Is Geographic Shift in SAR Flood Mapping Really a Resolution Problem?

Radar satellites map floods through clouds, and the models that read their images are known to fail unevenly across the world. A U-Net trained on the Sentinel-1 flood benchmark sen1floods11 scores intersection-over-union (IoU, the overlap between predicted and true water) anywhere from 0.53 to 0.81 depending on which flood event it is tested on (Deploying an Artificial Intelligence Application to Detect Flood from Sentinel 1 Data). The usual explanation is "geographic domain shift": terrain, climate and land use differ between regions, so a model trained in one place sees unfamiliar inputs in another. This proposal tests a narrower physical explanation. A radar can only separate two points if they are farther apart than its resolution cell, and that cell is set by antenna aperture and wavelength. Regions differ in how much of their flooded area is made of structures narrower than that cell: thin channels, flooded streets, water fringes under vegetation. If so, much of what looks like geographic shift is a resolution limit that shows up in different amounts in different regions. No amount of diverse training data or clever fine-tuning can repair that part, and it should be measured and reported apart from genuine appearance shift.

The concrete claim is that the per-event error of SAR flood segmentation models can be split into two parts. One is a resolution-limited part, predicted by a simple geometric covariate: how wide each piece of water is compared with the sensor's effective resolution cell. The other is a residual appearance-shift part, which is what geographic methods should actually target. We will build the decomposition, check it causally by coarsening the radar images on purpose, and train a model whose uncertainty keeps "this structure is too small for the sensor" separate from "this region is unfamiliar." If the claim holds, geographic hold-out scores for radar will need to be read differently, and it becomes possible to estimate, region by region, how much a sharper radar would actually help.

Why this intersection, and what is already known

The radar-perception literature has already shown that angular resolution, not model capacity, is often the binding constraint. On the automotive dataset CARRADA, the same segmentation network reaches 52.9% mean IoU on range-Doppler views but only 32.6% on range-angle views, because nearby targets merge within wide angular bins (CARRADA Dataset). Two methods attack this limit upstream of image formation. ADC-SRNet synthesizes missing virtual antennas from raw complex samples and lifts RADDet detection mAP from 0.764 to 0.802 (Azimuth Super-Resolution for FMCW Radar in Autonomous Driving). NEAR fits an untrained network with a low-rank wave-physics regularizer to a single snapshot (NEAR: Neural Electromagnetic Array Response). These papers supply the motivating physics. They also constrain what we can do: both need raw phase, which the processed intensity products in Earth-observation benchmarks no longer contain. We therefore do not propose super-resolving Sentinel-1. Instead, we estimate how much such super-resolution would be worth in each region.

On the geographic side, the position paper Mission Critical – Satellite Data is a Distinct Modality in Machine Learning argues that spatial autocorrelation and heterogeneity break the usual i.i.d. evaluation. STAR (Statistically-Guided Deep Network Transformation to Harness Heterogeneity in Space) responds by detecting spatial clusters of errors and branching the network by region. Trusting Small Training Dataset for Supervised Change Detection shows that geographic diversity beats proximity for optical change detection, and that a maximum-logit confidence score collapses on unfamiliar terrain. All three treat location as the cause. We ask whether, for active sensors, part of the location effect is mediated by a physical covariate that transfers across regions.

On the foundation-model side, CROMA (CROMA) and OmniSat (OmniSat) pre-train on paired radar and optical rasters and treat radar blur as a fixed property of the input. OmniSat explicitly names geometric distortion and coarse radar resolution as limits that attention is left to smooth over. Fine-tune Smarter, Not Harder shows that LoRA on the geospatial foundation model Prithvi 2.0 generalizes better to sen1floods11's geographic hold-out set than full fine-tuning does. No paper in the file asks which parts of the remaining hold-out error any adaptation method could plausibly remove. That question is the gap we target, and the file confirms no existing work spans resolution physics and geographic evaluation together.

Core Claim

For each labeled water pixel $x$, define its local width $w(x)$ as twice the radius of the largest disk that contains $x$ and fits entirely inside the water mask. This is computed by morphological opening at increasing radii. Define its resolvability as

$$\rho(x) = \frac{w(x)}{\delta},$$

where $\delta$ is the effective ground resolution of the radar image. We estimate $\delta$ from the width of the image's spatial autocorrelation peak rather than assuming it equals the 10 m pixel grid spacing, because the delivered grid can be finer than the true resolution. The claim is that the probability of misclassifying a pixel is, to first order, a function of $\rho$ and backscatter contrast that stays stable across flood events. Events differ mainly in their distribution of $\rho$. We will count the claim as supported if three things hold together. First, reweighting each held-out event to a common $\rho$ distribution removes a large share of the between-event IoU variance. Second, deliberately coarsening $\delta$ degrades events in proportion to their mass of low-$\rho$ water. Third, an uncertainty head trained to predict error from local resolvability stays calibrated on held-out regions, while generic confidence does not.

Research Program

The first component is the decomposition itself. Labels are needed to compute $\rho$, so this is an evaluation tool. On the 446 hand-labeled sen1floods11 tiles, we compute $\rho$ maps and bin pixels by $\rho$. For each trained model we fit a pooled error curve, the error rate as a function of $\rho$ bin and a local backscatter-contrast term. We then report two quantities for each event: the error that curve predicts from the event's own $\rho$ distribution, and the residual. The residual is the part that geography-aware methods such as diverse sampling, STAR-style branching or PEFT should be credited with or blamed for. We compare how much between-event variance is explained by $\rho$ versus by event identity.

