acceptodds
Under review as a conference paper at ICLR 2027

Learning Continuous Regional Temperature Fields with Lead-Time and Resolution Queries

Abstract

Accurate regional near-surface temperature forecasting is fundamental to short-range weather services and downstream risk assessment. Existing deep learning-based regional forecasters commonly produce a fixed set of future frames on a prescribed grid, limiting their use when forecast products must be evaluated at query-dependent lead times or display resolutions. To address this limitation, we formulate regional T2M forecasting as query-conditioned continuous spatiotemporal temperature field evaluation and propose the Continuous Spatiotemporal Temperature Forecaster (CSTF), a neural field that evaluates 2-m temperature (T2M) at queried lead times and spatial resolutions. Specifically, CSTF first encodes multi-variable ERA5 histories into latent meteorological states and then decodes T2M as a coordinate-based field. Accordingly, spatial location, forecast lead time, and output resolution are introduced as queries, enabling standard hourly forecasts, intermediate lead-time diagnostics, and resolution-controllable outputs within a unified field-evaluation framework. Furthermore, to maintain coherence across flexible field queries, we design spatial-gradient, temporal-difference, and scale-consistency objectives that regularize regional thermal structures, lead-wise evolution, and cross-resolution agreement. Experiments on the Southeast China 0–6 h ERA5-Land benchmark demonstrate that CSTF achieves the best aggregate deterministic skill, including a 17.0% reduction in Bias, with global-scope diagnostics further illustrating flexible lead-time and resolution-controllable inference.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Weather Forecasting, Continuous Spatiotemporal Modeling, Query-Conditioned Prediction, Regional Temperature Forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ogrmOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the limitation of existing deep learning-based regional weather forecasting models, which are constrained by fixed forecast lead times and predefined discrete output grids. The authors reformulate short-range 2-m temperature (T2M) forecasting as a query-conditioned continuous spatiotemporal field evaluation task and propose the Continuous Spatiotemporal Temperature Forecaster (CSTF).

The core pipeline consists of:

  1. Shared Latent State Construction: A multi-scale encoder compresses an input sequence of 3-hour, 9-variable historical ERA5 data into a latent meteorological state ZZ that is independent of fixed lead times and target grids.

  2. Multi-Scale and Continuous Query Decoding: Spatially, the latent representation ZZ is bilinearly interpolated to the requested query resolution (Ho,Wo)(H_o, W_o). It is then concatenated along the channel dimension with normalized spatial coordinates PP, grid scale encoding RR, and a continuous sinusoidal temporal embedding vector ϕ(τ)\phi(\tau). A lightweight decoder directly maps this combined tensor to the requested field.

  3. Cross-Scale and Spatiotemporal Regularization: The framework introduces a spatial-gradient loss, a temporal difference loss, and a scale-consistency loss that uses a detached, area-downsampled high-resolution prediction as a self-supervised pseudo-ground-truth.

For empirical validation, the authors evaluate CSTF on a self-constructed Southeast China benchmark (resized to 128×128128 \times 128) for 0–6 h forecasts. CSTF demonstrates lower aggregate RMSE (1.123°C) and reduced bias compared to baseline models at integer lead times. Furthermore, qualitative diagnostics over regional and global scopes are provided to illustrate smooth transitions under non-integer lead-time queries (e.g., 0.5 h, 4.5 h) and multiple output resolutions (from 96296^2 to 1922192^2).

Critical strengths
  1. Unified Query-Conditioned Neural Field Paradigm:

The paper departs from the conventional paradigm of hard-coding prediction lead times and output grids into discrete output channels. By treating short-range temperature forecasting as an implicit neural field evaluated conditionally on shared latent states, spatial coordinates, and temporal queries, it effectively decouples atmospheric background feature extraction from flexible downstream product generation.

  1. Lightweight Cross-Scale Consistency Regularization:

The proposed scale-consistency objective (Lscale\mathcal{L}_{scale}) leverages an area-downsampled high-resolution prediction with detached gradients as a self-supervised target. This effectively aligns evaluations across different sampling densities without introducing auxiliary networks or learnable parameters (ablation studies demonstrate that it reduces normalized cross-resolution error from 1.295 to 1.052), providing an efficient mechanism for multi-scale coherence.

