Global Spectral Context Measured in Meters, for Satellite Models That Cross Sensors and Continents

Satellite land-use models fail when deployed in places they were not trained on, and the benchmarks that measure this failure have mostly tested one thing: training objectives stacked on ordinary convolutional networks. Those networks see only a small neighborhood of pixels at a time. They have a second, quieter problem. A neighborhood measured in pixels covers a different patch of ground on every satellite, because each sensor has its own ground sample distance (GSD), the width of ground covered by one pixel. A filter that has learned that "texture at 1/8 cycles per pixel means rooftops" is learning a fact about one camera, not about the Earth.

We propose a single change that addresses both problems. We replace local feature mixing with a learnable global frequency-domain filter, a gate applied to the Fourier transform of a feature map. Every Fourier coefficient summarizes the whole image, so this gate sees the entire tile at once. Its cutoff frequencies are defined in cycles per meter of ground rather than cycles per pixel, and they are converted to pixel units at run time using each image's recorded GSD. One set of learned parameters then describes the same ground-scale structures, such as street grids, field boundaries, and parking lots, on every sensor. Our central question is whether global context expressed in physical units narrows the geographic out-of-distribution gap more than either global context alone or sensor normalization alone. We also ask whether larger context sometimes makes that gap worse.

Core Claim

The claim has two separable parts, and the experimental design tests each one on its own. The first is that wider spatial context, delivered cheaply through a frequency-domain gate, helps under geographic shift. We do not assume this. Distinguishing a factory from an office needs surrounding context, which is consistent with how weak image-only features are on Functional Map of the World. Yet Fraccaro et al. and Rolf et al. warn that models exploit spatially correlated backgrounds to memorize sites, with drops of up to 30% in new regions. Larger context could feed that shortcut. The second part is that context which is consistent across sensors helps beyond context alone. Expressing the gate in ground units removes a nuisance variable, the pixel scale, in the same way that the differential radio map cancels receiver gain by subtracting signal strengths between access points. That analogy supplies the method's logic: express the computation in a quantity in which the device-specific term cancels before the network ever sees it.

The Mechanism

We build on the anisotropic Butterworth gate of Butterworth as Attention. That work turns the classical smooth low/high-pass filter into a learnable layer with rotatable elliptical passbands, run through fast Fourier transforms at $O(N \log N)$ cost. It was applied only to pansharpening and evaluated by reconstruction fidelity. For an $N \times N$ feature map with frequency indices $(u, v)$ and recorded GSD $g$ in meters per pixel, the ground frequency is $(u, v)/(N g)$. Our gate is

$$ G(u,v) = \left[1 + \left(\frac{\lVert R_\theta\, (u,v)/(N g) \rVert_{A}}{c}\right)^{2n}\right]^{-1}. $$

Here $R_\theta$ is a learned rotation, $\lVert\cdot\rVert_A$ is an elliptical norm with learned axis ratio, $c$ is a learned cutoff in cycles per meter, and $n$ is the learned filter order. Several such gates per layer, each with its own output projection, form a block. These blocks replace the late-stage convolution stages of a ResNet-50 or DenseNet-161 backbone, while the early convolutional stem keeps local edge detail. Keeping that stem answers the worry that global attention acts as a low-pass filter and erases boundaries. Because $c$ lives in ground units, the gate's effective receptive radius in meters becomes an interpretable quantity we can sweep and constrain.

As a secondary ablation, we fold off-nadir angle into the ellipse. Oblique viewing stretches ground along the look direction, so we scale one axis by the cosine of the recorded off-nadir angle. The file lists sensor off-nadir but not the satellite's look azimuth, so the direction of that stretch is learned rather than given.

Relation to Prior Work

Three lines of work each touch two of the three directions. None of them puts these together.

FlexiMo handles varying resolution by resizing patch embeddings with a parameter-free interpolation inverse, and handles varying band sets with wavelength-generated kernels. It is evaluated only on single-region or random splits such as SegMunich and EuroSAT. It also keeps standard transformer attention. Our proposal tests on held-out geography and moves resolution handling into the global mixing operator itself.

WILDS and its unsupervised extension supply both the evaluation and the motivating negative result. Invariance losses such as CORAL and IRM rarely beat plain training across real shifts, and unsupervised adaptation gave no reliable gain on PovertyMap. Every model tested there was a standard local convolutional network, so the architecture axis is untested.

Functional Map of the World records GSD, angles, and coordinates for each image, but only concatenates them before the classifier. We use the same metadata to set what the feature extractor computes.

GeoFed separates regional class-frequency shift from appearance shift. It does this through federated losses on local backbones and assumes uniform sensors, and its partitioned datasets serve us as region-holdout segmentation tests.

CCNet provides an alternative cheap global operator: attention along rows and columns, applied twice so that every pixel reaches every other. It was shown only on Cityscapes and ADE20K. We use it as a competitor that has global context but no notion of physical units.

Research Program

We will build the gate block, insert it into matched backbones, and train under official benchmark splits. Every variant gets an identical recipe, ImageNet initialization, and an equal hyperparameter-search budget of a fixed number of trials per variant, selected on the official out-of-distribution validation split. Each configuration is run over multiple seeds. The program has three analyses.

The first is a factorial comparison crossing two factors: context type (local, global in pixel units, global in meters) and sensor handling (none, resample to a common GSD, metadata conditioning). This separates the two parts of the claim.

