acceptodds
Under review as a conference paper at ICLR 2027

MIRAGE: Multimodal Radar-Conditioned Graph Learning for Rainfall Interpolation and Forecasting

Abstract

Precipitation interpolation and forecasting over sparse weather station networks are challenging because rainfall exhibits highly localized and non-stationary dynamics, while ground observations are spatially limited. We propose MIRAGE, a radar-conditioned dynamic graph framework that integrates dense radar fields with sparse station observations for precipitation estimation at locations without local precipitation histories. MIRAGE constructs station-specific multimodal representations through cross-attention and uses them to adapt graph computation in two complementary ways: Dynamic Graph Construction determines where information should propagate, while a Meteorological Hypernetwork determines how propagated information should be transformed. Experiments on KMA and MeteoNet datasets show that MIRAGE achieves strong performance in both spatial interpolation and multi-step forecasting, with particularly consistent improvements at unobserved stations. MIRAGE also maintains competitive performance as observation sparsity increases. Ablation studies further demonstrate the complementary contributions of radar conditioning, cross-modal attention, and adaptive graph computation.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Precipitation Estimation, Precipitation Forecasting, Multimodal Learning, Graph Neural Networks, Weather Radar, Spatial Interpolation, Spatiotemporal Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_SpZ3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces MIRAGE, which addresses rainfall interpolation and short-term forecasting at stations without local precipitation histories. It employs a frozen pretrained radar encoder and station-specific cross-attention to fuse radar sequences, available station observations, coordinates, and temporal features into multimodal contexts. Experiments on KMA and two MeteoNet regions cover interpolation, forecasting up to 60 minutes, and robustness to missing station observations, yielding particularly strong mean results at target stations excluded from precipitation inputs and training supervision.

Critical strengths

1.The focus on stations lacking local precipitation histories is practically important and underexplored.

2.Using radar as external meteorological context to condition graph computation, rather than merely concatenating it as a node feature, is a sensible and novel design choice.

3.The evaluation spans three datasets, multiple forecast horizons, sparsity sweeps, and ablations, providing thorough empirical support.

Critical weaknesses

1.It is unclear whether the learning baselines receive comparable radar inputs or pretrained radar representations. Since MIRAGE additionally uses radar, the headline comparisons cannot separate the contribution of radar information from that of the proposed adaptive graph mechanisms.

2.The exact replacement in "w/o HyperNet" is unspecified, making it unclear whether the degradation stems from losing context-conditioned transformation or merely from reduced parameters. Moreover, Eqs. (21)–(23) allow target context to bypass graph messages entirely; a target-context-only control with the same radar representation is needed to quantify the net contribution of propagation.

3.Since precipitation stations are spatially correlated, randomly held-out targets likely retain observed neighbors from which graph propagation can easily borrow information, making the task easier than the practically relevant scenario. A spatially clustered holdout (e.g., withholding an entire subregion) is needed to test generalization to genuinely unobserved areas. 4.Figure 1's caption references "KMA and MeteoNet," but the figure only shows MeteoNet; Table 3 contains a typo in the RMSE column header for the 60-min KMA case ("RMSE (↑)" instead of "RMSE (↓)"); and the All Nodes vs. Unobserved Nodes comparison in Table 2 is largely redundant with Table 3.

Questions for authors

1.Do the learning baselines receive comparable radar inputs or pretrained radar representations? Since MIRAGE additionally uses radar, it is unclear whether the headline comparisons can separate the contribution of radar information from that of the proposed adaptive graph mechanisms.

2.What exactly replaces the Hypernetwork in the "w/o HyperNet" variant? Is it a normally trained, input-independent message transformation with an otherwise comparable architecture? Without this specification, it is unclear whether the degradation stems from losing context-conditioned transformation or merely from reduced parameters. Furthermore, since Eqs. (21)–(23) allow target context to bypass graph messages entirely, could a target-context-only control with the same radar representation help quantify the net contribution of propagation?

