CorrectCast:用于降水临近预报中尺度感知残差校正的选择性多模态上下文
摘要
该论文提出了 CorrectCast,一种解耦的残差校正框架。该框架冻结基础临近预报模型,仅训练一个以 QPE 和 PWV 观测数据为条件的轻量级校正器,以改进降水临近预报。论文报告,在 MRB 和 CONUS 数据集上,六种异构临近预报模型的 CSI 均有所提升。
arXiv:2611.22178v1 Announce Type: new
Abstract: Precipitation nowcasting models suffer from errors in intensity, location, and spatial structure. Addressing these errors by retraining nowcaster backbones can be computationally expensive and requires architecture-specific modifications. To overcome these limitations, we propose CorrectCast, a decoupled residual correction framework that refines completed nowcasts while keeping the base nowcaster frozen. By operating on predicted precipitation fields, CorrectCast requires training only a lightweight corrector, reducing computational overhead while providing a unified correction interface across different nowcasting models. Conditioned on the base nowcast and historical quantitative precipitation estimation (QPE) and precipitable water vapor (PWV) observations, the corrector selects multimodal context and employs scale-aware modulation to estimate residual corrections. On the Mississippi River Basin (MRB) and contiguous United States (CONUS) datasets, CorrectCast improves six base nowcasters spanning extrapolation-based, deterministic, and generative approaches, with absolute gains in mean CSI averaging 0.0233 and 0.0290 across the six models, respectively.
查看缓存全文
缓存时间: 2026/11/19 03:27
# CorrectCast: Selective Multimodal Context for Scale-Aware Residual Correction in Precipitation Nowcasting
Source: [https://arxiv.org/html/2611.22178](https://arxiv.org/html/2611.22178)
Shengwu XiongAffiliation:Wuhan, Hubei, ChinaYifang ZhangAffiliation:\{361332, xiongsw, yifangzhang, duanpf\}@wut\.edu\.cnPengfei Duan∗Affiliation:∗Corresponding author
###### Abstract
Precipitation nowcasting models suffer from errors in intensity, location, and spatial structure\. Addressing these errors by retraining nowcaster backbones can be computationally expensive and requires architecture\-specific modifications\. To overcome these limitations, we propose CorrectCast, a decoupled residual correction framework that refines completed nowcasts while keeping the base nowcaster frozen\. By operating on predicted precipitation fields, CorrectCast requires training only a lightweight corrector, reducing computational overhead while providing a unified correction interface across different nowcasting models\. Conditioned on the base nowcast and historical quantitative precipitation estimation \(QPE\) and precipitable water vapor \(PWV\) observations, the corrector selects multimodal context and employs scale\-aware modulation to estimate residual corrections\. On the Mississippi River Basin \(MRB\) and contiguous United States \(CONUS\) datasets, CorrectCast improves six base nowcasters spanning extrapolation\-based, deterministic, and generative approaches, with absolute gains in mean CSI averaging 0\.0233 and 0\.0290 across the six models, respectively\.
## 1Introduction
Precipitation nowcasting plays a critical role in providing timely warnings for rainfall\-related hazards\. Weather, climate and water disasters caused over two million deaths and $4\.3 trillion in losses during 1970–2021\[[12](https://arxiv.org/html/2611.22178#bib.bib12)\]\. Existing nowcasting approaches span extrapolation\-based methods\[[14](https://arxiv.org/html/2611.22178#bib.bib14)\], deterministic networks\[[11](https://arxiv.org/html/2611.22178#bib.bib11),[27](https://arxiv.org/html/2611.22178#bib.bib27),[8](https://arxiv.org/html/2611.22178#bib.bib8),[19](https://arxiv.org/html/2611.22178#bib.bib19)\], and generative models\[[15](https://arxiv.org/html/2611.22178#bib.bib15),[26](https://arxiv.org/html/2611.22178#bib.bib26),[4](https://arxiv.org/html/2611.22178#bib.bib4),[5](https://arxiv.org/html/2611.22178#bib.bib5),[16](https://arxiv.org/html/2611.22178#bib.bib16),[23](https://arxiv.org/html/2611.22178#bib.bib23)\]\. Despite advances, these approaches remain susceptible to errors in precipitation intensity, location, and spatial structure\. Extrapolation\-based methods can struggle with precipitation growth and decay, deterministic networks often produce overly smooth nowcasts at longer lead times, and generative models can exhibit spatial misalignment or structural discrepancies\. These limitations motivate a unified approach to refining nowcasts across diverse nowcasting architectures\.
Figure 1:Motivation and architectural design of CorrectCast\.As illustrated in Fig\.[1](https://arxiv.org/html/2611.22178#S1.F1)\(a\), extrapolation\-based methods, deterministic networks, and generative models employ different prediction mechanisms, making internal changes dependent on the specific architecture\. Improving their nowcasts may involve revising prediction modules or retraining network backbones\. Such strategies can require architecture\-specific adaptations and incur substantial computational costs\. DiffCast compares end\-to\-end training with a frozen\-backbone predictor\-corrector variant, reporting higher CSI under joint training for three deterministic backbones on SEVIR\[[24](https://arxiv.org/html/2611.22178#bib.bib24)\]\. However, this result does not establish whether output\-level correction can work effectively across heterogeneous, frozen base nowcasters\. We therefore investigate a complementary approach: a decoupled predictor\-corrector framework that learns residual corrections from completed nowcasts without accessing internal model features or updating the base nowcaster\.
However, learning effective residual corrections from completed nowcasts poses two key challenges, as illustrated in Fig\.[1](https://arxiv.org/html/2611.22178#S1.F1)\(b\)\. First, the corrector must identify historical information that is relevant to errors remaining in the base nowcast\. Historical QPE captures recent precipitation patterns, while PWV provides complementary information on atmospheric moisture conditions\[[7](https://arxiv.org/html/2611.22178#bib.bib7),[1](https://arxiv.org/html/2611.22178#bib.bib1),[20](https://arxiv.org/html/2611.22178#bib.bib20),[9](https://arxiv.org/html/2611.22178#bib.bib9)\]\. However, some QPE information may already be encoded in the base nowcast, and PWV features may include both precipitation\-relevant variations and signals unrelated to residual errors\. Directly combining these observations may therefore introduce redundancy and task\-irrelevant information, limiting their value for correction\. Second, the corrector must account for errors at different spatial scales\. Local discrepancies call for flexible, position\-dependent adjustments across neighboring locations, whereas broader structural errors involve the extent and organization of precipitation systems\. Addressing these larger\-scale discrepancies requires coordinated adjustments across spatially related regions, rather than independent corrections at individual locations\.
