acceptodds
Under review as a conference paper at ICLR 2027

PatchShot: Rare Precipitation Events Are Encoded Yet Hidden from Standard Readouts

Abstract

Rare precipitation events are difficult to recognize under severe label scarcity. We show that frozen 3DMAE radar representations already encode discriminative evidence for high-threshold events, yet standard readouts fail to recover it, revealing a representation accessibility gap. We introduce PatchShot, a case-disjoint inductive few-shot framework for parameter-efficient representation re-access. PatchShot freezes the weather encoder, learns a lightweight adapter through support-supervised contrastive learning, constructs support-derived prototypes, and calibrates the decision threshold using support data alone. Query patch scores are aggregated into case-level predictions without using query labels. We further propose PatchShot-Env, which conditions the re-access adapter on matched ERA5 low-mode context through lightweight expert routing. Across precipitation thresholds and support budgets, PatchShot substantially improves over the frozen QPE output and direct prototype readouts, while matched environmental routing provides further gains and shuffled-context controls weaken them. Probe and intervention analyses confirm that the improvement reflects recovered access to encoded rare-event evidence rather than encoder retraining.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Representation accessibility, Few-shot learning, Rare-event recognition, Contrastive learning, Weather radar

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_j8AYOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies rare precipitation recognition under limited supervision from the perspective of representation accessibility. The authors argue that frozen 3DMAE radar representations already retain useful evidence for high-threshold precipitation events, while standard readouts fail to effectively expose it. To address this, they propose PatchShot, which keeps the weather encoder frozen and learns a lightweight support-supervised contrastive adapter, followed by prototype-based prediction and support-only threshold calibration. An extension, PatchShot-Env, further conditions the readout on matched ERA5 environmental context through lightweight expert routing. Experiments across different precipitation thresholds and support budgets show improved rare-event recognition, with additional probe and intervention analyses used to support the representation re-access interpretation.

Critical strengths

The paper studies a problem of rare precipitation recognition under limited supervision and provides a clear distinction between representation failure and readout failure. The proposed PatchShot framework is simple and parameter-efficient, keeping the pretrained weather encoder frozen while adapting only a lightweight readout. I also appreciate the case-disjoint evaluation protocol and the effort to analyze the proposed “accessibility gap” beyond predictive performance.

Critical weaknesses

The main limitation is that the algorithmic novelty appears relatively modest. PatchShot largely combines supervised contrastive adaptation, prototype-based classification, and support-set calibration, and its advantage over some strong support-aware baselines is fairly small. In addition, the interpretation of the results as “recovering hidden information” is somewhat stronger than what the experiments directly establish, since the accessibility measure depends on the chosen readout family. For PatchShot-Env, it is also not entirely clear how much of the improvement comes from better access to radar representations versus additional predictive information introduced by ERA5 context. Finally, the method uses patch-level supervision derived from precipitation targets, so the practical meaning of the claimed few-shot setting could be clarified further.

Questions for authors

Could the authors clarify whether PatchShot requires dense patch-level precipitation labels for the support cases, and how the method would perform if only case-level rare-event labels were available?

For PatchShot-Env, can the authors disentangle the benefit of improved radar representation access from the additional predictive information directly provided by ERA5, e.g., with an ERA5-only or simple radar–ERA5 fusion baseline?

1 comment
AuthorsOpenReview ↗

Supervision and Environmental Contributions

We thank the reviewer for recognizing the distinction between representation failure and readout failure.

Scope and contribution. We agree that PatchShot's advantage over strong support-aware baselines is modest and that accessibility depends on the readout family. Our contribution is not a new algorithmic component but a controlled study of whether readout failure necessarily means that rare-event information is absent from a frozen weather representation; PatchShot is a lightweight instrument for this study. The empirical envelope in Table 6 is defined by three probes, not an information-theoretic bound. “Hidden” refers only to evidence not exposed by the original QPE head or direct prototype readout. Matched CE also improves accessibility (0.981/0.987 versus 0.981/0.996 for PatchShot), indicating that task-aligned supervision, rather than a uniquely contrastive mechanism, drives much of the improvement.

