and temper the claims accordingly; the citation to NowcastNet [48] does not justify this filtering for a reanalysis-based retrieval benchmark.
  2. [§4.1.2, §5.2, Table 5] PRE-Net is trained and fine-tuned on CLDAS-V2.0 labels, and the same CLDAS-V2.0 is used as the ground truth when computing the metrics in Table 5. PERSIANN-CCS, PDIR, and IMERG are independent products that were not trained on CLDAS. The comparison therefore conflates retrieval skill with the ability to reproduce CLDAS-specific biases. To support the 'outperforms IMERG' claim, the authors should provide an evaluation on independent gauge or radar observations over the study region, akin to the ISD station evaluation in Table 6, and report both filtered and unfiltered comparisons.
  3. [§3.4, §4.2] The Self-MaskTune algorithm has two free parameters, rho (the threshold factor in Eq. 8) and K (the number of warm-up epochs before masking is applied in Eq. 7), but neither value is given in the implementation details (Section 4.2). Without these values, the Full-Disc Adaptation stage cannot be reproduced, and the ablation in Table 4 cannot be independently checked. Please report rho and K, and preferably a sensitivity study over both.
  4. [§4.1.3, §5.2] The CSI-4 and CSI-8 'neighbor' metrics are defined only as 'max pooling with kernel sizes of 4 and 8' (Section 4.1.3). It is not specified whether the pooling is applied to the binary rain/no-rain masks, how the 0.1 mm/hr threshold interacts with pooling, or how the metric is intended to correct for product-resolution differences. The claim in Section 5.2 that IMERG's higher CSI-4 and CSI-8 scores 'stem from spatial smoothing during interpolation' is not supported by the given definition or by any analysis. The exact computation should be stated, and the interpretation should be backed by evidence.
minor comments (6)
  1. [Abstract, §5.2] The abstract states that PRE-Net is 'outperforming leading products like PERSIANN-CCS, PDIR, and IMERG', but Section 5.2 says 'comparable to IMERG' and Table 5 shows PRE-Net is worse than IMERG on CSI-4 and CSI-8; the abstract and conclusion should match the actual claims.
  2. [Table 5 caption, §4.1.2] PERSIANN-CCS, PDIR, and IMERG are described as 'NWP methods', but these are satellite-based retrieval products rather than numerical weather prediction models; please correct the terminology.
  3. [Table 6] The text says the evaluation uses 'station-level ground truth data from ISD', but the table reports region-wide RMSE/CC/POD/FAR/CSI; clarify whether the metrics are computed at station locations, on a common grid, and how many stations are used.
  4. [Figure 2 caption] The caption says 'Two identical PRE-Net models are trained to obtain both classification and regression results'; clarify whether the classification and regression tasks share weights or are trained independently.
  5. [§5.2] The sentence 'While IMERG exhibits higher CSI-4 (0.4176) and CSI-8 (0.4891) scores, this advantage stems from spatial smoothing during interpolation to 4 km resolution' is speculative, as no interpolation analysis or sensitivity test is provided.
  6. [§4.2, Tables 2-7] No information is given about the number of random seeds or variance across runs; given that some reported differences are small (e.g., Table 4 RMSE 0.8714 vs 0.8853), error bars or multiple-seed results would strengthen the claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: held-out CLDAS evaluation with independent comparison products; the shared label source is a standard supervised design, not a definitional loop.

full rationale

The paper's central derivation is the PRE-Net training pipeline, not a closed-form theory. The teacher is trained on multimodal inputs with CLDAS-V2.0 labels; the student is distilled from the teacher and fine-tuned on full-disc IR with the same CLDAS product used as supervision. The final comparison in Table 5 is against external products (PERSIANN-CCS, PDIR, IMERG) on a held-out 2023 test partition of CLDAS. Because the test labels are not used in training, the reported RMSE, CC, POD, FAR, and CSI are genuine generalization measurements rather than fitted values. The use of CLDAS both for training and evaluation is a label-source choice common to supervised retrieval benchmarks; it does not make the predictions equal to the inputs by construction. The only self-citation is reference [35], the CLDAS-V2.0 dataset paper co-authored by Bin Xu, used as data provenance and validated against station observations; it is not invoked as a uniqueness theorem or to forbid alternatives. The exclusion of dry patches (Section 4.1.2) is a potential validity threat to the full-disc claim, but it affects the fairness of the product comparison, not the circularity of the derivation. No equation in the paper reduces to its own inputs.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