The second component is a causal check. Correlation between thin water and error could be confounded, for example if urban events are both thin and differently textured. We therefore coarsen the radar input itself by factors of 1.5, 2, 3 and 4, using block averaging followed by resampling to the original grid, and retrain or re-evaluate each model at each level. This yields a per-event resolution dose-response curve. The prediction is that the slope of each event's curve is set by how much of its water sits near $\rho \approx 1$, not by its region. Extrapolating these curves toward finer $\delta$ gives an honest per-region estimate of what an aperture or super-resolution upgrade of the kind ADC-SR and NEAR provide would buy.

The third component is a model that uses this at test time, when labels are unavailable. On a Prithvi 2.0 backbone adapted with LoRA, we add two lightweight heads. The first predicts the sub-cell water fraction, trained with a soft target obtained by averaging the hand label over the estimated resolution cell. This lets the model say "partly water" instead of being forced into a hard mistake. The second predicts the probability of error. It is trained on in-distribution folds only and receives local image features that proxy resolvability, such as edge density and texture scale. We compare this error head against the maximum-logit indicator from Trusting Small Training Dataset. The question is whether separating a resolution-driven uncertainty from a residual out-of-distribution score lets a user tell "the sensor cannot see this" apart from "the model has not seen this place."

Baselines

The naive reference point is a ResNet U-Net on Sentinel-1 VV, VH and VV/VH composites, as in Deploying an Artificial Intelligence Application to Detect Flood from Sentinel 1 Data. That paper reports per-event IoU between 0.53 and 0.81 on sen1floods11. This baseline establishes the spread the decomposition has to explain. It will be trained on the same splits as every other model.

The strongest existing method is Prithvi 2.0 300M adapted with LoRA, following Fine-tune Smarter, Not Harder. That paper reports 87.57% mIoU on the sen1floods11 geographic hold-out set, against 82.07% for full fine-tuning of the same backbone. It is the strongest competitor because it is the best reported geographic-generalization result on the exact benchmark. We will reproduce it with its published recipe and the same GPU budget, and give our heads the same number of tuning trials. Full fine-tuning and a CROMA backbone (CROMA) serve as additional backbones, so that the decomposition is not tied to one model.

Beating this baseline means different things for different parts of the program. On the decomposition, the target is not a higher mIoU. It is showing that $\rho$ explains a larger share of LoRA's own remaining per-event variance than event identity does. On the model, beating it means matching its hold-out mIoU within noise while cutting the calibration error of error-probability estimates on held-out events. It also means achieving a higher area under the ROC curve for flagging wrong pixels than the maximum-logit indicator provides. For fractional outputs, which have no reported number on Lacuna, we report mean absolute error of predicted water fraction on low-$\rho$ pixels.

The key ablation removes the sub-cell soft target and the resolvability features while keeping the extra heads and parameter count. This isolates whether physics-aware supervision matters or whether any added head helps. A second ablation replaces $\rho$ in the decomposition with plain distance to the water boundary. This tests whether width relative to the resolution cell carries information beyond "near an edge."

Why not just…?

Why not just train on more geographically diverse data? Diversity helps with appearance shift, as Trusting Small Training Dataset showed for optical imagery (Kappa 0.32 diverse versus 0.12 localized). It cannot, however, make a 20 m-wide flooded street visible to a sensor whose resolution cell is wider than the street. The coarsening experiment makes this testable. If adding events from other regions flattens the dose-response slopes, then our resolution account is wrong, and we will say so.

Why not just use a boundary-weighted loss or blur augmentation? This is a real threat to the model component. Upweighting edges may recover most of the IoU gain of the soft sub-cell target. The distance-to-boundary ablation settles this. If it matches, the contribution narrows to the decomposition and the split uncertainty, which boundary weighting does not provide, because it still forces hard labels on pixels the sensor cannot resolve.

Why not just super-resolve the Sentinel-1 images in image space? Azimuth Super-Resolution for FMCW Radar in Autonomous Driving found that operating on post-formation maps discards phase information and needs far larger networks than working on raw signals. On intensity-only products, an image-space super-resolver can only redraw plausible edges learned from the training regions, which is exactly the kind of geography-specific hallucination we want to measure. We will include one image-space super-resolver followed by the U-Net as a baseline. We expect it to improve apparent sharpness but not low-$\rho$ error on held-out events.

Why not just let region-specific branches absorb the heterogeneity, as STAR does? STAR improved California crop F1 from 0.49 to 0.67 by splitting layers per region. This does not contradict our claim: region branches could be picking up a hidden physical covariate. We test this directly by comparing variance explained by event identity against variance explained by $\rho$. We also check whether conditioning on $\rho$-proxies removes the error clustering that STAR's scan statistic would detect.

Proposed Experiments

All experiments use sen1floods11, with each flood event held out in turn plus the official geographic hold-out, so that results line up with the reported LoRA number. The first decisive check is variance attribution. Across leave-one-event-out folds, we compute the fraction of between-event IoU variance removed by reweighting to a common $\rho$ distribution, for the U-Net, LoRA-Prithvi and CROMA. A share near zero falsifies the core claim.

The second check is the dose-response test. We regress each event's IoU slope under coarsening on its mass of water with $\rho < 2$, and on event identity. The resolution account predicts that the first explains the slopes and the second adds little.

The third check is uncertainty separation. We label held-out error pixels as low-$\rho$ or high-$\rho$ and ask whether the resolvability-error head and the residual confidence score each flag their own type better than maximum logit flags either. Reporting all three checks together tells us whether sen1floods11's geographic hold-out mainly measures geography or the aperture of the sensor.