Critical weaknesses

1. Limited Data Scale and Lack of Standardized Benchmark Comparisons (e.g., WeatherBench)

Modern data-driven atmospheric models typically benefit from training on multi-decadal reanalysis datasets (e.g., 40 years of ERA5 data) to establish robust dynamical representations that generalize across diverse seasons, synoptic regimes, and extreme weather events. In contrast, the empirical evaluation in this work relies on a relatively small, custom data subset (comprising only 4,615 training samples from 2024 to 2025, uniformly rescaled to a 128×128128 \times 128 resolution).

Recommendation: To establish fair and standardized comparisons with the broader literature, the paper would be significantly strengthened if the authors could evaluate the proposed method on recognized community benchmarks, such as the official WeatherBench evaluation suite (https://sites.research.google/gr/weatherbench/deterministic-scores/) and its regional or East Asian subsets. Demonstrating consistent performance gains on standard public benchmarks would instill much greater confidence in the method's effectiveness across the research community.

2. Generalization Concerns Regarding the Minimalist Architecture (High Train/Test Distribution Similarity and Small Test Set Size)

The paper achieves continuous-time and multi-resolution forecasting through a highly streamlined design combining sinusoidal time embeddings ϕ(τ)\phi(\tau), bilinear interpolation in the latent feature space, and a lightweight decoder. While this minimalist pipeline yields encouraging numerical metrics on the 0–6 hour South China benchmark, it inevitably raises a gentle concern: could the strong performance of such a simple "interpolation + temporal modulation" design be largely attributed to the minimal distributional shift between the current training and test sets, coupled with a very small test sample size (only 714 test samples, drawn from the exact same brief 2024–2025 temporal window as the training set)? When the training and test sets are temporally adjacent and limited in scale, the background climatology and synoptic patterns remain highly similar. Under these conditions, a lightweight decoder can easily achieve low test errors by fitting local smooth manifolds, which does not necessarily indicate that the model has genuinely captured complex, non-linear atmospheric physical evolution. Validating this minimalist architecture on a larger-scale test set and diverse climatic regimes would help dispel potential skepticism that the superior metrics stem primarily from overly close train/test distributions and a limited evaluation sample size.

3. Positioning the Research Motivation Against the Direct Regional Cropping Paradigm of Global High-Resolution Models

While overcoming the fixed-grid limitations of regional models is a valuable goal, state-of-the-art global weather foundation models (e.g., FengWu-GHR, Pangu-Weather, GraphCast) are increasingly capable of directly generating global high-resolution forecasts. Downstream users can simply crop any target region directly from these global outputs, thereby obtaining fine-grained regional predictions while naturally preserving the physical consistency of large-scale global circulation. The motivation of this paper would be much more solid and inspiring if the authors could more explicitly articulate and discuss the irreplaceable advantages of training a dedicated regional continuous neural field, specifically in terms of computational overhead, unique application scenarios, and fine-scale physical benefits, compared to directly deploying or fine-tuning a global high-resolution foundation model and cropping on demand.

Questions for authors

See Weakness.

Reviewer_kSHiOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a continuous forecasting method using a query-conditioned approach. In this field, continuous prediction has long been discussed as an important core topic, and the query-conditioned approach makes sense.

Critical strengths

The most critical strength is that this method follows current research trends, and we can reasonably expect this combination of methods to work well in some cases.

Critical weaknesses

This paper has the potential to become a good paper if several critical weaknesses are addressed.

  1. The methodological details are insufficient to understand the proposed method. For example, in the multi-scale meteorological encoder, how many encoders and decoders are used? What is the bottleneck structure? What are the channel dimensions? Without these details, even the model size is unclear. Regardless of how good the performance is, it needs to be assessed in relation to the model size.

  2. The paper uses the term “continuous latent representation” without explaining it. For the latent representation to be continuous, I would expect a query to produce the corresponding latent representation. However, the description suggests that the decoder performs this role. If I have misunderstood the method, the current description is highly confusing.

  3. The evaluation protocol is unclear. Why are probabilistic and deterministic forecasting models compared under the same settings? Probabilistic models should be evaluated using ensemble predictions. Otherwise, they are naturally at a disadvantage when evaluated on deterministic forecasting metrics.

  4. The proposed method does not include any temperature-specific components or regularization, so why is it evaluated only on temperature forecasting? This appears to be dataset cherry-picking. Please provide quantitative, rather than only qualitative, experimental results for era2era5 and describe the evaluation protocol rigorously.