To address these challenges, we propose CorrectCast, a decoupled residual correction framework that integrates selective multimodal context construction with scale\-aware feature modulation \(SAFM\), as illustrated in Fig\.[1](https://arxiv.org/html/2611.22178#S1.F1)\(c\)\. Given a completed nowcast from a frozen base nowcaster, CorrectCast estimates a residual correction conditioned on historical QPE and PWV observations\. To construct informative context, we extract multiscale precipitation features from QPE and characterize PWV in terms of moisture state, temporal variation, and spatial distribution\. We then employ a two\-stage information selection strategy: cross\-modal orthogonal decomposition suppresses PWV components linearly aligned with QPE features, and a spatiotemporal variational information bottleneck \(VIB\) subsequently constrains the information retained in the fused representation\. To address errors at different spatial scales, Spatial SAFM enables flexible, position\-dependent modulation of nowcast features, while Wavelet SAFM applies low\-frequency wavelet modulation to encourage coordinated adjustments across broader regions\. The resulting features are decoded into a residual correction that refines the original nowcast\.
Our main contributions are summarized as follows:
- •We introduce CorrectCast, an output\-level correction framework that refines completed nowcasts without modifying base nowcasters or accessing their internal features, providing a common correction interface across different nowcasting paradigms\.
- •We develop a correction\-oriented representation strategy that combines precipitation and moisture information through cross\-modal orthogonal decomposition and a spatiotemporal VIB to reduce redundancy and encourage task\-relevant context\.
- •We design Spatial SAFM and Wavelet SAFM to support flexible local adjustments and coherent large\-scale structural refinement, and demonstrate the effectiveness of CorrectCast across six base nowcasters on two datasets with different geographic coverage\.
## 2Related Work
### 2\.1Precipitation Nowcasting Paradigms
Precipitation nowcasting spans numerical extrapolation, deterministic, and generative methods\. Numerical methods such as PySTEPS\[[14](https://arxiv.org/html/2611.22178#bib.bib14)\]efficiently advect observed precipitation but struggle with growth and decay\. Deterministic models learn precipitation evolution from historical observations but may produce smooth nowcasts at longer lead times\[[17](https://arxiv.org/html/2611.22178#bib.bib17),[13](https://arxiv.org/html/2611.22178#bib.bib13),[3](https://arxiv.org/html/2611.22178#bib.bib3),[6](https://arxiv.org/html/2611.22178#bib.bib6),[18](https://arxiv.org/html/2611.22178#bib.bib18),[22](https://arxiv.org/html/2611.22178#bib.bib22),[21](https://arxiv.org/html/2611.22178#bib.bib21)\]\. Generative models, including GAN\-based\[[15](https://arxiv.org/html/2611.22178#bib.bib15),[26](https://arxiv.org/html/2611.22178#bib.bib26),[10](https://arxiv.org/html/2611.22178#bib.bib10)\]and diffusion\-based approaches\[[4](https://arxiv.org/html/2611.22178#bib.bib4),[24](https://arxiv.org/html/2611.22178#bib.bib24)\], can preserve fine\-scale structures but may exhibit spatial displacement or structural discrepancies\[[26](https://arxiv.org/html/2611.22178#bib.bib26),[24](https://arxiv.org/html/2611.22178#bib.bib24)\]\.
### 2\.2Generative Refinement and Output Correction
DiffCast combines a deterministic nowcasting backbone with a residual diffusion model, favoring end\-to\-end optimization while also evaluating a frozen backbone predictor\-corrector variant\[[24](https://arxiv.org/html/2611.22178#bib.bib24)\]\. MoCast\+ extends the physics\-guided MoCast predictor with diffusion\-based residual post\-processing\[[23](https://arxiv.org/html/2611.22178#bib.bib23)\]\. These approaches demonstrate the value of residual refinement but primarily study diffusion\-based residual refinement within specific nowcasting frameworks\. WaveC2R employs wavelet\-guided diffusion to refine satellite radar retrievals, demonstrating the potential of refinement with frequency awareness for precipitation structures\[[16](https://arxiv.org/html/2611.22178#bib.bib16)\]\. CorrectCast instead investigates residual correction as an independent stage applied to completed nowcasts from heterogeneous base models\. CorrectCast trains a lightweight corrector conditioned on the base nowcast and historical QPE and PWV, without accessing internal model features or updating the base nowcaster\. By separating correction from prediction, CorrectCast provides a common output\-level interface across nowcasting paradigms, with additional training restricted to the corrector\.
Figure 2:Overall architecture of the proposed CorrectCast framework\.
## 3Problem Formulation
LetY0∈ℝT×H×WY^\{0\}\\in\\mathbb\{R\}^\{T\\times H\\times W\}denote a completed nowcast produced by a base nowcaster, and letY∈ℝT×H×WY\\in\\mathbb\{R\}^\{T\\times H\\times W\}denote the corresponding ground\-truth future precipitation sequence\. Historical QPE and PWV observations are denoted byQ∈ℝK×H×WQ\\in\\mathbb\{R\}^\{K\\times H\\times W\}andP∈ℝK×H×WP\\in\\mathbb\{R\}^\{K\\times H\\times W\}, respectively\. Here,KKandTTdenote the numbers of historical frames and nowcast frames, whileHHandWWrepresent the spatial dimensions\. CorrectCast predicts a residual correction:
ΔY^=𝒢Θ\(Y0,Q,P\),\\widehat\{\\Delta Y\}=\\mathcal\{G\}\_\{\\Theta\}\\left\(Y^\{0\},Q,P\\right\),\(1\)where𝒢Θ\\mathcal\{G\}\_\{\\Theta\}denotes the learnable corrector parameterized byΘ\\Theta, and the target residual is defined asΔY=Y−Y0\\Delta Y=Y\-Y^\{0\}\. The corrected nowcast is then obtained as
Y^=Π\+\(Y0\+ΔY^\),\\widehat\{Y\}=\\Pi\_\{\+\}\\left\(Y^\{0\}\+\\widehat\{\\Delta Y\}\\right\),\(2\)whereΠ\+\(⋅\)\\Pi\_\{\+\}\(\\cdot\)denotes element\-wise projection onto the nonnegative domain\. Throughout training, the base nowcaster remains fixed, and only the corrector parametersθ\\thetaare optimized\. CorrectCast thus learns to refine completed nowcasts using historical multimodal observations without updating the base nowcaster\.