3.Why are the missing-station experiments limited to random splits? Since precipitation stations are spatially correlated, randomly held-out targets likely retain observed neighbors from which graph propagation can easily borrow information, making the task easier than the practically relevant scenario. Would a spatially clustered holdout (e.g., withholding an entire subregion) better test generalization to genuinely unobserved areas?

4.Is the main figure AI-generated? It contains several visual artifacts and inconsistencies that give the impression of being directly produced by an AI image generator, which raises concerns about the overall presentation quality.

Reviewer_DiSkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Generally precipitation forecasting is an extremely challenging field that from what I can tell deep learning has struggled with substantially in the past. In this worrk, the authors propose a frozen, pretrained radar encoder that is fused with rain-gauge station data through cross-attention, giving each station a "multimodal context". That context drives a dynamic mask on a kNN graph (where messages go) and a hypernetwork that sets the message weights (how they are transformed). The model is evaluated on data from Korea and France for rainfall interpolation and 10–60 min forecasting at 30% held-out stations.

Critical strengths

Strengths:

  • Clear pretraining objective using strict station-level holdout, with random station masking during training
  • it's a pretty good and well motivated idea to use radar as conditioning context, not as a direct Z–R substitute.
  • I like how many meaningful experiments were carried out to test the approach e.g. three regions, a 10–70% sparsity sweep, regional blackout tests, ablations, CSI by intensity, fast inference, ...
Critical weaknesses

Weaknesses:

  • I don't have a full overview on which domain baselines are the most relevant but would have expected to read about Z–R with bias correction, persistence for nowcasting, IGNNK, GRIN and SATCN, which the paper cites but doesn't compare against, ...
  • Not sure about the inputs to the baselines: Which of the baselines actually also get radar input? Plain kriging nearly ties MIRAGE (KMA RMSE 1.461 vs 1.442).
  • Overall the novelty feels rather incremental novelty: adaptive graphs, hypernetworks and cross-attention fusion are all known. The hypernetwork uses one globally pooled context, so it doesn't adapt locally. Units of the metrics are unclear (they look like log scale), and NaNs are set to 0.
Questions for authors

Questions:

  • Do the baselines get the same radar input?
  • Are RMSE/MAE in mm/h after inverting the log transform?
  • Which years was the radar encoder pretrained on, do they overlap the test years?
Reviewer_PUArOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies contemporaneous precipitation interpolation and short-range multi-step forecasting when a subset of weather stations has no local precipitation history. MIRAGE uses station-to-radar cross-attention to construct station-level multimodal contexts, modulates propagation weights on a fixed geographical kNN graph with a dynamic mask, and uses a hypernetwork to generate sample-conditioned message-transformation parameters. Experiments on KMA, MeteoNet-NW, and MeteoNet-SE use fixed held-out stations and evaluate interpolation, future forecasting, increasing spatial sparsity, and component ablations. MIRAGE reports generally competitive mean RMSE and MAE at unobserved nodes, but the current evaluation does not adequately disentangle the benefit of additional radar information, model capacity, and the proposed adaptive graph mechanisms.

