When a network fills in a rain or wind map from a few hundred ground stations, the usual training signal is error at the stations. Error at the stations rewards smooth maps. Adding an adversarial critic, a second network trained to tell real maps from generated ones, is the standard fix for blur. It creates a new failure mode when inputs are this sparse. A typical critic looks at small patches, so it only checks that each neighborhood looks like rain. A generator can then satisfy it by painting plausible local texture around each station while leaving the large-scale pattern incoherent: bands that should connect across a 100 km gap do not, and cells cluster around gauges. We propose that when stations are sparse, the critic's field of view matters more than the generator's. A critic that sees the whole map can reject fields that are locally realistic and globally wrong. Such a critic can also be trained on archives of dense maps that never had paired inputs. This makes it a way to learn spatial structure without many paired labels.
In two sentences, the idea is this. Train a single-pass conditional generative adversarial network (GAN) to turn sparse station readings plus satellite imagery into a dense field. Give its discriminator an efficient whole-grid attention mechanism, and feed that discriminator unpaired dense radar or simulation fields, so it supplies a large-scale structural prior. That prior replaces most of the scarce paired dense ground truth. The testable prediction is that the gain from a global critic over a local one grows as station spacing grows past the local critic's patch size.
What exists and what is missing
The closest prior work attacks sparse field reconstruction in three ways, and none puts a global receptive field in an adversarial critic. STA-GANN uses an adversary, but only as a domain classifier that makes node embeddings indistinguishable between seen and unseen stations. It predicts values at discrete graph nodes, never a continuous map, and never judges whether an output field is realistic. GenDA is the nearest in spirit. It trains a diffusion model on simulated urban wind, uses it as an unconditional prior, and guides it toward sensor values, cutting relative RMSE by 25–57% over MeshGraphNet and POD-3DVar at 0.1% sensor coverage. It needs many denoising steps per sample and gets its long-range reach from a two-level mesh, not from the critic. Station2Radar reconstructs radar-like precipitation from stations and GK2A infrared imagery with station-anchored anisotropic Gaussian splats. It reports a 50% RMSE improvement over IMERG and MSWEP. Its splats are local, and it trains with regression losses.
Two regression approaches motivate the blur concern. Dense Air Pollution Estimation predicts full NO₂ grids in one pass with a bottlenecked convolutional autoencoder. It found that skip connections carried noise from sparse inputs, and it reached 4.98 µg/m³ MAE. SPIN enforces texture by matching satellite aerosol gradients in its loss and uses graph kernels bounded by geographic radius. Both show that deterministic losses can be made structure-aware without an adversary, so a plain gradient or spectral loss is a baseline we must beat.
The attention mechanism comes from CCNet. Each location attends along its own row and column. Two passes with shared weights give every location a path to every other, at 11× less memory than dense non-local blocks. CCNet was built for fully labeled natural-image segmentation. We reuse it inside a discriminator on sparse-input fields.
The file's own note on GANs names the gap directly: CNN-based discriminators struggle to judge realism across large unobserved voids. GraPhy constrains the design. It reports a "sparsity floor," below which local message passing stops helping. We read this as evidence that locality, not model capacity, is what fails, and it predicts where our effect should appear.
Core Claim
In sparse-station field reconstruction, a discriminator with a whole-grid receptive field is what suppresses globally incoherent generator outputs, such as cells clustered around stations and broken long-range structure. A generator with global attention trained against a local critic will not suppress them. Because the discriminator only needs realistic dense fields, not paired examples, it can learn from unpaired archives. With 1–10% of the paired dense ground truth, it should then approach the structural quality of full supervision while keeping station error at the level of strong regression models.
Research Program
The generator architecture is fixed across all arms. It is a convolutional encoder-decoder, with recurrent criss-cross attention at the bottleneck. Its inputs are the satellite image, a rasterized station-value channel, a station-mask channel, and a distance-to-nearest-station channel. Its output is a dense field. The discriminator is conditioned on the same inputs and sees either a real or a generated field. We build it in two versions that differ only in receptive field. The local version is a patch critic with a roughly 70-pixel field of view. The global version adds two criss-cross attention passes before its output head.