## 4Methodology
### 4\.1Overview
CorrectCast consists of two stages: selective multimodal context construction and scale\-aware residual correction\. As illustrated in Fig\.[2](https://arxiv.org/html/2611.22178#S2.F2), historical QPE and PWV are first encoded into candidate correction features\. Cross\-modal orthogonal decomposition first suppresses PWV components linearly aligned with QPE representations, followed by a spatiotemporal VIB that encourages a compact fused representation with reduced input dependence while preserving information relevant to residual correction\. Guided by this context, Spatial SAFM applies flexible, position\-dependent modulation to the completed nowcast, while Wavelet SAFM subsequently introduces low\-frequency wavelet modulation to support coordinated refinement of broader precipitation structures\. Finally, a decoder predicts the residual correctionΔY^\\widehat\{\\Delta Y\}, which is combined withY0Y^\{0\}according to Eq\. \([2](https://arxiv.org/html/2611.22178#S3.E2)\)\.
### 4\.2Selective Multimodal Context Construction
#### 4\.2\.1Historical QPE Encoder
The base nowcaster already uses historical QPE observations to model the evolution of the precipitation field\. Rather than repeating the full nowcasting process, the QPE branch of CorrectCast extracts compact multiscale features to support residual correction\. Given a historical QPE sequenceQ=\{Qt\}t=1KQ=\\\{Q\_\{t\}\\\}\_\{t=1\}^\{K\}, dilated convolutionsϕs\(⋅\)\\phi\_\{s\}\(\\cdot\)are applied to each frameQtQ\_\{t\}with dilation rates𝒮=\{2,4\}\\mathcal\{S\}=\\\{2,4\\\}:
Ft\(s\)=ϕs\(Qt\),s∈𝒮\.F\_\{t\}^\{\(s\)\}=\\phi\_\{s\}\(Q\_\{t\}\),\\qquad s\\in\\mathcal\{S\}\.\(3\)The resulting features are concatenated to form the multiscale representation of each frame:
Ft=Concats∈𝒮\(Ft\(s\)\)\.F\_\{t\}=\\operatorname\{Concat\}\_\{s\\in\\mathcal\{S\}\}\\left\(F\_\{t\}^\{\(s\)\}\\right\)\.\(4\)A ConvGRU then captures temporal dependencies across\{Ft\}t=1K\\\{F\_\{t\}\\\}\_\{t=1\}^\{K\}to produce the QPE representationHQH^\{Q\}\.
#### 4\.2\.2Moisture\-Variation PWV Encoder
PWV provides information about atmospheric moisture conditions that is complementary to the historical QPE\. To characterize the moisture state and its temporal and spatial variations, each PWV framePtP\_\{t\}is augmented with a temporal\-difference descriptorDttdD\_\{t\}^\{\\mathrm\{td\}\}and a spatial\-Laplacian descriptorDtslD\_\{t\}^\{\\mathrm\{sl\}\}\. We further apply spatial normalization to the PWV field and its derived descriptors\. The resulting frame\-level input is
Gt=Concat\(Pt,ℳ\(Pt\),ℳ\(Dttd\),ℳ\(Dtsl\)\),G\_\{t\}=\\operatorname\{Concat\}\\left\(P\_\{t\},\\mathcal\{M\}\(P\_\{t\}\),\\mathcal\{M\}\(D\_\{t\}^\{\\mathrm\{td\}\}\),\\mathcal\{M\}\(D\_\{t\}^\{\\mathrm\{sl\}\}\)\\right\),\(5\)whereℳ\(⋅\)\\mathcal\{M\}\(\\cdot\)denotes spatial normalization\. A convolutional block first transformsGtG\_\{t\}into a learned feature representationEtE\_\{t\}\. A ConvGRU then captures temporal dependencies across\{Et\}t=1K\\\{E\_\{t\}\\\}\_\{t=1\}^\{K\}, producing the PWV representationHPH^\{P\}from its final hidden state\.
#### 4\.2\.3Cross\-Modal Linear Redundancy Reduction
QPE and PWV representations may contain overlapping information, and directly retaining such information can introduce redundancy into the residual correction process\. To reduce this overlap, we suppress the PWV components linearly aligned with the corresponding QPE features\. Specifically, for each sample and channel, we flatten the spatial dimensions ofHQH^\{Q\}andHPH^\{P\}into vectorshQ,hP∈ℝHWh^\{Q\},h^\{P\}\\in\\mathbb\{R\}^\{HW\}\. The projection ofhPh^\{P\}ontohQh^\{Q\}is defined as
h∥P=⟨hP,hQ⟩‖hQ‖22\+εhQ,h\_\{\\parallel\}^\{P\}=\\frac\{\\left\\langle h^\{P\},h^\{Q\}\\right\\rangle\}\{\\left\\\|h^\{Q\}\\right\\\|\_\{2\}^\{2\}\+\\varepsilon\}h^\{Q\},\(6\)whereε\>0\\varepsilon\>0ensures numerical stability\. Subtracting this projection yields the remaining PWV componenth⟂P=hP−h∥Ph\_\{\\perp\}^\{P\}=h^\{P\}\-h\_\{\\parallel\}^\{P\}, which is approximately orthogonal tohQh^\{Q\}whenε\\varepsilonis sufficiently small\. After restoring the spatial dimensions, we obtainH⟂PH\_\{\\perp\}^\{P\}and construct the fused representationX=Concat\(HQ,H⟂P\)X=\\operatorname\{Concat\}\\left\(H^\{Q\},H\_\{\\perp\}^\{P\}\\right\)\. This operation reduces channel\-wise linear alignment between QPE and PWV features without assuming complete statistical independence\.