Questions for authors

N/A

Reviewer_WgqpOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies regional 2-meter air temperature forecasts for the next 1–6 hours. CSTF uses three input frames of nine ERA5 variables to build a shared weather representation. It then predicts ERA5-Land 2-meter temperature at a requested lead time and output grid. During training, the authors add losses for spatial temperature changes, changes between adjacent hours, and agreement between predictions on different grids.

Critical strengths
  1. Hourly forecasts improve at leads of 3–6 hours. CSTF has the lowest RMSE and MAE among the listed methods at these four leads in Table 1. It also has the lowest mean errors in Table 2. This supports its basic forecasting ability, especially at later leads.
  2. The paper tests the added loss terms. In Table 3, the mean RMSE falls from 1.213°C with only the forecast loss to 1.123°C with the full method. Table 4 also shows that the scale-consistency loss reduces differences between predictions on different grids. These ablations help explain the role of each loss.
Critical weaknesses
  1. The hourly results do not establish the value of the query-based design. CSTF has a good mean score in Table 2, but its RMSE at hours 1 and 2 is 0.823°C and 0.909°C, compared with 0.739°C and 0.870°C for ARROW. A fixed-output model can also predict at hourly leads, and Table 1 evaluates only a 128×128 grid. The paper does not compare against strong architectures such as FourCastNet or Pangu-Weather retrained for the same hourly task. It also lacks a fixed-output control with a similar model size and training setup. It is therefore hard to tell how much of the mean improvement comes from the continuous query design.
  2. The evidence for variable lead times and resolutions is incomplete, and the global figures appear incorrect. The half-hour predictions have no reference temperature fields. Figures 5–8 change either the lead time or the output grid, but do not test what happens when both change together. Table 4 measures how close predictions on different grids are to the model's own 128×128 prediction. It does not measure their error against the true temperature. The experiments do not quantify whether small-scale temperature patterns are preserved, and they lack comparisons with interpolation after forecasting or methods designed for arbitrary output scales on the same task. Figure 4 raises another serious concern: the valid temperature regions in both the reference and CSTF panels appear to be displaced from the coastlines by about 180° of longitude. The authors should check the longitude convention and alignment of the data and land mask.
Questions for authors
  1. Could you hold out some hourly lead times during training and test the model at these unseen times? (e.g. training at T+1, T+2, T+4 and T+5, and setting the predict lead time at T+3 and T+6). Please compare it with linear interpolation between predictions at the neighboring trained hours. ERA5-Land provides hourly fields; if independent half-hour station observations are available, errors at station locations would also be useful. Such tests would change how I judge the continuous-time claim.
  2. For outputs at 96×96, 160×160, and 192×192, could you report errors against reference temperature fields? Please compare with a simple baseline that predicts at 128×128 and then resamples, as well as with arbitrary-scale downscaling or super-resolution methods trained on the same data and forecast task. Beyond MAE and RMSE, please compare spatial power spectra of temperature anomalies, errors at different spatial scales, and temperature-gradient errors on a common physical grid. This would help show whether small-scale patterns are oversmoothed. The ERA5-Land grid supplied by CDS is 0.1°×0.1°. The study region spans 20°×20°, so its original grid is about 200×200. Since the data are reduced to 128×128 before training, could you keep the original grid as a reference for the 192×192 output? Comparing only with the model's own 128×128 output does not show whether the new grid is more accurate.
  3. Could you add strong architectures retrained on the same data and hourly forecast task, such as FourCastNet or Pangu-Weather? A fixed-output model with a similar number of parameters and training conditions would also help.
  4. Please discuss how this work relates to the coordinate-based continuous ERA5 fields in Huang and Hoefler (ICLR 2023). That work compresses existing data rather than forecasting the future, but it uses a related idea of querying a continuous field by space and time coordinates.
  5. Please check Figure 4 and the other global figures.
  6. Why were these nine input variables chosen? Could you compare forecasts using only past ERA5 T2M, using all variables except ERA5 T2M, and using all nine variables? Please report errors at each lead time. The current ablations keep the input variables fixed, so they do not show how much the other weather variables help.
  7. The paper says the lead time and output grid can be requested together, but Figures 5–8 vary only one of them at a time. Could you report a joint “lead time × output grid” experiment, for example using 96, 128, 160, and 192 grids at several hourly and half-hour leads? Hourly leads could be checked against original ERA5-Land fields. For half-hour leads, you could at least test agreement across grids and compare with a baseline that interpolates in both time and space.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.