Q1. Does PatchShot require dense patch-level labels?

No manual patch annotation is required. Precipitation targets originate from regional automatic weather station (AWS) observations provided by CMDSC. After quality control, observations from 2,856 stations are interpolated via Kriging into 700 × 900 precipitation grids spanning 108.50°–117.50°E and 19.04°–26.04°N. Patch event labels are generated automatically using Eq. (2).

“Few-shot” refers to 19–959 labeled radar support cases, each paired with a spatial precipitation target, not to a small total number of annotated patches. Gridded targets are used to derive labels for support cases; query targets are withheld from training and inference and used only for evaluation. This is denser supervision than a single binary event label per case. The gridded targets permit spatial alignment of radar structure and rainfall, although interpolation introduces uncertainty. Learning from case-level labels alone, e.g., through multiple-instance learning, remains future work.

Q2. Radar accessibility versus additional ERA5 information

We agree that ERA5 supplies additional predictive information, which must be distinguished from improved access to radar representations. Table 4 already includes simple radar–ERA5 fusion. At H5/H10, radar-only PatchShot achieves CSI of 23.29/16.65, simple fusion 30.27/24.30, PatchShot-Env 30.95/25.59, and case-shuffled routing 30.01/23.58.

The substantial gain from simple fusion supports an important contribution from additional environmental inputs. Moreover, ERA5-conditioned adapters use 3,000 separate development cases, so the total gain cannot be attributed to environmental inputs alone. Routing adds a modest improvement over simple fusion; case-shuffled routing performs worse, consistent with the usefulness of case-aligned conditioning. An ERA5-only readout, which we have not yet evaluated, would further isolate the independent contribution of atmospheric context. Neither comparison establishes that the environmental gains arise from improved radar representation access. We therefore ground our accessibility conclusions in radar-only experiments and present PatchShot-Env as a complementary extension.

ERA5 provides humidity, temperature, winds, and geopotential (q, t, u, v, z) at pressure levels, but it is a delayed reanalysis product. Real-time AWS, radiosondes, and NWP analyses or forecasts are potential operational substitutes; whether they reproduce these gains remains unverified.

We also examine AGRI geostationary satellite observations. In Table 5, matched cloud states outperform month-shuffled controls, suggesting useful case-aligned information, but gains over the unconditioned readout are modest and nonuniform (H5 CSI: 28.94 versus 28.43). We treat AGRI as a complementary extension, not part of the final model.

Summary. Radar-only experiments support readout-dependent recoverability under a restricted empirical accessibility measure. Spatial precipitation targets provide support-case supervision; ERA5 and AGRI add complementary context. We do not attribute their gains exclusively to improved radar access.

Reviewer_usAjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether rare precipitation information retained in pretrained weather representations can be recovered without retraining the encoder. It introduces PatchShot, a few-shot framework that uses a lightweight adapter trained with supervised contrastive learning, and PatchShot-Env, which incorporates ERA5 atmospheric information through expert routing. Experiments show improvements over standard readouts, supported by probing and intervention analyses.

Critical strengths
  1. The paper provides empirical evidence that pretrained weather representations retain discriminative information about rare precipitation events even when the original QPE readout fails, highlighting the practical importance of representation accessibility under limited label availability.
  2. The paper provides meaningful experimental analyses, including controlled comparisons with matched CE and shuffled-label training (Table 2) to isolate the role of task-aligned supervision, and support-removal interventions (Figures 4 and 6) to demonstrate that influential support examples materially affect final predictions.
Critical weaknesses
  1. Limited methodological novelty and unclear advantage over existing methods. PatchShot performs similarly to matched CE and PaCo, while strong baselines are not compared at the smallest support budgets. Moreover, the ablations show the importance of task supervision but do not clearly establish what the proposed accessibility framework reveals beyond conventional supervised readout adaptation.
  2. Limited evaluation scope. Experiments rely on a single regional radar dataset and one frozen encoder. It remains unclear whether the observed accessibility gap and the benefits of PatchShot generalize to other datasets or pretrained representations.
Questions for authors
  1. How does PatchShot compare with matched CE and PaCo at 1% and 5% support under identical evaluation conditions? These comparisons would help clarify whether PatchShot offers a distinctive advantage under the severe label scarcity that motivates the paper.
  2. Does the observed accessibility gap and its reduction through adaptation persist on another dataset or with a different pretrained encoder? Evidence from either setting would help assess whether the central finding generalizes beyond the current experimental setting.
3 comments
AuthorsOpenReview ↗