#### 4\.2\.4Task\-Level Redundancy Reduction
Since the fused representationXXmay contain information irrelevant to residual correction, we employ a spatiotemporal VIB to learn a compact, correction\-oriented representationZZthrough a diagonal gaussian posterior:
qθ\(Z∣X\)=𝒩\(μ,diag\(σ2\)\),q\_\{\\theta\}\(Z\\mid X\)=\\mathcal\{N\}\\left\(\\mu,\\operatorname\{diag\}\(\\sigma^\{2\}\)\\right\),\(7\)where\(μ,logσ2\)=fθ\(X\)\(\\mu,\\log\\sigma^\{2\}\)=f\_\{\\theta\}\(X\), andfθf\_\{\\theta\}denotes a learnable convolutional block\. The latent representation is sampled using the reparameterization trick:
Z=μ\+σ⊙ϵ,ϵ∼𝒩\(0,I\)\.Z=\\mu\+\\sigma\\odot\\epsilon,\\qquad\\epsilon\\sim\\mathcal\{N\}\(0,I\)\.\(8\)
The information bottleneck objective encouragesZZto retain correction\-relevant information, while limiting its dependence onXX:
maxθI\(Z;ΔY∣Y0,HQ\)−βI\(Z;X\),β\>0\.\\max\_\{\\theta\}\\;I\(Z;\\Delta Y\\mid Y^\{0\},H^\{Q\}\)\-\\beta I\(Z;X\),\\qquad\\beta\>0\.\(9\)Directly computingI\(Z,X\)I\(Z;X\)is difficult because the marginal distribution ofqθ\(Z\)q\_\{\\theta\}\(Z\)is generally intractable\. To obtain a computable alternative, we introduce a standard gaussian priorp\(Z\)=𝒩\(0,I\)p\(Z\)=\\mathcal\{N\}\(0,I\)\. By the definition of mutual information, the expected KL divergence satisfies
𝔼X\[DKL\(qθ\(Z∣X\)∥p\(Z\)\)\]\\displaystyle\\mathbb\{E\}\_\{X\}\\left\[D\_\{\\mathrm\{KL\}\}\\left\(q\_\{\\theta\}\(Z\\mid X\)\\,\\\|\\,p\(Z\)\\right\)\\right\]\(10\)=I\(Z;X\)\+DKL\(qθ\(Z\)∥p\(Z\)\)\.\\displaystyle=I\(Z;X\)\+D\_\{\\mathrm\{KL\}\}\\left\(q\_\{\\theta\}\(Z\)\\,\\\|\\,p\(Z\)\\right\)\.Since the second term is non\-negative, the mutual information is upper\-bounded by
I\(Z;X\)≤𝔼X\[DKL\(qθ\(Z∣X\)∥p\(Z\)\)\]\.I\(Z;X\)\\leq\\mathbb\{E\}\_\{X\}\\left\[D\_\{\\mathrm\{KL\}\}\\left\(q\_\{\\theta\}\(Z\\mid X\)\\,\\\|\\,p\(Z\)\\right\)\\right\]\.\(11\)This provides an alternative to directly minimizingI\(Z,X\)I\(Z;X\)\. However, minimizing the KL divergence alone may also remove information needed for residual correction\. Therefore, we superviseZZthrough the complete downstream correction network, together withY0Y^\{0\}andHQH^\{Q\}, to predict the precipitation residual:
ΔY^=ℛϕ\(Y0,HQ,Z\),\\widehat\{\\Delta Y\}=\\mathcal\{R\}\_\{\\phi\}\\left\(Y^\{0\},H^\{Q\},Z\\right\),\(12\)whereℛϕ\\mathcal\{R\}\_\{\\phi\}comprises scale\-aware residual correction and residual decoding\. To measure the correction accuracy, we define the residual reconstruction loss as
ℒrec=‖ΔY−ΔY^‖22\.\\mathcal\{L\}\_\{\\mathrm\{rec\}\}=\\left\\\|\\Delta Y\-\\widehat\{\\Delta Y\}\\right\\\|\_\{2\}^\{2\}\.\(13\)Assuming a fixed\-variance isotropic gaussian residual likelihood, minimizing the expected loss in Eq\.[13](https://arxiv.org/html/2611.22178#S4.E13)maximizes a variational lower bound onI\(Z;ΔY∣Y0,HQ\)I\(Z;\\Delta Y\\mid Y^\{0\},H^\{Q\}\)\.
Finally, the reconstruction loss and KL regularization are combined to jointly encourage correction accuracy and information compression:
ℒVIB=λmseℒrec\+λklDKL\(qθ\(Z∣X\)∥p\(Z\)\),\\mathcal\{L\}\_\{\\mathrm\{VIB\}\}=\\lambda\_\{\\mathrm\{mse\}\}\\mathcal\{L\}\_\{\\mathrm\{rec\}\}\+\\lambda\_\{\\mathrm\{kl\}\}D\_\{\\mathrm\{KL\}\}\\left\(q\_\{\\theta\}\(Z\\mid X\)\\,\\\|\\,p\(Z\)\\right\),\(14\)whereλmse\\lambda\_\{\\mathrm\{mse\}\}andλkl\\lambda\_\{\\mathrm\{kl\}\}balance residual fidelity and latent information compression, respectively\.
### 4\.3Scale\-Aware Residual Correction
#### 4\.3\.1Spatial SAFM for Local Adaptation
Letfpriorf\_\{\\mathrm\{prior\}\}denote features extracted from the completed base nowcastY0Y^\{0\}\. We first combine the historical QPE representationHQH^\{Q\}with the VIB latent representationZZto construct a fused correction context, which is then injected into the nowcast features:
ZA=HQ\+𝒥\(Z\),finit=fprior\+ZA,Z\_\{A\}=H^\{Q\}\+\\mathcal\{J\}\(Z\),\\qquad f\_\{\\mathrm\{init\}\}=f\_\{\\mathrm\{prior\}\}\+Z\_\{A\},\(15\)where𝒥\\mathcal\{J\}denotes a learned projection with feature scaling\.
Guided byZAZ\_\{A\}, Spatial SAFM generates spatially varying scale and shift parameters to refine the initialized features:
γspatial\\displaystyle\\gamma\_\{\\mathrm\{spatial\}\}=gγ\(ZA\),βspatial=gβ\(ZA\),\\displaystyle=g\_\{\\gamma\}\(Z\_\{A\}\),\\qquad\\beta\_\{\\mathrm\{spatial\}\}=g\_\{\\beta\}\(Z\_\{A\}\),\(16\)fspatial\\displaystyle f\_\{\\mathrm\{spatial\}\}=finit⊙\(1\+γspatial\)\+βspatial\.\\displaystyle=f\_\{\\mathrm\{init\}\}\\odot\(1\+\\gamma\_\{\\mathrm\{spatial\}\}\)\+\\beta\_\{\\mathrm\{spatial\}\}\.By varying the modulation parameters across spatial locations, Spatial SAFM enables localized enhancement or suppression of feature responses, allowing different adjustments at neighboring positions\.
Table 1:Quantitative results on the MRB and CONUS datasets\. Within each base nowcaster pair,boldfaceindicates the better value\. For LPIPS and MSE, the improvement row reports the absolute reduction\.
#### 4\.3\.2Wavelet SAFM for Coherent Refinement
To complement local adaptation, Wavelet SAFM applies a discrete wavelet transform tofspatialf\_\{\\mathrm\{spatial\}\}, separating the low\-frequency componentLLLLfrom high\-frequency detail components\. Guided by the correction contextZAZ\_\{A\}, the low\-frequency component is modulated to support coordinated adjustments across broader spatial regions\. The modulatedLLLLand unchanged high\-frequency components are then recombined through the inverse discrete wavelet transform, incorporating broader structural adjustments while passing the high\-frequency subbands of the spatially modulated features through the wavelet stage unchanged\. The reconstructed features are decoded into the residual correctionΔY^\\widehat\{\\Delta Y\}, which is combined withY0Y^\{0\}and projected onto the nonnegative domain according to Eq\. \([2](https://arxiv.org/html/2611.22178#S3.E2)\)\.