The second is a context-radius curve. We cap the gates' effective cutoffs to allow only structures larger than a given ground scale, sweep that scale from tens to thousands of meters, and plot in-distribution and out-of-distribution performance against it. If out-of-distribution performance peaks and then falls while in-distribution performance keeps rising, we have measured the memorization effect that Rolf et al. and Fraccaro et al. describe, and we report the safe radius.

The third is a control-dataset prediction. PovertyMap-WILDS uses Landsat imagery at a single GSD, so the meter-unit conditioning is mathematically a no-op there. Any gain on PovertyMap must come from global context alone. If the metric-unit version beats the pixel-unit version on PovertyMap, something other than our mechanism is responsible, and we will say so.

Baselines

The naive reference is plain empirical risk minimization with the WILDS backbones, a DenseNet-121 or ResNet-50 for FMoW-WILDS and a ResNet-18 adapted to multispectral input for PovertyMap. On PovertyMap, WILDS reports Pearson correlation of 0.60 in distribution and 0.45 out of distribution for this baseline WILDS. On FMoW-WILDS the file gives in-distribution accuracy of 59.7%, but the out-of-distribution and worst-region figures are truncated; no reported result was found on Lacuna for FMoW-WILDS out-of-distribution accuracy. We therefore rerun this baseline ourselves under our protocol.

The strongest existing methods on these benchmarks come from WILDS 2.0. That paper names FMoW-WILDS as one of only two image tasks where unsupervised adaptation (for example Noisy Student or FixMatch) clearly beat empirical risk minimization, but no reported number was found on Lacuna for those FMoW results. We rerun the best of them using the official unlabeled target data. Beating it means exceeding its worst-region accuracy by more than seed-to-seed spread without using any unlabeled target data. We also test whether the two methods stack.

The metadata baseline follows the original fMoW recipe of late concatenation, which on the original fMoW split raised average F1 from 0.679 to 0.722 Functional Map of the World. This tests whether metadata is merely useful as a feature or useful as a control of the filters.

The resample baseline rescales every image to a common GSD before an otherwise identical model. We discuss it below because it is the main threat.

The global-context competitors are a CCNet head and a pixel-unit Butterworth gate in the style of ABFNet, both on the same backbone. They have matched parameter counts and FLOPs within about ten percent. The pixel-unit gate is the key ablation, since it differs from our model only in whether cutoffs are expressed in meters.

Finally, a FlexiMo-style resolution-aware transformer with matched initialization is included. The released FlexiMo is reported separately, labelled as having different pretraining.

Why not just…?

Why not just resample every image to a common ground resolution before training? This is the real threat. When going from fine to coarse resolution, resampling achieves the same unit consistency as our gate. Where they differ: resampling discards detail on fine sensors when downsampling and invents none when upsampling coarse ones. It is also applied once to the input, whereas our gates adjust at every layer and leave fine-resolution features intact. We will not argue past this alternative; the factorial design settles it. If "resample plus pixel-unit global gate" matches "metric global gate" on FMoW-WILDS worst-region accuracy, the honest conclusion is that global context is what matters and sensor normalization can be the simplest one available. That would still answer the open question from WILDS about architecture. We would report it that way and drop the metric-unit claim.

Why not just use a vision transformer or a remote-sensing foundation model, which already has global attention? Transformers do provide global context. FlexiMo even adds resolution awareness, but at the token-embedding stage, and it has never been tested under geographic holdout. Its attention is quadratic and, as the Butterworth paper argues, smooths away boundaries. More importantly, large pretrained models confound architecture with data. Our matched-initialization transformer baseline isolates the architecture question. The off-the-shelf FlexiMo result tells practitioners whether pretraining scale overwhelms the architectural effect. If it does, the gate is worth less than it seems, and we will state that.

Why not just add an invariance or robustness loss? WILDS tested CORAL, IRM, and Group DRO across ten real shifts and found they rarely beat plain training. Group DRO helped only with explicit categorical groups, as on CivilComments. We include the strongest unsupervised adaptation method instead, because it is the one family the file reports working on FMoW.

Why not just feed the metadata into the classifier, as fMoW did? Late concatenation helps on an independent split, but it cannot change which spatial frequencies the backbone extracts from an unfamiliar sensor. The comparison between late metadata and metric-conditioned gates, at equal parameters, tests exactly that difference.

Proposed Experiments

The primary benchmark is FMoW-WILDS, with worst-region accuracy and average out-of-distribution accuracy measured on the official region and year splits. Here both sensor variation and geographic shift are present. PovertyMap-WILDS, scored by out-of-distribution Pearson correlation across countries, serves as the single-sensor control. For dense prediction, which tests whether the gate keeps boundaries, we train centrally on the climate-zone partitions of FedCASID and the city partitions of FedInria with one zone or city held out. We score mIoU and boundary F1 there. GeoFed's improvements over federated baselines, 2.17% global on FedCASID and 1.03% on FedInria, belong to a different, federated protocol and are context only, not a target.

Three decisive checks determine the outcome. The first is the FMoW-WILDS worst-region gap between the metric gate and the pixel-unit gate, with the resample baseline as the bar both must clear. The second is the PovertyMap control: we predict no difference between metric and pixel versions, and gains over empirical risk minimization only if global context itself helps. The third is the shape of the context-radius curve, which either confirms or refutes the claim that unbounded context invites site memorization. We will also stratify FMoW test images by GSD bin. If the mechanism works, its advantage should be largest for the sensors most underrepresented in training, and absent where sensors match. A result without that pattern would mean the gains come from somewhere else.