All free parameters are hyperparameters; none are fitted to test data by the paper's own account, but the missing rho and K reduce reproducibility. The key assumptions are the reliability of CLDAS as ground truth and the validity of excluding non-precipitating data. No new physical entities are postulated.

free parameters (6)
  • alpha (initial mask ratio in RMKD) = 0.25
    Set in Section 4.2 to balance masking and reconstruction; not derived.
  • n (number of generated masks in RMKD) = 3
    Section 4.2; high mask ratio (n+alpha-1)/n = 0.75 is achieved by construction.
  • lambda (feature distillation loss weight) = 0.2
    Section 4.2, Eq. 6.
  • gamma (feature reconstruction loss weight) = 50
    Section 4.2, Eq. 6.
  • rho (Self-MaskTune threshold parameter) = not specified
    Appears in Eq. 8 as Pmax = rho * max(L_task); value omitted from implementation details.
  • K (number of warm-up epochs for Self-MaskTune) = not specified
    Eq. 7 uses t <= K versus t > K; K is never reported.
assumptions (6)
  • domain assumption CLDAS-V2.0 is a reliable ground truth for precipitation retrieval
    The paper treats CLDAS-V2.0 as ground truth for training and testing. Its own biases are inherited by PRE-Net and may disadvantage untrained products like IMERG. See Section 4.1.2.
  • ad hoc to paper Excluding low-precipitation swaths and patches does not bias evaluation
    Swaths and full-disc patches with minimal or no precipitation are excluded before training and testing, which changes the prevalence of rain pixels used to compute POD, FAR, and CSI. See Section 4.1.2.
  • domain assumption IR has weak correlation with precipitation while PMW and PR have strong correlation
    This motivates the teacher-student transfer from multimodal data to IR-only inputs. See Section 1.
  • domain assumption The multimodal teacher, trained on the same ground truth, is a valid knowledge source
    The student aligns its features to the teacher's features via distillation, so the teacher's quality and its training target directly shape the student. See Section 3.3.
  • ad hoc to paper CSI-4 and CSI-8 neighbor metrics fairly account for product resolution differences
    Max-pooling-based neighbor metrics are introduced to compare PRE-Net at 4 km with products at 0.04 to 0.1 degrees, but the paper does not validate that this weighting is neutral across products. See Section 4.1.3.
  • domain assumption ISD station measurements are an appropriate ground truth for the Australia generalization test
    Regional generalization is evaluated against NOAA ISD station data, but station distribution and representativeness are not analyzed. See Section 5.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Swath to Full-Disc: Advancing Precipitation Retrieval with Multimodal Knowledge Expansion." pith.science (2026). https://pith.science/paper/H5N4VRUG

@misc{pith2026250607050,
  author       = {Pith},
  title        = {Pith review of: From Swath to Full-Disc: Advancing Precipitation Retrieval with Multimodal Knowledge Expansion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H5N4VRUG}},
  note         = {Machine review of arXiv:2506.07050}
}
read the original abstract

Accurate near-real-time precipitation retrieval has been enhanced by satellite-based technologies. However, infrared-based algorithms have low accuracy due to weak relations with surface precipitation, whereas passive microwave and radar-based methods are more accurate but limited in range. This challenge motivates the Precipitation Retrieval Expansion (PRE) task, which aims to enable accurate, infrared-based full-disc precipitation retrievals beyond the scanning swath. We introduce Multimodal Knowledge Expansion, a two-stage pipeline with the proposed PRE-Net model. In the Swath-Distilling stage, PRE-Net transfers knowledge from a multimodal data integration model to an infrared-based model within the scanning swath via Coordinated Masking and Wavelet Enhancement (CoMWE). In the Full-Disc Adaptation stage, Self-MaskTune refines predictions across the full disc by balancing multimodal and full-disc infrared knowledge. Experiments on the introduced PRE benchmark demonstrate that PRE-Net significantly advanced precipitation retrieval performance, outperforming leading products like PERSIANN-CCS, PDIR, and IMERG. The code will be available at https://github.com/Zjut-MultimediaPlus/PRE-Net.