### 4\.4Training Objective
The overall training objective combines the VIB loss in Eq\. \([14](https://arxiv.org/html/2611.22178#S4.E14)\) with two auxiliary losses\. A threshold\-based weighted binary cross\-entropy lossℒbce\\mathcal\{L\}\_\{\\mathrm\{bce\}\}encourages accurate detection of high\-intensity precipitation, while a spectral\-consistency lossℒspec\\mathcal\{L\}\_\{\\mathrm\{spec\}\}penalizes differences between the amplitude spectra of the corrected nowcast and ground truth\. The total loss is defined as
ℒ=ℒVIB\+λbceℒbce\+λspecℒspec\.\\mathcal\{L\}=\\mathcal\{L\}\_\{\\mathrm\{VIB\}\}\+\\lambda\_\{\\mathrm\{bce\}\}\\mathcal\{L\}\_\{\\mathrm\{bce\}\}\+\\lambda\_\{\\mathrm\{spec\}\}\\mathcal\{L\}\_\{\\mathrm\{spec\}\}\.\(17\)
## 5Experiments
### 5\.1Experimental Setup
#### 5\.1\.1Datasets
We evaluate CorrectCast on two precipitation nowcasting datasets, MRB and CONUS, which integrate Multi\-Radar Multi\-Sensor \(MRMS\) QPE\[[25](https://arxiv.org/html/2611.22178#bib.bib25)\]with Global Navigation Satellite System \(GNSS\)\-derived PWV from NGL\[[2](https://arxiv.org/html/2611.22178#bib.bib2)\]\. MRMS QPE is provided at0\.01∘0\.01^\{\\circ\}\( 1 km\) spatial resolution, and both variables are temporally aligned at 10\-min intervals\.
MRB: QPE is cropped to1024×10241024\\times 1024grids over32\.0∘32\.0^\{\\circ\}–42\.23∘42\.23^\{\\circ\}N and93\.5∘93\.5^\{\\circ\}–83\.25∘83\.25^\{\\circ\}W, with PWV from 1,526 quality\-controlled GNSS stations bilinearly interpolated onto the same grid\. The training, validation, and test periods are Nov\. 2020–Jun\. 2022, Jul\.–Aug\. 2022, and Sep\. 2022–Aug\. 2023, respectively\.
CONUS: Following the importance\-sampling strategy of DGMR\[[15](https://arxiv.org/html/2611.22178#bib.bib15)\], precipitation events lasting over 4 h are sampled from the full MRMS domain and cropped into1024×10241024\\times 1024patches\. PWV from 8,229 GNSS stations is bilinearly interpolated and spatially aligned with the QPE patches\. The training, validation, and test periods are Nov\. 2020–Aug\. 2022, Sep\.–Oct\. 2022, and Nov\. 2022–Oct\. 2023, respectively\.
#### 5\.1\.2Evaluation Protocol
Following NowcastNet\[[26](https://arxiv.org/html/2611.22178#bib.bib26)\]and DiffCast\[[24](https://arxiv.org/html/2611.22178#bib.bib24)\], we evaluate categorical nowcast skill using the critical success index \(CSI\), heidke skill score \(HSS\), CSI\-pool4, and CSI\-pool16 at precipitation thresholdsτ∈\{16\.0,32\.0,64\.0\}\\tau\\in\\\{16\.0,32\.0,64\.0\\\}mm/h\. We additionally report the structural similarity index measure \(SSIM\), learned perceptual image patch similarity \(LPIPS\), and mean squared error \(MSE\) to characterize structural, perceptual, and pixelwise agreement, respectively\. The extra details are reported in the supplementary material\.
Figure 3:Qualitative comparison for a heavy precipitation event in the MRB dataset\. Due to space limitations, the results of PhyDNet, ExPreCast, and DiffCast, together with the CONUS case, are provided in the supplementary material\.Figure 4:CSI and HSS across lead times on the MRB and CONUS datasets\. For clarity, only PySTEPS, PhyDNet, and NowcastNet are shown, before and after applying CorrectCast\.
### 5\.2Experimental Results
Table[1](https://arxiv.org/html/2611.22178#S4.T1)reports the quantitative results on MRB and CONUS\. CorrectCast improves CSI and HSS for all six base nowcasters on both datasets, with mean absolute CSI gains of 0\.0233 and 0\.0290, and corresponding HSS gains of 0\.0279 and 0\.0452, respectively\. On MRB, AlphaPre improves from 0\.3244 to 0\.3617 in CSI and from 0\.4748 to 0\.5167 in HSS\. The gains also extend to extrapolation\-based methods, with PySTEPS achieving a CSI improvement of 0\.0160\. On CONUS, NowcastNet exhibits particularly large gains, increasing CSI from 0\.0135 to 0\.0902 and HSS from 0\.0260 to 0\.1479\.
The improvements extend to spatially pooled and image\-quality metrics\. Both CSI\-pool4 and CSI\-pool16 increase for every base nowcaster on the two datasets\. Notably, CSI\-pool16 improves by 0\.1304 for PhyDNet on MRB and 0\.1668 for AlphaPre on CONUS\. Furthermore, SSIM increases for all six models on both datasets, while LPIPS and MSE consistently decrease\. Although the improvement magnitudes vary across models, the consistently positive results support the applicability of CorrectCast across different nowcasting paradigms\. The qualitative comparison in Fig\.[3](https://arxiv.org/html/2611.22178#S5.F3)further illustrates the effects of residual correction\. The base nowcasts exhibit varying degrees of smoothing, rainband displacement, and local intensity errors\. CorrectCast adjusts precipitation intensity and spatial extent while retaining the local structure of precipitation cells\.
Table 2:Quantitative comparison of generative refinement models on the MRB dataset\. The best results are highlighted inboldand the second\-best results areunderlined\.Table[2](https://arxiv.org/html/2611.22178#S5.T2)compares CorrectCast with representative diffusion\-based refinement approaches on MRB\. CorrectCast achieves the best results on six of the seven metrics\. Compared with DiffCast, it improves CSI by 0\.0400 and HSS by 0\.0347, while reducing LPIPS from 0\.0609 to 0\.0539 and MSE from 12\.2483 to 8\.0004\. DiffCast performs better on CSI\-pool16\. Overall, CorrectCast performs better under most reported metrics, although DiffCast remains superior at the largest pooling scale\. Details of the comparison setup are provided in the supplementary material\.
Fig\.[4](https://arxiv.org/html/2611.22178#S5.F4)reports CSI and HSS as functions of nowcast lead time\. Both metrics generally decrease at longer horizons, reflecting the increasing difficulty of precipitation nowcasting\. CorrectCast nevertheless improves the corresponding base nowcasts over most evaluated lead times\. The benefit is therefore not restricted to the first few nowcast frames\.