Critical strengths
  1. The problem setting is practically relevant, and the target-station holdout is reasonably strict. Sections 2.1, 2.7, and 3.3 define a fixed unobserved set V_u before training and exclude its ground-level precipitation histories throughout training, validation, and testing; its labels are also excluded from the training loss. This is a more meaningful test of spatial generalization than masking target stations only at test time. Additional random masking of observed stations during training is also consistent with the intended setting.
  2. The architecture has a clear functional decomposition. Equations (2)–(4) and Appendix Equations (13)–(23) assign where information propagates to the dynamic graph mask and how propagated messages are transformed to the hypernetwork. The gated fusion of a target's radar, coordinate, and temporal context with propagated station information is well motivated for locations without local histories.
  3. The experiments cover multiple regions, tasks, and forecast horizons. Table 1 evaluates contemporaneous interpolation at unobserved stations on KMA, MeteoNet-NW, and MeteoNet-SE. Table 2 evaluates 30- and 60-minute forecasts on the two MeteoNet regions, while Table 3 evaluates 10–60-minute forecasts on KMA. The forecasting tables distinguish All Nodes from Unobserved Nodes, and the paper additionally studies spatial sparsity, component ablations, regional block masking, CSI, and inference latency.
  4. The reported mean performance at unobserved nodes is reasonably consistent. MIRAGE obtains the lowest interpolation RMSE and MAE in all three regions in Table 1 and the best mean KMA results at unobserved nodes across all forecasting horizons in Table 3. It is also generally strongest at unobserved MeteoNet nodes in Table 2. The paper appropriately acknowledges that MIRAGE is not uniformly best on All-Node RMSE for MeteoNet-NW.
Critical weaknesses
  1. The main baseline comparison does not establish input-modality or task-adaptation parity The paper states that all methods share the same station partitions, temporal splits, and evaluation protocol, but it does not specify, model by model, whether the baselines receive radar observations, station histories, coordinates, temporal features, and missingness indicators. MIRAGE can access radar covering the target locations, whereas Kriging is explicitly described as using spatial covariance among observed stations. Appendix D likewise does not describe how the other baselines ingest radar representations. SINR and SUICA are introduced in Section 3.2 as neural spatial reconstruction methods but are evaluated for future multi-step forecasting in Tables 2–3. Conversely, forecasting models such as ReDYNet, DualCast, MIGN, and STGFormer are evaluated for contemporaneous interpolation in Table 1. Except for a brief description of STGFormer, the paper does not sufficiently document the input heads, output heads, history windows, and objectives used for these cross-task adaptations. Tables 1–3 can therefore be read as comparisons among end-to-end systems, but the current evidence does not support attributing their differences specifically to DynMask, the hypernetwork, or cross-attention. Per-model input and adaptation details, together with modality-matched controls, are needed.
  2. The isolation protocol for radar pretraining and missing-data handling is incomplete Appendix A.1 states that the raw KMA radar and AWS data span 2014–2023, whereas the experiments use June–September data from 2019–2023 with a 2019–2021/2022/2023 train/validation/test split. These statements are not contradictory. The unresolved issue is that the paper does not report which years were used for self-supervised radar-encoder pretraining, station filtering by missingness, fitting normalization statistics, or selecting the pretrained checkpoint. Leakage would occur only if validation or test years were used in those steps, but the current manuscript does not establish that they were restricted to the training period. The raw NaNs in both KMA and MeteoNet are replaced with zero, while zero is also the physically meaningful value for no precipitation. The main text mentions a binary missingness indicator for fixed target stations, but Appendix E.2 and Table 7 do not clearly show where this indicator enters the model. The paper also does not state whether ordinary raw NaNs have their own mask or whether zero-filled missing labels are excluded from the training loss and evaluation. This may conflate missing observations with genuine dry conditions. True zero rainfall should be retained, while originally missing labels should be excluded with an explicit mask.
  3. The HyperNet ablation does not yet support the claimed mechanism Table 4 reports higher RMSE in all 20 settings after removing HyperNet, indicating that this component is strongly associated with performance in the current implementation. However, the paper does not define the operator that replaces the hypernetwork-generated W_t and b_t in the w/o HyperNet variant, nor does it report trainable parameter counts, FLOPs, learning curves, or variation across random initializations for the ablations. The w/o HyperNet variant is even worse than Precip-only in every reported setting. These two rows change both Radar and HyperNet, however, so they cannot support a single-factor conclusion such as radar being harmful without HyperNet. The pattern may instead reflect module interactions, optimization differences, or the particular ablation implementation. Without a capacity-matched static transformation, an ordinary MLP control, a complete factorial ablation, and training diagnostics, the experiment establishes only that the current system depends on HyperNet. It does not isolate sample-conditioned meteorological adaptation from capacity, optimization, or replacement-operator effects. Table 4 also reports no uncertainty or significance analysis. If the authors wish to support this attribution with a physical mechanism, they should at least provide a visualization or quantitative test showing that the learned graph or attention directions agree with precipitation movement; directly removing a component cannot establish that capability.