Figures

Figures reproduced from arXiv: 2506.07050 by the authors.

Figure 1
Figure 1. (a) The Precipitation Retrieval Expansion (PRE) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The multimodal knowledge expansion pipeline of PRE-Net. The framework of PRE-Net is detailed in Stage I. Two [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. The illustration of DAWE. It decomposes input im [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Visual comparison on diverse scanning swaths. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: The visual comparison examples on different full-disc precipitation events. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 50 canonical work pages

  1. [48]

    Shuai Wang, Zipei Yan, Daoan Zhang, Haining Wei, Zhongsen Li, and Rui Li

    work page

  2. [1]

    Phillip A Arkin and Bernard N Meisner. 1987. The relationship between large- scale convective rainfall and cold cloud over the western hemisphere during 1982-84.Monthly Weather Review115, 1 (1987), 51–74

    work page 1987

  3. [2]

    Malarvizhi Arulraj and Ana P Barros. 2021. Automatic detection and classification of low-level orographic precipitation processes from space-borne radars using machine learning.Remote Sensing of Environment257 (2021), 112355

    work page 2021

  4. [3]

    Mamoudou B Ba and Arnold Gruber. 2001. GOES multispectral rainfall algorithm (GMSRA).Journal of Applied Meteorology and Climatology40, 8 (2001), 1500– 1514

    work page 2001

  5. [4]

    Peter Bauer, Alan Thorpe, and Gilbert Brunet. 2015. The quiet revolution of numerical weather prediction.Nature525, 7567 (2015), 47–55

    work page 2015

  6. [5]

    Ali Behrangi, Kuo-lin Hsu, Bisher Imam, Soroosh Sorooshian, and Robert J Kuligowski. 2009. Evaluating the utility of multispectral information in de- lineating the areal extent of precipitation.Journal of Hydrometeorology10, 3 (2009), 684–700

    work page 2009

  7. [7]

    Haonan Chen, V Chandrasekar, Haiming Tan, and Robert Cifelli. 2019. Rainfall estimation from ground radar and TRMM precipitation radar using hybrid deep neural networks.Geophysical Research Letters46, 17-18 (2019), 10669–10678

    work page 2019

  8. [8]

    Jiayi Chen and Aidong Zhang. 2020. HGMF: Heterogeneous Graph-based Fusion for Multimodal Data with Incompleteness. InKDD ’20: The 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Virtual Event, CA, USA, August 23-27, 2020, Rajesh Gupta, Yan Liu, Jiliang Tang, and B. Aditya Prakash (Eds.). ACM, 1295–1305

    work page 2020

Show all 55 references
  1. [9]

    Yuanyuan Chen, Yongsheng Pan, Yong Xia, and Yixuan Yuan. 2023. Disentangle First, Then Distill: A Unified Framework for Missing Modality Imputation and Alzheimer’s Disease Diagnosis.IEEE Trans. Medical Imaging42, 12 (2023), 3566–

    2023

  2. [10]

    Reuben Dorent, Samuel Joutard, Marc Modat, Sébastien Ourselin, and Tom Ver- cauteren. 2019. Hetero-Modal Variational Encoder-Decoder for Joint Modal- ity Completion and Segmentation. InMedical Image Computing and Computer Assisted Intervention - MICCAI 2019 - 22nd Internationa...