### 5\.3Analysis and Discussions
#### 5\.3\.1Which Components Contribute to CorrectCast?
Table 3:Component ablation of CorrectCast on MRB using NowcastNet as the base nowcaster\. All learned variants use the base nowcast, historical QPE, and moisture\-guided PWV as inputs\. Best results are highlighted inbold\.Table[3](https://arxiv.org/html/2611.22178#S5.T3)examines the contributions of information selection, scale\-aware modulation, and residual correction\. Orthogonal decomposition and spatiotemporal VIB individually improve CSI over the base nowcaster, with their combination further increasing CSI to 0\.2060\. Adding Spatial SAFM raises CSI to 0\.2074, whereas adding Wavelet SAFM alone does not yield an additional gain\. Combining both modulation modules achieves the best overall performance, with CSI reaching 0\.2158 and HSS reaching 0\.3206\. Under identical inputs and modules, direct prediction obtains a CSI of 0\.0964, substantially below the 0\.2158 achieved by residual correction\. These results support the combined design of CorrectCast, with additional task and PWV controls reported in the supplementary material\.
Table 4:Ablation study of the training losses on the MRB dataset\. The best results are highlighted inboldand the second\-best results areunderlined\.
#### 5\.3\.2Training Objective Analysis
The loss weights are selected by a two\-stage validation procedure \(see supplementary material\) and then fixed for all experiments\. The selected values areλmse=1\.0\\lambda\_\{\\mathrm\{mse\}\}=1\.0,λbce=0\.5\\lambda\_\{\\mathrm\{bce\}\}=0\.5,λkl=10−3\\lambda\_\{\\mathrm\{kl\}\}=10^\{\-3\}, andλspec=0\.3\\lambda\_\{\\mathrm\{spec\}\}=0\.3\. Table[4](https://arxiv.org/html/2611.22178#S5.T4)shows that both auxiliary objectives contribute to performance\. Removing the spectral loss reduces CSI\-pool16 from 0\.4150 to 0\.3396 and increases LPIPS from 0\.0586 to 0\.0755, consistent with its role in constraining spatial frequency structure\. Removing the BCE term also degrades all reported metrics and increases MSE from 9\.3595 to 10\.8713\. The two terms therefore provide complementary supervision to the residual reconstruction and VIB objectives\.
Figure 5:Mechanism analysis of CorrectCast\. \(a\) Absolute cosine similarity between QPE and PWV features before/after orthogonal decomposition\. \(b\) FSS over neighborhood sizes: base nowcast, CorrectCast, and CorrectCast without Wavelet SAFM\. \(c\) Residual\-PSD reduction over the base nowcast across spatial frequencies; inset gives the mean reduction in low\-/mid\-/high\-frequency bands, positive values indicate lower residual\-error\.
#### 5\.3\.3Do the Modules Behave as Intended?
Fig\.[5](https://arxiv.org/html/2611.22178#S5.F5)\(a\) shows that orthogonal decomposition reduces the mean absolute cosine similarity between QPE and PWV features from 0\.5299 to5\.97×10−85\.97\\times 10^\{\-8\}, indicating effective suppression of channel\-wise linear alignment\. To assess the predictive utility of the retained PWV information, we evaluate residual prediction probes with frozen QPE and PWV encoders\. As shown in Table[5](https://arxiv.org/html/2611.22178#S5.T5), adding the QPE\-orthogonal component \(H⟂PH\_\{\\perp\}^\{P\}\) achieves the best CSI, HSS, and CSI\-pool16\. In contrast, the QPE\-aligned component \(H∥PH\_\{\\parallel\}^\{P\}\) provides only marginal gains, while including both components reduces performance on these three metrics\. However, none of the PWV\-augmented probes improves MSE\. These results suggest that the retained orthogonal component is particularly useful for improving categorical nowcast skill, although this benefit does not extend to pixelwise MSE in the probe setting\. Further comparisons of orthogonalization strategies are provided in the supplementary material\.
Fig\.[5](https://arxiv.org/html/2611.22178#S5.F5)\(b\) assesses spatial agreement using the fractions skill score \(FSS\) across different neighborhood sizes\. CorrectCast achieves higher FSS than the base nowcast at all evaluated scales\. Removing Wavelet SAFM reduces FSS across these scales, with a more pronounced decrease at larger neighborhood sizes\. This pattern suggests that low\-frequency wavelet modulation complements spatial adaptation by improving agreement over broader precipitation structures\.
Fig\.[5](https://arxiv.org/html/2611.22178#S5.F5)\(c\) examines changes in residual power spectral density \(PSD\) relative to the base nowcast\. Positive values indicate reductions in prediction\-error power at the corresponding spatial frequencies\. CorrectCast reduces residual power across most of the spectrum, with mean decreases of 68\.93%, 25\.19%, and 8\.56% in the low\-, mid\-, and high\-frequency bands, respectively\. The greater reduction at low frequencies suggests improved correction of large\-scale precipitation errors, consistent with the intended role of Wavelet SAFM\.
We further assess statistical dependence using the normalized Hilbert–Schmidt independence criterion \(nHSIC\)\. As shown in Fig\.[6](https://arxiv.org/html/2611.22178#S5.F6), spatiotemporal VIB reduces input dependence from 0\.7747 to 0\.6718 while increasing target dependence from 0\.7496 to 0\.8561, relative to direct fusion\. The paired differences are−0\.1029\-0\.1029and\+0\.1065\+0\.1065, respectively, and the corresponding 95% bootstrap confidence intervals both exclude zero\. These results indicate a more favorable balance between input compression and target relevance, consistent with the intended role of spatiotemporal VIB in learning correction\-oriented representations\.
Table 5:Predictive utility of PWV components under frozen\-encoder residual probes\. The best results are highlighted inboldand the second\-best results areunderlined\.Figure 6:Information\-dependence analysis of the spatiotemporal VIB module\. \(a\) VIB information plane; preferred direction is lower input dependence with higher target relevance\. \(b\) Paired deltas over direct fusion \(95% bootstrap CI\);Δ\\Deltainput <0 andΔ\\Deltatarget \>0 denote the intended spatiotemporal VIB behavior\.
#### 5\.3\.4Does CorrectCast Preserve Motion Consistency?
We estimate optical flow from consecutive precipitation frames and evaluate motion agreement using endpoint error \(EPE\) between the base nowcast and the CorrectCast\-corrected nowcast\. As shown in Fig\.[7](https://arxiv.org/html/2611.22178#S5.F7), the base nowcast exhibits a fragmented flow pattern with inconsistent motion directions across neighboring regions\. In contrast, the corrected nowcast produces a more coherent flow field that better matches the motion estimated from ground\-truth precipitation\. The corresponding EPE maps show a concentrated region of large errors in the base nowcast, which is substantially reduced after correction\. The mean EPE decreases from 4\.7610 to 1\.3641 pixels, indicating improved agreement with the reference motion field\.