Support-Budget Comparisons and Generalization of Representation Accessibility

We thank the reviewer for recognizing the evidence of recoverable rare-event information and the value of our controlled supervision and support-removal analyses. We agree that label scarcity, strong baselines, and generalization require careful examination, and have strengthened the analysis accordingly.

Q1. How does PatchShot compare with Matched CE and PaCo at 1% and 5% support?

The revised Figure 3 compares radar-only and ERA5-conditioned readouts across five support budgets (1%–50%) using frozen 3DMAE representations and a case-disjoint evaluation protocol.

At 1% support (19 cases), PatchShot achieves H5/H10 CSI of 9.86/6.87, compared with 12.28/8.33 for Matched CE, 13.07/8.72 for PaCo-Proto, 15.62/10.77 for frozen prototypes, and 15.45/8.90 for Balanced Linear. Thus, PatchShot does not outperform these simpler or alternative readouts under extreme label scarcity.

At 5% support (96 cases), PatchShot, Matched CE, and PaCo-Proto achieve comparable H5 CSI (18.18/18.38/17.40) and H10 CSI (12.50/12.64/12.56), respectively. Balanced Linear remains competitive (19.18/13.59).

Figure 3 shows that adaptation is unstable at 1% support, while gains relative to the frozen-prototype reference become more consistent with additional supervision. At 20% and 50% support, all five displayed PatchShot seeds exceed this reference at both thresholds. We therefore do not claim universal superiority under severe label scarcity.

At the standard operating point, Table 2 compares adaptation objectives under matched architecture, optimization, prototype decoding, and calibration. Matched CE achieves H5/H10 CSI of 22.55/16.84, versus 23.29/16.65 for PatchShot. Shuffled-label SupCon returns H5 CSI close to frozen prototypes (16.64 vs. 16.02) with lower AP (17.50 vs. 21.62). These controls indicate that task-aligned supervision, rather than adapter capacity or a uniquely necessary contrastive objective, drives much of the recovered discrimination.

Beyond predictive comparisons, our accessibility framework distinguishes readout failure from representation information loss. Multiple readout families recover rare-event discrimination despite zero CSI from the original QPE head. Table 6 quantifies empirical accessibility: at H5/H10, the accessibility fraction increases from 0.788/0.823 for the frozen representation to 0.981/0.996 after radar-only PatchShot adaptation, while Matched CE reaches 0.981/0.987. This demonstrates that improved accessibility is not unique to contrastive learning.

The support-derived prototype readout further enables exact decision-level interventions without retraining the adapter. Removing the 10 most influential of 19 support cases flips approximately 40% of true-positive query patches, versus fewer than 8% under random removal (Figures 4 and 6). This reveals dependence on historical support evidence without implying physical causality.

Q2. Does the accessibility gap generalize to other datasets or encoders?

We agree that broader validation would strengthen generality. Our experiments involve one regional 3D radar dataset and one frozen 3DMAE encoder; cross-dataset and cross-encoder generalization therefore remains unverified.

Rare precipitation events are difficult to collect. Even with radar observations covering a 900 × 700 km region, high-threshold events remain scarce, motivating the study of representation accessibility under limited supervision.

The accessibility gap may vary across weather regimes and observation conditions, including terrain and radar sampling geometry. Where event-relevant information is genuinely missing from observations, readout adaptation alone cannot recover it.