    2019

  3. [11]

    Vesta Afzali Gorooh, Ata Akbari Asanjan, Phu Nguyen, Kuolin Hsu, and Soroosh Sorooshian. 2022. Deep neural network high SpatioTEmporal resolution pre- cipitation estimation (Deep-STEP) using passive microwave and infrared data. Journal of Hydrometeorology23, 4 (2022), 597–617

    2022

  4. [12]

    Olson, Stephen Joseph Munchak, Sarah Ringerud, Liang Liao, Ziad Haddad, Bartie L

    Mircea Grecu, William S. Olson, Stephen Joseph Munchak, Sarah Ringerud, Liang Liao, Ziad Haddad, Bartie L. Kelley, and Steven F. McLaughlin. 2016. The GPM Combined Algorithm.Journal of Atmospheric and Oceanic Technology33, 10 (2016), 2225 – 2245

    2016

  5. [13]

    Webster Gumindoga, Tom HM Rientjes, Alemseged Tamiru Haile, Hodson Makurira, and Paolo Reggiani. 2019. Performance of bias-correction schemes for CMORPH rainfall estimates in the Zambezi River basin.Hydrology and Earth System Sciences23, 7 (2019), 2915–2938

    2019

  6. [14]

    Mohammad Havaei, Nicolas Guizard, Nicolas Chapados, and Yoshua Bengio

  7. [15]

    Girshick

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross B. Girshick. 2022. Masked Autoencoders Are Scalable Vision Learners. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022. IEEE, 15979–15988

    2022

  8. [16]

    Hinton, Oriol Vinyals, and Jeffrey Dean

    Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. 2015. Distilling the Knowl- edge in a Neural Network.CoRRabs/1503.02531 (2015)

    2015 arXiv

  9. [17]

    Yang Hong, Kuo-Lin Hsu, Soroosh Sorooshian, and Xiaogang Gao. 2004. Pre- cipitation estimation from remotely sensed imagery using an artificial neural network cloud classification system.Journal of Applied Meteorology43, 12 (2004), 1834–1853

    2004

  10. [18]

    Arthur Y Hou, Ramesh K Kakar, Steven Neeck, Ardeshir A Azarbarzin, Christian D Kummerow, Masahiro Kojima, Riko Oki, Kenji Nakamura, and Toshio Iguchi

  11. [19]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-Rank Adaptation of Large Language Models. InThe Tenth International Conference on Learning Representa- tions, ICLR 2022, Virtual Event, April 25-29, 2...

    2022

  12. [20]

    Robert J Joyce, John E Janowiak, Phillip A Arkin, and Pingping Xie. 2004. CMORPH: A method that produces global precipitation estimates from pas- sive microwave and infrared data at high spatial and temporal resolution.Journal of hydrometeorology5, 3 (2004), 487–503

    2004

  13. [21]

    Christian Kummerow, Y Hong, WS Olson, S Yang, RF Adler, J McCollum, R Ferraro, G Petty, Dong-Bin Shin, and TT Wilheit. 2001. The evolution of the Goddard Profiling Algorithm (GPROF) for rainfall estimation from passive microwave sensors.Journal of Applied Meteorology and Clima...

    2001

  14. [22]

    Shanshan Lao, Guanglu Song, Boxiao Liu, Yu Liu, and Yujiu Yang. 2023. Masked Autoencoders Are Stronger Knowledge Distillers. InIEEE/CVF International Con- ference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023. IEEE, 6361–6370

    2023

  15. [23]

    Vadim Lebedev, Vladimir Ivashkin, Irina Rudenko, Alexander Ganshin, Alexander Molchanov, Sergey Ovcharenko, Ruslan Grokhovetskiy, Ivan Bushmarinov, and Dmitry Solomentsev. 2019. Precipitation nowcasting with satellite imagery. InProceedings of the 25th ACM SIGKDD international...

    2019

  16. [24]

    Kangwen Liu, Jieying He, and Haonan Chen. 2022. Precipitation retrieval from Fengyun-3D MWHTS and MWRI data using deep learning.IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing15 (2022), 7619– 7630

    2022

  17. [25]

    Mengmeng Ma, Jian Ren, Long Zhao, Sergey Tulyakov, Cathy Wu, and Xi Peng

  18. [26]

    Phu Nguyen, Mohammed Ombadi, Soroosh Sorooshian, Kuolin Hsu, Amir AghaK- ouchak, Dan Braithwaite, Hamed Ashouri, and Andrea Rose Thorstensen. 2018. The PERSIANN family of global satellite precipitation data: A review and evalua- tion of products.Hydrology and Earth System Scie...