Figure 7:\(a\)\-\(c\) Estimated motion fields from ground\-truth, base nowcast, and corrected nowcast \(arrows: direction and speed\)\. \(d\) EPE maps between the base nowcast and the CorrectCast\-corrected nowcast\.
#### 5\.3\.5How Well Does CorrectCast Transfer Across Base Nowcasters?
We evaluate cross\-model transfer on CONUS by applying correctors trained with one base nowcaster to other nowcasters without retraining\. Fig\.[8](https://arxiv.org/html/2611.22178#S5.F8)shows that improvements are less consistent than in the matched\-model setting, with positive gains occurring more frequently for spatially pooled CSI than for pixelwise CSI\. Notably, the AlphaPre\-trained corrector improves all three CSI metrics for every other base nowcaster, whereas the NowcastNet\-trained corrector degrades performance for several target models\. These results show that CorrectCast can be applied across different nowcasting architectures without modification, although successful transfer depends on the source and target models\.
Figure 8:Cross\-model transfer on CONUS\. X\-axis: target base nowcaster receiving the corrector without retraining; Y\-axis: source base nowcaster used to train the corrector\. Off\-diagonal cells report absolute CSI, CSI\-pool4, and CSI\-pool16 changes over the corresponding target baseline; diagonals omitted\.
#### 5\.3\.6How Well Does CorrectCast Transfer from MRB to CONUS?
To evaluate cross\-dataset transfer, we train CorrectCast on MRB and apply it to CONUS without retraining\. As shown in Fig\.[9](https://arxiv.org/html/2611.22178#S5.F9), CSI improves for five of the six base nowcasters, with gains in spatially pooled CSI also observed for AlphaPre, NowcastNet, PhyDNet, and ExPreCast\. However, PySTEPS exhibits decreases across all three CSI metrics, while DiffCast shows a slight decline in CSI\-pool16\. These mixed results suggest partial transferability between the two datasets, while indicating that the effectiveness of transferred corrections varies with the base nowcaster and target\-domain conditions\.
Figure 9:Cross\-region transfer from MRB to CONUS\. The base nowcast and cross\-region corrected nowcast are compared using CSI, CSI\-pool4, and CSI\-pool16\.Table 6:Computational efficiency comparison under the same evaluation environment\. The CorrectCast row reports only the additional overhead introduced by the correction module\.
#### 5\.3\.7How Much Computational Overhead Does CorrectCast Introduce?
Table[6](https://arxiv.org/html/2611.22178#S5.T6)reports the computational costs of different nowcasting models\. CorrectCast trains only its correction module, requiring 14\.27 hours compared with 25\.40–50\.75 hours for the evaluated deep\-learning models\. It introduces only 0\.985 M trainable parameters, with 70\.29 ms inference latency and 184 MB peak GPU memory\. These results demonstrate that CorrectCast offers an efficient alternative to full\-model retraining when existing nowcasting models underperform\. PySTEPS is a non\-learning extrapolation method and is excluded from the training\-cost comparison\.
## 6Conclusion
We presented CorrectCast, a decoupled residual correction framework that refines completed nowcasts using historical QPE and PWV without updating base nowcasters\. Experiments across six nowcasters on MRB and CONUS demonstrate consistent improvements in categorical skill, spatially pooled CSI, and image quality\. Ablation and mechanism analyses support the proposed multimodal context selection and scale\-aware correction, while transfer experiments reveal both generalization potential and limitations\. Overall, CorrectCast offers a lightweight approach to improving existing nowcasts without retraining the base nowcaster\.
## LLM Usage Statement
The authors used ChatGPT\(OpenAI\) solely to assist in refining the language, improving readability, and ensuring clarity in various sections of the paper\. It is important to note that the LLM was not involved in the ideation, research methodology, or experimental design\. All research concepts, ideas, and analyses were developed and conducted by the authors\. The authors take full responsibility for the content of the manuscript\. We have ensured that the LLM\-generated text adheres to ethical guidelines and does not contribute to plagiarism or scientific misconduct\.
## References
- \[1\]Hylke E Beck, Albert IJM Van Dijk, Vincenzo Levizzani, Jaap Schellekens, Diego G Miralles, Brecht Martens, and Ad De Roo\.Mswep: 3\-hourly 0\.25 global gridded precipitation \(1979–2015\) by merging gauge, satellite, and reanalysis data\.*Hydrology and Earth System Sciences*, 21\(1\):589–615, 2017\.
- \[2\]Geoffrey Blewitt, William Hammond, and Corn Kreemer\.Harnessing the gps data explosion for interdisciplinary science\.*Eos*, 99\(2\):e2020943118, 2018\.
- \[3\]Zhihan Gao, Xingjian Shi, Hao Wang, Yi Zhu, Yuyang Bernie Wang, Mu Li, and Dit\-Yan Yeung\.Earthformer: Exploring space\-time transformers for earth system forecasting\.*Advances in Neural Information Processing Systems*, 35:25390–25403, 2022\.
- \[4\]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*, 36:78621–78656, 2023\.
- \[5\]Junchao Gong, Lei Bai, Peng Ye, Wanghan Xu, Na Liu, Jianhua Dai, Xiaokang Yang, and Wanli Ouyang\.Cascast: skillful high\-resolution precipitation nowcasting via cascaded modelling\.In*Proceedings of the 41st International Conference on Machine Learning*\. JMLR\.org, 2024\.
- \[6\]Vincent Le Guen and Nicolas Thome\.Disentangling physical dynamics from unknown factors for unsupervised video prediction\.In*Proceedings of the IEEE/CVF conference on computer vision and pattern recognition*, pages 11474–11484, 2020\.
- \[7\]David Kitzmiller, Dennis Miller, Richard Fulton, and Feng Ding\.Radar and multisensor precipitation estimation techniques in national weather service hydrologic operations\.*Journal of Hydrologic Engineering*, 18\(2\):133–142, 2013\.
- \[8\]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\.In*2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\)*, pages 17841–17850, 2025\.
- \[9\]Mengjie Liu, Weixing Zhang, Yidong Lou, Xingping Dong, Zhenyi Zhang, and Xiaohong Zhang\.A deep learning\-based precipitation nowcasting model fusing gnss\-pwv and radar echo observations\.*IEEE Transactions on Geoscience and Remote Sensing*, 2025\.
