Phase Up to Sign: Geometry Anchors for Cross-Band Remote Sensing Adaptation

A common trick for adapting segmentation models across imaging conditions rests on the Fourier transform. Split an image into amplitude, which records how much energy sits at each spatial frequency, and phase, which records where edges and shapes are. Then change the amplitude to restyle the image and keep the phase so the object boundaries, and therefore the label masks, stay in place. This works well when the shift is in lighting or color. Aerial imagery often breaks the assumption behind it, because domains there frequently differ in which spectral bands were recorded. A Potsdam tile in red-green-blue and a Vaihingen tile in infrared-red-green show the same kinds of roofs and trees, but vegetation is dark in red light and bright in near-infrared. Every edge between a tree and a road therefore reverses its contrast. In the language of local phase, that edge keeps its position and orientation, but its phase shifts by π. A method that freezes phase also freezes the wrong contrast, and a method that frees phase lets edges drift and labels scramble.

This note proposes to treat geometry as phase modulo π. Edge location and orientation are anchored, while contrast polarity may flip region by region. The first step is to measure, on co-registered RGB and infrared-red-green Potsdam tiles, whether band substitution really appears as class-dependent π-flips. If it does, a band translator trained under a polarity-invariant local-phase constraint should close part of the gap between same-band and cross-band transfer. On Potsdam RGB → Vaihingen IRRG, a frequency-decomposition method reaches 63.09% mean intersection-over-union (mIoU). The same method reaches 68.09% when the source is already IRRG (De-GLGAN). Those five points are the target.

Core Claim

Amplitude-only adaptation cannot express cross-band shift in remote sensing, and full-phase preservation forbids it. Both fail for the same reason: band substitution is a spatially varying, class-dependent sign change in local contrast. A translator that may remix bands freely, but is penalized for moving edges and not for flipping them, should do three things. It should match amplitude-swap methods on same-band shifts. It should beat them and adversarial translators on cross-band shifts. Its translations should be measurably closer to real infrared imagery, which Potsdam provides as paired ground truth.

Grounding in Prior Work

Three papers supply the phase-anchoring technique this proposal modifies. FDA swaps a low-frequency window of target amplitude into a source image and keeps the source phase, reaching 50.45 mIoU on GTA5 → Cityscapes with no trained generator. Phase Consistent Ecological Domain Adaptation adds a loss that penalizes phase drift in a learned translator, so the translator cannot map roofs to vegetation, and reports 50.5% mIoU. DAC shows that ordinary normalization layers distort phase and builds phase-consistent normalization for domain generalization. All three treat the full phase as the thing to preserve, and all assume the source and target share the same three color channels. The file notes explicitly that cross-band phase behavior is unexplored. Our change targets exactly this point: we keep their insistence on fixed edge locations and drop their insistence on fixed sign.

The remote sensing papers motivate the problem and supply the strongest competitors. De-GLGAN (TGRS version; Decomposition-Based Unsupervised Domain Adaptation for Remote Sensing Image Semantic Segmentation, arXiv version) splits features into "low" and "high" frequency streams using attention and convolutions. It never separates Fourier phase, and it relies on adversarial min-max training. Its own numbers expose the band-mismatch penalty that we target. BiFDANet translates between sensors with bidirectional generative adversarial networks (GANs). It is exactly the kind of unconstrained translator that phase-consistency work shows can scramble labels. CrossEarth injects statistically simulated "Earth styles" into a DINOv2 backbone and generalizes across 32 shift settings without phase anchoring. The closest use of phase in remote sensing is the Domain Matching module for reference-based super-resolution (Min et al.). It replaces amplitude and restores phase, but only on RGB reconstruction, and it first converts images to grayscale. That conversion quietly avoids the polarity question. This proposal differs from all of these in one respect: it identifies sign-flipping local phase as the specific signature of band shift and constrains a translator to permit that flip and nothing else.

Research Program

The work starts with a diagnostic that can kill the idea cheaply. Potsdam ships co-registered RGB and IRRG versions of every tile. We filter both versions with a bank of complex log-Gabor filters, which measure local phase (the position of an edge within a small oscillating window) at several scales and orientations. We then compute the phase difference at every pixel where both versions have strong local amplitude, and histogram these differences separately for each ground-truth class. The hypothesis predicts specific peaks: near π for tree and low vegetation in the channel where NIR replaces red, near 0 for impervious surfaces and buildings, and little mass elsewhere. If most differences sit near 0, then ordinary phase preservation is already correct, and the proposal reduces to applying FDA to remote sensing. We would report that result and stop. If most mass is spread uniformly, phase is simply not a geometry anchor across bands, and that is also worth reporting.

If the diagnostic passes, we build a translator $G$ that maps a source image to a target-band image. To keep it from moving edges, its capacity is restricted to a spatially smooth band-mixing field $\tilde{x}(p) = A(p)\,x(p) + b(p)$. Here $x(p)$ is the source pixel vector at location $p$, and $A(p)$ is a band-mixing matrix and $b(p)$ an offset, both predicted at low resolution by a small network and upsampled. Negative entries of $A$ let vegetation invert while roads do not. The translator is pulled toward the target by a lightweight patch discriminator and by matching low-frequency amplitude statistics. A second guard keeps edges in place: a polarity-invariant local-phase loss,

$$ \mathcal{L}_{\pm} = \sum_{k}\sum_{p} w_k(p)\,\big[1 - \cos\big(2(\phi_k^{G}(p) - \phi_k^{s}(p))\big)\big], $$

where $\phi_k^{s}(p)$ and $\phi_k^{G}(p)$ are the local phases of filter $k$ at pixel $p$ in the source and translated images, and $w_k(p)$ is the product of their local amplitudes, so the loss attends only to real edges. Doubling the angle makes a π shift cost nothing, while any other displacement is penalized. Replacing $2(\cdot)$ with $(\cdot)$ recovers a full-phase consistency loss in the spirit of Lao et al. That gives a one-symbol ablation. A segmentation network is then trained on translated source images with source labels, followed by standard pseudo-label self-training on the target.