Training combines three signals:
$$\mathcal{L}_G = \mathcal{L}_{\text{stn}} + \lambda_{p}\,\mathbb{1}[\text{paired}]\,\lVert G(c) - y \rVert_1 + \lambda_{a}\,\mathcal{L}_{\text{adv}}(D_{\text{glob}}, G(c))$$
Here $c$ is the conditioning input (satellite, station channels), $G(c)$ is the generated field, and $\mathcal{L}_{\text{stn}}$ is absolute error at a random 20% of stations. Those stations are hidden from the input each step, following the masking idea of IGNNK. The second term applies only on the scarce timestamps that have a paired dense field $y$. The third term uses a hinge adversarial loss with gradient penalty; the penalty is used for stability, as in the WGAN-GP usage of Algaba et al.. The discriminator sees two kinds of real samples. On paired timestamps it sees correctly matched (input, field) pairs. It also sees fields drawn from the unpaired archive, scored by an unconditional head. The two heads share the global trunk, so structure learned from the archive informs the conditional judgment.
The analysis centers on one diagnostic, which we call the station-halo profile. It is the radially averaged spectral power of the error field, together with forecast skill binned by distance to the nearest active station. A local critic should leave excess error and missing power at wavelengths longer than the typical station spacing. A global critic should close that gap.
Baselines
The naive reference is classical interpolation on station values alone, specifically ordinary kriging and inverse distance weighting. De Felice argues these break down as spacing grows, and we include them to show where learning starts to pay off at each density. They get no tuning beyond variogram fitting on training stations.
The strongest existing method on our primary benchmark is Station2Radar, because it was built for exactly this input-output pair. The file reports only a relative figure for it, a 50% RMSE improvement over IMERG and MSWEP, and gives no absolute RMSE. We therefore retrain it on our splits with the authors' configuration and the same data and epochs. Beating it means lower RMSE against radar and higher skill for heavy rain at all station densities, in the full-label and the 10%-label regimes. Matching it on RMSE while improving structural scores counts only as a partial win.
For the diffusion alternative we reimplement a GenDA-style model. It is an unconditional diffusion prior trained on the same dense archive, with classifier-free guidance toward station values, on the same generator backbone. On the Bristol wind benchmark we also compare against GenDA's own reported 25–57% RRMSE and 23–33% SSIM improvements over MeshGraphNet and POD-3DVar. The file gives no absolute GenDA RRMSE. Beating it means comparable RRMSE and SSIM at one forward pass, where diffusion needs tens of denoising passes. We report quality against wall-clock time, not just a single point.
The strongest deterministic competitor is our own generator trained without any adversary. We give it the best available structure-aware loss: L1 plus a spectral-power matching term plus gradient matching in the manner of SPIN. Its hyperparameter search budget equals the GAN's. This arm answers whether adversarial training is needed at all.
Two ablations isolate the key ingredient. The first keeps everything and swaps the global critic for the local patch critic. The second crosses the two factors: a local generator against a global critic, and a global generator against a local critic. If the second combination does as well as the full model, the claim is false. A further ablation removes the unpaired archive from the discriminator, which tests whether the label-efficiency gain comes from the archive or from the architecture.
Why not just…?
Why not just add global attention to a regression model and skip the GAN? This is the most serious threat. It is the deterministic baseline above, with the same criss-cross generator and a spectral-plus-gradient loss. Regression losses reach for the conditional mean. Where stations are far apart the mean is smooth, so we expect this arm to match the GAN on station MAE and lose on heavy-rain skill and long-wavelength power. If it matches on the halo profile as well, the adversary adds nothing and we would drop it. The experiment is designed to tell us so.
Why not just use a diffusion prior as GenDA does? Diffusion gives ensembles and tends to cover more of the distribution than GANs. Mode collapse is a real risk for us, and the file flags it. We do not claim better quality. We claim similar structural quality at single-pass cost, which matters for hourly national-scale maps or mobile fleets like OpenSense Zurich. We measure diversity directly, as the spread across noise seeds compared with radar variability, and report it. If diffusion wins on quality at every compute budget we test, we will say so.