  4. The physical interpretation of DynMask and the module attribution in the appendix are not adequately supported Equation (13) defines DynMask using both multimodal-context similarity and a relative-geometry encoding. Equation (14) defines the effective dynamic adjacency as: A_dyn,t = Norm(A_base ⊙ M_t ⊙ S_t) Because A_base is a fixed inverse-distance kNN graph, and absent an unreported thresholding operation, the sigmoid mask primarily reweights edges within a fixed candidate support; S_t masks unavailable sources. The method cannot create propagation edges outside that support. Appendix I Figure 9 visualizes the raw mask M_t rather than the effective adjacency A_dyn,t. It provides only one station-index matrix per dataset, without identifying the sample, time, graph head, or aggregation procedure, and without overlaying radar echoes, wind vectors, or HyperNet parameters. The figure shows that the learned masks differ numerically from the predefined adjacency, but it does not demonstrate temporal adaptation to weather systems or physical alignment with wind and advection. Because Equation (13) already contains an explicit geometry term, the observed heatmap structure may also arise primarily from positional bias. Appendix G further attributes online reconstruction of the adjacency and edge weights to HyperNet and states that this occurs without a fixed spatial prior. This is inconsistent with the role assignment in the main method and Appendix E: Equation (14) constructs adjacency from the fixed A_base and DynMask, whereas Equation (16) uses HyperNet to generate message-transformation parameters rather than adjacency. The authors should clarify whether the regional experiment used an undocumented implementation and revise the attribution.
  5. The evaluation and statistical reporting contain issues that affect confidence in the conclusions The main RMSE and MAE results are computed over all samples. When dry cases dominate, unconditional averages may obscure predictive differences on wet events. Appendix F adds CSI, but the reported numerical advantage is limited. I therefore recommend reporting MSE, together with the corresponding RMSE and MAE, conditioned on ground-truth precipitation greater than zero, to assess forecasting performance during actual precipitation events more directly. Appendix Table 8 reports an All-Node 60-minute MAE of 0.342 ± 0.8530 for STGFormer and states that the statistics are computed over four runs. For four nonnegative MAE values with mean 0.342, the maximum possible sample standard deviation is 2 × 0.342 = 0.684; therefore, 0.8530 cannot be the standard deviation of the stated four runs. Several other baseline standard deviations also increase abruptly at 60 minutes, so the authors should verify the table and provide the per-run results.
Questions for authors
  1. Baseline inputs and adaptations Please list the exact inputs used by each baseline, including radar, station histories, coordinates, temporal features, missingness indicators, and history lengths. How were SINR and SUICA adapted for multi-step forecasting, and how were ReDYNet, DualCast, MIGN, and STGFormer adapted for contemporaneous interpolation? If some baselines do not use radar, can the authors add station-only MIRAGE, a radar-only model, a simple radar-plus-station fusion model, and radar-augmented baselines?
  2. Data isolation and missing values Which years and regions were used to pretrain the frozen radar encoder? Were station filtering, normalization statistics, and checkpoint selection restricted strictly to the training period? Do zero-filled raw NaNs have a separate missingness mask, and are their labels excluded from the loss and evaluation? Where does the binary missingness indicator enter the network?
  3. HyperNet attribution What operator replaces W_t and b_t in the w/o HyperNet variant? What are the trainable parameter counts and computational costs of the ablation variants? Can the authors add a capacity-matched static transformation or MLP, a complete Radar × HyperNet × DynMask factorial ablation, learning curves, and variation across random initializations?
  4. Dynamic graph and physical interpretation Which sample, time, and graph head does Figure 9 represent, and is it averaged across samples or time? Can the authors visualize the effective A_dyn,t geographically over time and quantify its directional agreement with observed winds, radar optical flow, or rain-band advection? How should Appendix G's attribution of adjacency reconstruction to HyperNet without a fixed prior be reconciled with Equations (14) and (16)?
  5. Statistics and rainfall-event evaluation Please verify the standard deviations in Table 8 that cannot hold under the stated definition or increase anomalously, and provide the four per-run values. Is Figure 3 evaluated on All Nodes or Unobserved Nodes? Can the authors report across-split uncertainty, category support, and wet-event continuous errors for Table 1, Table 4, Figure 3, and Figure 6?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.