Within the evaluated dataset, Figure 7 audits temporal dependence: two-hour blocking changes radar-only PatchShot H5/H10 CSI from 0.2128/0.1757 to 0.2278/0.1630. Discrimination remains recoverable, although storm-level independence is not established.

ERA5-conditioned variants use adapters trained on 3,000 separate development cases; their performance with 19 local support cases should not be interpreted as training the entire model from only 19 labels.

Future work will extend representation accessibility to precipitation retrieval under environmental shifts, using sparse matching of historical rare-event evidence for lightweight adaptation across regions and weather regimes.

Summary. We do not claim universal few-shot superiority or cross-encoder generalization. Our contribution is a controlled study of representation accessibility: rare-event evidence can remain recoverable from frozen weather representations even when the original QPE readout fails. This provides a basis for diagnosing when to adapt the readout, incorporate additional observations, or improve the encoder.

Thank you for the response. The additional results clarify the performance under low support budgets. I also understand that the main contribution is to demonstrate the recoverability of rare-event evidence from frozen weather representations, rather than to establish generalization across datasets or encoders.

However, I remain concerned about the broader scientific takeaway of this accessibility study. While the experiments confirm that rare-event signals remain recoverable, it is still unclear what actionable insights or guiding principles this provides for future work beyond conventional readout adaptation, or how these findings translate into a sufficiently distinctive methodological contribution for PatchShot.

Overall, I appreciate the additional analyses and more carefully scoped claims. However, because these core concerns regarding scientific insight and methodological novelty remain unaddressed, I will maintain my original score.

AuthorsOpenReview ↗

Scientific Takeaways and Practical Implications

We thank the reviewer for the follow-up. Our main takeaway is that, for rare-precipitation retrieval, 3D radar provides the primary evidence, ERA5 atmospheric context the largest complementary gain, and AGRI cloud-top states the least in our evaluated setting, and that aligned environmental context becomes more important as precipitation intensity increases. The analyses below support this interpretation and provide practical guidance on observation selection, event intensity, and readout design. All CSI values are reported as percentages (%), and absolute CSI differences as percentage points (pp).

1. Observation sources provide different kinds of precipitation evidence.

  • 3D radar provides the primary precipitation-sensitive representation. Although the original QPE head achieves zero event CSI, lightweight readouts recover 23.29%/16.65% CSI at H5/H10 from the same frozen radar features.
  • ERA5 provides substantial complementary atmospheric context. Simple radar–ERA5 fusion improves CSI by 6.98/7.65 pp (approximately 30%/46% relative improvement; Table 4). However, the ERA5-conditioned adapters also use 3,000 development cases, so these gains cannot be attributed solely to additional environmental variables.
  • AGRI provides more limited additional information in our evaluated setting. Matched cloud states outperform shuffled controls but improve H5 CSI by only 0.51 pp over the corresponding unconditioned readout, with no consistent H10 improvement (Table 5).

These observations are physically consistent with radar sensing precipitation structures, ERA5 characterizing the atmospheric column, and AGRI primarily providing cloud-state information. They motivate radar-centered, observation-aware fusion rather than treating all sources as interchangeable. We do not claim a universal sensor ranking, since ERA5-only was not evaluated and the fusion experiments use different settings.

2. Environmental sensitivity changes with precipitation intensity.

The relative improvement from radar-only to simple ERA5 fusion increases from approximately 30% at H5 to 46% at H10. Compared with case-shuffled routing, aligned routing improves CSI by 0.94 pp at H5 and 2.01 pp at H10. These results suggest greater sensitivity to properly aligned environmental context at the higher event threshold.

We also observe an interesting tendency across environmental variable groups: thermodynamic routing (humidity and temperature) performs best at H5 (31.16% CSI), while at H10, dynamic routing (winds and geopotential; 25.34% CSI) and broad low-order vertical modes (25.59% CSI) outperform thermodynamic routing (25.13% CSI). This suggests a possible shift in the relative usefulness of thermodynamic and dynamical information with event intensity. However, these differences are within approximately one standard deviation, so we regard this as a hypothesis for further investigation rather than an established physical mechanism.