    2018

  19. [27]

    Phu Nguyen, Eric J Shearer, Mohammed Ombadi, Vesta Afzali Gorooh, Kuolin Hsu, Soroosh Sorooshian, William S Logan, and Marty Ralph. 2020. PERSIANN Dynamic Infrared–Rain rate model (PDIR) for high-resolution, real-time satellite precipitation estimation.Bulletin of the American...

    2020

  20. [28]

    Taikan Oki and Shinjiro Kanae. 2006. Global hydrological cycles and world water resources.science313, 5790 (2006), 1068–1072

    2006

  21. [29]

    Yongsheng Pan, Ming-Xia Liu, Yong Xia, and Dinggang Shen. 2022. Disease- Image-Specific Learning for Diagnosis-Oriented Neuroimage Synthesis With Incomplete Multi-Modality Data.IEEE Trans. Pattern Anal. Mach. Intell.44, 10 (2022), 6839–6853

    2022

  22. [30]

    Giulia Panegrossi, Stefano Dietrich, Frank S Marzano, Alberto Mugnai, Eric A Smith, Xuwu Xiang, Gregory J Tripoli, Pao K Wang, and JPV Poiares Baptista

  23. [31]

    Rajani K Pradhan, Yannis Markonis, Mijael Rodrigo Vargas Godoy, Anahí Villalba- Pradas, Konstantinos M Andreadis, Efthymios I Nikolopoulos, Simon Michael Papalexiou, Akif Rahim, Francisco J Tapiador, and Martin Hanel. 2022. Review of GPM IMERG performance: A global perspective...

    2022

  24. [32]

    Catherine Prigent. 2010. Precipitation retrieval from space: An overview.Comptes Rendus Geoscience342, 4-5 (2010), 380–389

    2010

  25. [33]

    Mojtaba Sadeghi, Phu Nguyen, Kuolin Hsu, and Soroosh Sorooshian. 2020. Im- proving near real-time precipitation estimation using a U-Net convolutional neural network and geographical information.Environ. Model. Softw.134 (2020), 104856

    2020

  26. [34]

    Mark J. Shensa. 1992. The discrete wavelet transform: wedding the a trous and Mallat algorithms.IEEE Trans. Signal Process.40, 10 (1992), 2464–2482

    1992

  27. [35]

    Chunxiang Shi, Lipeng Jiang, Tao Zhang, Bin Xu, and Shuai Han. 2014. Status and plans of CMA land data assimilation system (CLDAS) project. InEGU General Assembly Conference Abstracts. 5671

    2014

  28. [36]

    Wangbin Sun, Fei Ma, Yang Li, Shao-Lun Huang, Shiguang Ni, and Lin Zhang

  29. [37]

    Kevin E Trenberth, Aiguo Dai, Roy M Rasmussen, and David B Parsons. 2003. The changing character of precipitation.Bulletin of the American Meteorological Society84, 9 (2003), 1205–1218

    2003

  30. [38]

    Cunguang Wang, Guoqiang Tang, and Pierre Gentine. 2021. PrecipGAN: Merg- ing microwave and infrared data for satellite precipitation estimation using generative adversarial network.Geophysical Research Letters48, 5 (2021), e2020GL092032

    2021

  31. [39]

    Cunguang Wang, Jing Xu, Guoqiang Tang, Yi Yang, and Yang Hong. 2020. Infrared Precipitation Estimation Using Convolutional Neural Network.IEEE Trans. Geosci. Remote. Sens.58, 12 (2020), 8612–8625. Precipitation Retrieval with Multimodal Knowledge Expansion KDD ’25, August 3–7,...

    2020

  32. [41]

    Thompson, and Jiayu Zhou

    Qi Wang, Liang Zhan, Paul M. Thompson, and Jiayu Zhou. 2020. Multimodal Learning with Incomplete Modalities by Knowledge Distillation. InKDD ’20: The 26th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Virtual Event, CA, USA, August 23-27, 2020, Rajesh Gupta, Ya...