- \[10\]Cuixian Lu, Xindi Luo, Yuxin Zheng, Quanfei Wang, Jiafeng Li, and Zhuo Wang\.Rsg\-gan: A gan\-based precipitation nowcasting model integrating radar qpe, goes\-16 swd, and gnss ztds\.*IEEE Transactions on Geoscience and Remote Sensing*, 63:1–17, 2025\.
- \[11\]Zhifeng Ma, Hao Zhang, and Jie Liu\.Mm\-rnn: A multimodal rnn for precipitation nowcasting\.*IEEE Transactions on Geoscience and Remote Sensing*, 61:1–14, 2023\.
- \[12\]World Meteorological Organization\.Atlas of mortality and economic losses from weather, climate and water\-related hazards \(1970–2021\)\.2021\.
- \[13\]Young\-Jae Park, Doyi Kim, Minseok Seo, Hae\-Gon Jeon, and Yeji Choi\.Data\-driven precipitation nowcasting using satellite imagery\.In*Proceedings of the AAAI Conference on Artificial Intelligence*, pages 28284–28292, 2025\.
- \[14\]Seppo Pulkkinen, Daniele Nerini, Andrés A Pérez Hortal, Carlos Velasco\-Forero, Alan Seed, Urs Germann, and Loris Foresti\.Pysteps: An open\-source python library for probabilistic precipitation nowcasting \(v1\. 0\)\.*Geoscientific Model Development*, 12\(10\):4185–4219, 2019\.
- \[15\]Suman Ravuri, Karel Lenc, Matthew Willson, Dmitry Kangin, Remi Lam, Piotr Mirowski, Megan Fitzsimons, Maria Athanassiadou, Sheleem Kashem, Sam Madge, et al\.Skilful precipitation nowcasting using deep generative models of radar\.*Nature*, 597\(7878\):672–677, 2021\.
- \[16\]Chunlei Shi, Han Xu, Yinghao Li, Yi\-Lin Wei, Yongchao Feng, Yecheng Zhang, and Dan Niu\.Wavec2r: Wavelet\-driven coarse\-to\-refined hierarchical learning for radar retrieval\.In*Proceedings of the AAAI Conference on Artificial Intelligence*, pages 8951–8959, 2026\.
- \[17\]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*, 28, 2015\.
- \[18\]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*, 30, 2017\.
- \[19\]Changhoon Song, Teng\-Yuan Chang, and Youngjoon Hong\.Extreme weather nowcasting via local precipitation pattern prediction\.In*International Conference on Learning Representations*, pages 109146–109177, 2026\.
- \[20\]Joël Van Baelen, Mathieu Reverdy, Frédéric Tridon, Laurent Labbouz, Galina Dick, Michael Bender, and Maria Hagen\.On the relationship between water vapour field evolution and the life cycle of precipitation systems\.*Quarterly Journal of the Royal Meteorological Society*, 137\(S1\):204–223, 2011\.
- \[21\]Yunbo Wang, Zhifeng Gao, Mingsheng Long, Jianmin Wang, and Philip S Yu\.Predrnn\+\+: Towards a resolution of the deep\-in\-time dilemma in spatiotemporal predictive learning\.In*International conference on machine learning*, pages 5123–5132\. PMLR, 2018\.
- \[22\]Yunbo Wang, Haixu Wu, Jianjin Zhang, Zhifeng Gao, Jianmin Wang, Philip S Yu, and Mingsheng Long\.Predrnn: A recurrent neural network for spatiotemporal predictive learning\.*IEEE Transactions on Pattern Analysis and Machine Intelligence*, 45\(2\):2208–2225, 2022\.
- \[23\]Binqing Wu, Weiqi Chen, Shiyu Liu, Zongjiang Shang, Haiou Wang, Liang Sun, and Ling Chen\.Mocast: Learning turbulent motions under physical guidance for precipitation nowcasting\.In*Proceedings of the AAAI Conference on Artificial Intelligence*, pages 15950–15958, 2026\.
- \[24\]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\.In*2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\)*, pages 27758–27767, 2024\.
- \[25\]Jian Zhang, Kenneth Howard, Carrie Langston, Brian Kaney, Youcun Qi, Lin Tang, Heather Grams, Yadong Wang, Stephen Cocks, Steven Martinaitis, Ami Arthur, Karen Cooper, Jeff Brogden, and David Kitzmiller\.Multi\-radar multi\-sensor \(mrms\) quantitative precipitation estimation: Initial operating capabilities\.*Bulletin of the American Meteorological Society*, 97\(4\):621 – 638, 2016\.
- \[26\]Yuchen Zhang, Mingsheng Long, Kaiyuan Chen, Lanxiang Xing, Ronghua Jin, Michael I Jordan, and Jianmin Wang\.Skilful nowcasting of extreme precipitation with nowcastnet\.*Nature*, 619\(7970\):526–532, 2023\.
- \[27\]Kun Zheng, Long He, Huihua Ruan, Shuo Yang, Jinbiao Zhang, Cong Luo, Siyu Tang, Jiaolong Zhang, Yugang Tian, and Jianmei Cheng\.A cross\-modal spatiotemporal joint predictive network for rainfall nowcasting\.*IEEE Transactions on Geoscience and Remote Sensing*, 62:1–23, 2024\.相似文章
RECAST:用于粗网格PDE求解器校正与超分辨率的机器学习框架
RECAST是一个机器学习框架,结合了学习校正与超分辨率技术,以恢复粗网格PDE求解器的精度,在六个一维PDE系统上将误差降低50-92%,从而能够在保持保真度的同时使用更粗的网格进行模拟。
SimCast-S2S:一种通过气候模拟迁移学习实现高效亚季节性降水预报的生成模型
SimCast-S2S是一种用于概率性亚季节性降水预报的生成式潜扩散框架,它利用气候模拟的迁移学习,在性能上超越了深度学习基线模型,并与业务预报系统具备竞争力。
AirCast-SR: 基于潜在一致性扩散的公里尺度大气超分辨率基础模型
AirCast-SR 是一个基于扩散模型的基础模型,能够将全球AI天气预报从0.25°降尺度到1公里分辨率,每小时生成一次预报,可产生67小时预报,具有接近零偏差和结构真实感,同时在单个商用GPU上仅需数分钟即可完成推理。
极端降水降尺度的多分位数回归
介绍了Q-SRDRN,一种使用分位数损失的多分位数超分辨率网络,用于改进极端降水降尺度,在保持整体精度的同时,显著提高了强降雨事件的检测率。
时间序列的检索校正共形预测
介绍了检索校正共形预测(RCCP),一种用于时间序列预测区间的检索增强校准方法,它选择相似的过去残差并应用标量共形校正,以低开销实现目标覆盖率。