Baselines

The naive reference is a segmenter trained on Potsdam RGB and applied directly to Vaihingen IRRG with no adaptation. Alongside it we run the matched oracle: a segmenter trained on real Potsdam IRRG and tested on Vaihingen IRRG. That oracle defines the ceiling for perfect band translation, because the geometry is identical and only the bands differ.

The strongest existing method on this exact task is De-GLGAN. It reports 63.09% mIoU on Potsdam RGB → Vaihingen IRRG and 68.09% on Potsdam IRRG → Vaihingen IRRG (De-GLGAN). We rerun it from released code where possible, on our splits, with its published settings. We also report its paper numbers, so that differences in reproduction are visible. Beating it means exceeding 63.09% on RGB → IRRG while staying within about one point of it on IRRG → IRRG, where our method should offer no advantage. We also run BiFDANet as the bidirectional-translation competitor. The file reports no absolute mIoU for it on these transfers; it reports only a 7-point gain over its unidirectional version on Vaihingen → Potsdam.

The phase-based baselines are FDA amplitude swapping, with $\beta$ tuned on a held-out source split, and full-phase consistency translation, which is our translator with $\cos(\Delta\phi)$ in place of $\cos(2\Delta\phi)$. No reported result was found on Lacuna for either FDA or phase-consistency adaptation on ISPRS or LoveDA. These baselines therefore carry no file-reported number, which is itself part of the gap.

For foundation models, we run CrossEarth with its Earth-Style Injection on the same transfer. The file reports only its relative gain of 7.2 mIoU over Rein on Potsdam → Vaihingen, with no absolute value. We also run a variant in which our translated images replace its style injection.

Fairness is enforced in four ways. Every method uses the same segmentation backbone, two backbones are tested in total (a DeepLab-style network and a frozen-then-tuned DINOv2), and every method gets the same self-training stage. Every method receives the same number of tuning trials, and hyperparameters are chosen on the source side only, so no method sees target labels. The decisive ablations are the full-phase versus modulo-π loss, the band-mixing translator with no phase loss at all, and an unrestricted convolutional translator with the modulo-π loss. Together these separate the contribution of the restricted translator from the contribution of the sign-free phase anchor.

Why not just…?

Why not just apply FDA, which is training-free? An amplitude swap keeps every source edge's sign. A tree-road edge that is dark-to-bright in red therefore stays dark-to-bright, even though it is bright-to-dark in NIR. FDA can shift the global brightness of the NIR channel but cannot invert vegetation locally. We expect it to help on same-band shifts and on LoVEDA Urban ↔ Rural, and we report it there honestly. Paired Potsdam data lets us check this directly by measuring how close FDA-translated RGB comes to real IRRG.

Why not just invert channels at random, or invert intensity inside vegetation masks during augmentation? This is the real threat. Label-conditional polarity augmentation uses source masks to flip one band only within vegetation, and it is a cheap, hand-specified version of what our translator learns. We run it as a baseline. If it matches our method, the honest conclusion is that a physics prior plus augmentation suffices. The remaining value would then lie in the diagnostic and in band pairs where nobody knows in advance which classes flip, such as GF-1 ↔ GF-1B NIR differences or seasonal shifts. We test those pairs to see whether learned polarity generalizes where hand rules do not exist.

Why not just use an unconstrained GAN translator, or fine-tune a foundation model? Lao et al. document label scrambling in unconstrained translators. Barbier et al. report that low-shot parameter-efficient tuning of GLIP is weak on DIOR (Barbier et al.), and SpectralX reports LoRA at 52.1% against 59.7% for spectral adapters on WHUOHS. Neither result shows that a strong backbone cannot absorb the band shift on its own, however. CrossEarth in both its plain and modified forms is our test of that possibility.

Decisive Experiments

The experiments run on ISPRS Potsdam RGB → Vaihingen IRRG and Vaihingen IRRG → Potsdam RGB as the main cross-band transfers. Potsdam IRRG → Vaihingen IRRG and LoveDA Rural ↔ Urban serve as same-band controls (LoveDA), and the 4-band Gaofen GF-1 ↔ GF-1B pair adds a sensor-change case. We report mIoU, per-class IoU, and boundary F-score, the last because geometry preservation should show up most clearly at edges. Translation fidelity against real Potsdam IRRG is measured by per-band correlation and edge-location error.

Three outcomes would confirm the idea together. First, the class-wise phase-difference histogram concentrates at 0 and π. Second, the modulo-π loss beats the full-phase loss on cross-band transfer but not on same-band controls. Third, the method recovers a substantial part of the 63.09 → 68.09 gap, with the largest gains on tree and low vegetation, while label-conditional polarity augmentation recovers less.

Three outcomes would refute it. A diagnostic centered at 0 means ordinary phase preservation was already correct. Equal results for the full-phase and modulo-π losses mean sign freedom contributes nothing. Hand-specified polarity augmentation matching the method means the learned translator adds nothing beyond a physics prior.