Why not just pretrain a masked autoencoder on the unpaired dense archive and fine-tune a regressor? Masked Record Modeling shows masked pretraining can give strong results from 1% labels, for example 88.5% mean AUC on CheXpert. This is the natural alternative way to use the archive. We run it with the same archive, backbone, and compute. Our argument is that a pretrained encoder shapes the features but not the output loss. At fine-tuning time the regressor is still pulled toward smooth means, whereas the critic keeps penalizing unrealistic structure throughout training. If masked pretraining plus the deterministic loss matches us in the 1–10% regime, the adversarial part of our proposal is unnecessary.
Why not just use a standard local patch discriminator, as most conditional GANs do? This is our main ablation. The prediction is specific. When stations are dense, with spacing under the patch size, local and global critics should tie. The gap should open as spacing exceeds the patch. A uniform gap, or no gap, would falsify the receptive-field explanation.
Proposed Experiments
The primary testbed is the KMA benchmark from Station2Radar. It covers the Korean Peninsula for 2019–2023, with automatic weather stations and GK2A infrared imagery as inputs and 0.5 km radar as dense truth. We use it because it has real sparse inputs and real dense truth. We split by year and simulate scarcity along two axes. The first is station density: we randomly subsample stations to 100%, 50%, 25%, and 10%. The second is paired-label fraction: we keep radar pairing for 1%, 10%, or 100% of training timestamps. The remaining radar fields go into the unpaired archive with their timestamps shuffled, so they carry no pairing information.
A geographic-holdout variant withholds all radar from one subregion during training. This tests whether a structural prior learned elsewhere transfers. It targets the unseen-geography failure the file lists under label-efficient learning.
The secondary testbed is the Bristol wind benchmark from GenDA, with sensor counts down to 300. It gives a head-to-head comparison with diffusion on dense simulated truth. We do not use the Beijing-Tianjin-Hebei PM2.5 data or the NO₂ data as primary benchmarks. They have no dense truth for training or evaluating a field critic, and using satellite aerosol fields as "real" samples would confuse column and surface quantities. The method as proposed is therefore scoped to variables with a dense observational or simulated proxy.
Metrics are RMSE and MAE against dense truth and at held-out stations. We also report critical success index at 1, 10, and 30 mm/h, SSIM, the radially averaged power spectrum, and the station-halo profile, plus wall-clock inference time and seed-to-seed spread.
Three checks decide the question. First, in the density sweep, the error ratio of global critic to local critic should fall steadily as spacing grows, with a tie at full density. Second, at 10% paired labels, the global-critic GAN with the archive should come within a small margin of the full-label deterministic model on heavy-rain skill and long-wavelength power. It should also beat masked pretraining followed by regression. Third, the local-generator/global-critic arm should retain most of the full model's gain and the global-generator/local-critic arm should not. That would show the critic's view is the ingredient that matters. If the first and third checks fail, the core claim is wrong regardless of headline RMSE.
References
- STA-GANN: A Valid and Generalizable Spatio-Temporal Kriging Approach
- GenDA: Generative Data Assimilation on Complex Urban Areas via Classifier-Free Diffusion Guidance
- Station2Radar: Query‑Conditioned Gaussian Splatting for Precipitation Field
- Dense Air Pollution Estimation from Sparse in-situ Measurements and Satellite Data
- Physics-Guided Inductive Spatiotemporal Kriging for PM2.5 with Satellite Gradient Constraints
- CCNet: Criss-Cross Attention for Semantic Segmentation
- Graph-Based Physics-Guided Urban PM2.5 Air Quality Imputation with Constrained Monitoring Data
- Inductive Graph Neural Networks for Spatiotemporal Kriging
- LUCID-GAN: Conditional Generative Models to Locate Unfairness
- Graph-based Virtual Sensing from Sparse and Partial Multivariate Observations
- Sensing the Air We Breathe - The OpenSense Zurich Dataset
- Advancing Radiograph Representation Learning with Masked Record Modeling