3. Readout complexity should depend on label availability.

At 5% support, PatchShot improves over frozen prototypes by 2.31 pp at H5 but only 0.74 pp at H10. Under the smallest support budgets, simpler readouts can be more competitive. Thus, a practical lesson is to evaluate lightweight frozen-feature probes before adopting more complex supervised adaptation, particularly for very rare events.

4. Support dependence can be diagnosed explicitly.

PatchShot's support-conditioned prototypes make predictions traceable to historical cases. Removing the 10 most influential among 19 support cases flips approximately 40% of true-positive query patches, compared with under 8% under random removal (Figures 4 and 6). This reveals strong dependence on particular support examples, although it does not establish a mechanism unique to PatchShot.

Overall takeaway. Our findings suggest a practical strategy for rare-event weather learning: first diagnose representation versus readout failure; then identify complementary observations; and finally adapt readout and environmental conditioning to event intensity and label availability.

Reviewer_wbRBOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This manuscript focuses on rare-precipitation information remains recoverable from frozen 3DMAE radar representations when the original prediction head performs poorly. PatchShot learns a lightweight adapter using support-supervised contrastive learning, following by prototype construction and threshold calibration. PatchShot-Env additionally incorporates ERA5 environmental context throught export routing. Experiments at two precipitation thresholds, together with component ablations, probe comparisons, and support-removal interventions, investigate the distinction between information retained in a representation and information accessible to a particular readout.

Critical strengths

1.The distinction between representation quality and readout effectiveness is relevant to adapting pretrained weather models. The study provides useful empirical evidence that poor precipitation-head performance does not imply an absence of discriminative information in the frozen features.

2.The comparisons include meaningful supervised alternatives, and the matched CE ablation helps clarify that task-specific supervision, rather than a uniquely effective contrastive objective, drives much of the improvement.

Critical weaknesses

1.The contribution beyond established supervised adaptation is limited. PatchShot combines a lightweight adapter, supervised contrastive learning, prototype classification, and threshold calibration. Table 2 shows that matched CE achieves similar performance, while Table 1 and 3 show that balanced linear and PaCo readouts are competitive. These results support the usefulness of supervised adaptation, but leave unclear what additional methodological or scientific insight PatchShot contributes beyond that established observation.

2.The accessibility analysis does not sufficiently establish a distinct explanation for the gains. The empirical envelope is the best performance among three probes, and training an adapter with a prototype-compatible objective naturally makes the prototype readout more competitive. Likewise, removeing influential support examples demonstrates dependence on prototype construction, but does not distinguish the proposed “re-access” account from ordinary supervised feature adaptation. The study therefore offers a plausible interpretation without clearly showing what new understanding that interpretation provides.

Questions for authors

1.Given the competitive balances linear, matched CE, and PaCo results, what is the principal new finding beyond the benefit of task-specific supervision on frozen features?

2.What conclusion from the probe and intervention analyses distinguishes representation “re-access” from conventional supervised adaptation? Please clarify which evidence supports that distinction.

2 comments
AuthorsOpenReview ↗

Representation Accessibility Beyond Supervised Adaptation

We thank the reviewer for the careful critique. We agree that improved CSI alone cannot distinguish representation re-access from ordinary supervised adaptation. We have strengthened the controls and narrowed the claims.

Q1. What is the finding beyond task-specific supervision?

To our knowledge, this is the first systematic investigation of whether failure of a QPE readout implies missing rare-event information in its frozen upstream radar representation, using controlled comparisons of diverse readout families on the same encoder. Strong baselines are part of the evidence, not merely competitors: their success reveals information hidden from the deployed head.

With frozen 3DMAE fixed, QPE obtains zero H5/H10 event CSI; direct prototypes achieve 0.1647/0.1222, Balanced Linear 0.2235/0.1697, PaCo-Proto 0.2315/0.1701, and PatchShot 0.2329/0.1665 (Tables 1, 3). Balanced Linear and PaCo were already present; the revised evaluation also reports SimpleShot, Shrinkage QDA, and controlled Matched CE analyses. We do not claim universal classifier superiority. Their strong performance is evidence that QPE failure need not mean information loss.