    2020

  33. [42]

    InIEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2021, Toronto, ON, Canada, June 6-11, 2021

    Semi-Supervised Multimodal Image Translation for Missing Modality Imputation. InIEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2021, Toronto, ON, Canada, June 6-11, 2021. IEEE, 4320–4324

    2021

  34. [43]

    Shicai Wei, Chunbo Luo, and Yang Luo. 2023. MMANet: Margin-Aware Distilla- tion and Modality-Aware Regularization for Incomplete Multimodal Learning. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023. IEEE, 20039–20049

    2023

  35. [44]

    Songkun Yan, Ziqiang Ma, Xiaoqing Li, Hao Hu, Jintao Xu, Qingwen Ji, and Fuzhong Weng. 2023. PCSSR-DNNWA: A physical constraints based surface snowfall rate retrieval algorithm using deep neural networks with attention module.Geophysical Research Letters50, 13 (2023), e2023GL103923

    2023

  36. [45]

    Zhendong Yang, Zhe Li, Mingqi Shao, Dachuan Shi, Zehuan Yuan, and Chun Yuan. 2022. Masked Generative Distillation. InComputer Vision - ECCV 2022 - 17th European Conference, Tel A viv, Israel, October 23-27, 2022, Proceedings, Part XI (Lecture Notes in Computer Science, Vol. 13...

    2022

  37. [46]

    Bo Zhang, Zuheng Ming, Wei Feng, Yaqian Liu, Liang He, and Kaixing Zhao

  38. [47]

    Qi Zhang, Yifei Wang, and Yisen Wang. 2022. How Mask Matters: Towards Theoretical Understandings of Masked Autoencoders. InAdvances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA,...

    2022

  39. [49]

    Siyu Zhu and Ziqiang Ma. 2022. PECA-FY4A: Precipitation Estimation using Chromatographic Analysis methodology for full-disc multispectral observations from FengYun-4A/AGRI.Remote Sensing of Environment282 (2022), 113234. KDD ’25, August 3–7, 2025, Toronto, ON, Canada Zheng Wan...

    2022

  40. [54]

    doi:10.48550/ ARXIV.2303.13101 arXiv:2303.13101

    MMFormer: Multimodal Transformer Using Multiscale Self-Attention for Remote Sensing Image Classification.CoRRabs/2303.13101 (2023). doi:10.48550/ ARXIV.2303.13101 arXiv:2303.13101

    2023 doi arXiv

  41. [56]

    Yuchen Zhang, Mingsheng Long, Kaiyuan Chen, Lanxiang Xing, Ronghua Jin, Michael I Jordan, and Jianmin Wang. 2023. Skilful nowcasting of extreme precip- itation with NowcastNet.Nature619, 7970 (2023), 526–532

    2023

  42. [1998]

    Journal of the atmospheric sciences55, 9 (1998), 1644–1673

    Use of cloud model microphysics for passive microwave-based precipitation retrieval: Significance of consistency between model and measurement manifolds. Journal of the atmospheric sciences55, 9 (1998), 1644–1673

    1998

  43. [2014]

    The global precipitation measurement mission.Bulletin of the American meteorological Society95, 5 (2014), 701–722

    2014

  44. [2016]

    HeMIS: Hetero-Modal Image Segmentation. InMedical Image Computing and Computer-Assisted Intervention - MICCAI 2016 - 19th International Conference, Athens, Greece, October 17-21, 2016, Proceedings, Part II (Lecture Notes in Computer Science, Vol. 9901), Sébastien Ourselin, Leo...

    2016

  45. [2021]

    SMIL: Multimodal Learning with Severely Missing Modality. InThirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Sympo- sium on Educational Advances in Artific...

    2021

  46. [2023]

    InIEEE International Conference on Acoustics, Speech and Signal Processing ICASSP 2023, Rhodes Island, Greece, June 4-10, 2023

    Prototype Knowledge Distillation for Medical Segmentation with Miss- ing Modality. InIEEE International Conference on Acoustics, Speech and Signal Processing ICASSP 2023, Rhodes Island, Greece, June 4-10, 2023. IEEE, 1–5

    2023

  47. [3578]

    doi:10.1109/TMI.2023.3295489

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.