The revised Figure 3 compares additional radar-only readouts across five support budgets (1%–50%). In the displayed curves, PatchShot increases with supervision and ranks highest among radar-only support-based readouts at 50%, but is lowest at 1% (19 cases), including below frozen prototypes at H5; simpler readouts are more robust there. This indicates budget-dependent accessibility, not uniform superiority. Figure 3 is descriptive; paired method comparisons are reported in Tables 1 and 3.

Table 2 isolates supervision. With matched architecture, training budget, prototype decoder, and support calibration, Matched CE achieves 0.2255/0.1684 versus PatchShot's 0.2329/0.1665. Shuffled-label SupCon returns H5 CSI near the unadapted-prototype level (16.64 vs. 16.02), with lower AP (17.50 vs. 21.62). Thus task-aligned labels, rather than adapter capacity alone or a uniquely necessary contrastive objective, drive much of the gain. We call this readout-dependent recoverability the representation accessibility gap.

Q2. How does re-access differ from conventional supervised adaptation?

We do not claim a novel optimization principle. The revised probes and interventions establish two narrower properties.

Accessibility. Table 6 measures diagnostic CSI under a shared probe protocol, not the deployed CSI in Tables 1–3. At H5/H10, prototype-to-envelope accessibility rises from 0.788/0.823 (frozen) to 0.981/0.996 (radar-only PatchShot). Matched CE also reaches 0.981/0.987 despite using a CE rather than contrastive objective. This argues against the improvement being explained solely by a prototype-compatible loss: task supervision improves accessibility under different objectives. PatchShot exhibits smaller normalized cross-probe ranges (0.061/0.058 vs. 0.103/0.099 for Matched CE; Table 7), with higher k-NN CSI. The envelope is empirical, not an information-theoretic bound, and these results do not isolate a uniquely contrastive mechanism.

Support traceability. We agree that removing support examples demonstrates dependence on prototype construction, not on SupCon itself. The added value is exact, case-specific intervention: after fixing the adapter, we remove a support case and recompute prototypes to measure its effect without retraining. This property belongs to the support-derived readout, unlike a conventional fitted CE/linear head without an equivalent prototype-recomputation operation. Figures 4 and 6 use development-trained ERA5 adapters (H5: Thermodynamic; H10: Spectral). Removing the 10 most influential of 19 support cases flips about 40% of true-positive query patches, versus under 8% for equal-size random removal, with stronger traced-removal effects across five paired seeds. We claim decision-level dependence, not physical causality.

Temporal dependence. New Figure 7 audits leakage, input overlap, short-range proximity, and atmospheric persistence. With identical queries, two-hour blocking changes radar-only PatchShot H5/H10 CSI from 0.2128/0.1757 to 0.2278/0.1630; blocked frozen prototypes obtain 0.1795/0.1406. This does not establish storm-level independence.

Summary. The contribution is a systematic diagnosis of prediction failure versus representation information loss. Strong readout comparisons reveal recoverable rare-event evidence; budget, probe, and intervention analyses characterize its accessibility and traceability.

Thanks for the authors' responses, I will keep my initial score.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Meteorological Motivation and Representation Accessibility

We would like to clarify the meteorological motivation underlying our study, as illustrated in the revised Figure 1.

Radar reflectivity and surface rainfall are imperfectly related: strong radar echoes do not necessarily correspond to heavy surface rainfall, while intense rainfall may occur under comparatively weak echoes. Figure 1(a) illustrates these contrasting cases alongside a matched one. Because event labels are defined by the surface rain target rather than by radar intensity, echo strength alone may be insufficient for reliable rare-event recognition.

This motivates our central research question: whether a frozen radar representation already retains event-relevant information that its original QPE readout fails to expose. Rather than treating prediction failure as evidence of missing information, we investigate the accessibility of rare-event evidence through controlled readout comparisons.

The examples in Figure 1 are illustrative and do not establish specific physical mechanisms. Our conclusions about representation accessibility rest on the experimental analyses, not on these schematic examples.

AuthorsOpenReview ↗

Physical Motivation and Interpretation of Context-Guided Routing

PatchShot-Env uses atmospheric context to guide how frozen radar evidence is read, rather than introducing additional precipitation labels or explicit storm identities. This reflects a key challenge in precipitation retrieval: similar radar observations may correspond to different surface rainfall under different environmental conditions.

1. Direct evidence and indirect environmental constraints.

Three-dimensional radar provides relatively direct evidence of precipitating hydrometeors, whereas ERA5 characterizes the surrounding thermodynamic and dynamical environment, and AGRI provides complementary cloud-state observations.

These auxiliary sources do not supply precipitation ground-truth labels. Instead, they provide indirect constraints on how radar representations should be interpreted. This motivates environment-conditioned readout, where atmospheric context influences the readout mapping rather than being treated solely as an additional concatenated feature.

2. Continuous weather systems and event independence.

Extreme precipitation is concentrated in rainy seasons, when weather systems may persist, regenerate, split, or merge within evolving atmospheric environments. Consequently, defining strictly independent meteorological events is difficult, even when radar cases and target times are disjoint.

Our case-disjoint protocol and additional 2-hour temporal-blocking analysis address sample overlap and immediate temporal dependence, but do not establish complete independence of the underlying weather systems.

Routing does not resolve this dependence. Rather, it conditions the readout on observed atmospheric states without requiring explicitly defined storm identities or discrete event categories.

3. Context-guided expert routing in PatchShot-Env.

As illustrated in Figure 2, temporally matched ERA5 descriptors, encoded as signed low-order vertical modes, determine soft weights over lightweight experts through MoE-inspired routing. Their weighted outputs are added residually to the frozen radar tokens before support-supervised adaptation, while the 3DMAE encoder remains unchanged.

The underlying hypothesis is that similar radar representations may require different readout mappings under different environmental conditions. Since such context-dependent mappings are difficult to specify manually, PatchShot-Env learns the conditioning of the readout from data rather than relying on predefined meteorological rules. ERA5 serves as a case-aligned conditioning signal rather than an independent precipitation predictor. AGRI is examined separately as a complementary cloud-state conditioning extension.

4. Experimental evidence and interpretation.

At H5/H10, PatchShot-Env achieves 30.95%/25.59% CSI, compared with 30.27%/24.30% for simple radar–ERA5 fusion, 30.39%/24.63% for Uniform Routing, and 30.01%/23.58% for case-shuffled routing (Table 4).

The comparisons provide complementary evidence:

  • Beyond simple fusion: PatchShot-Env improves CSI by 0.68/1.29 percentage points (pp), indicating a modest additional benefit from conditional routing.
  • Beyond expert capacity: Uniform Routing uses the same expert bank with equal weights and performs 0.56/0.96 pp below PatchShot-Env.
  • Environmental alignment: Shuffled Routing preserves the full-spectrum descriptor and routing architecture but disrupts case-level ERA5–radar correspondence. Compared with matched Full-Spectrum routing (30.90%/25.31%), CSI decreases by 0.89/1.73 pp. Shuffled Routing also falls below Uniform Routing by 0.38/1.05 pp, suggesting that misaligned context can be less useful than uniform expert weighting.

PatchShot-Env learns to use indirect environmental information to condition the readout of rare-precipitation evidence from frozen radar representations, without requiring predefined atmospheric regimes or explicit storm identities. The complementary controls further highlight the importance of case-aligned environmental information beyond simply increasing expert capacity.

5. Practical implications and future directions.

Improving extreme-rainfall detection remains important for disaster-risk assessment, even at larger support budgets, while more reliable readouts can facilitate failure attribution. Future work will investigate elevation-aware calibration of sparse radar–surface rainfall correspondences to distinguish terrain-related artifacts from genuine precipitation signals and improve retrieval accuracy.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.