Seam-Free Satellite Dehazing by Splitting Scene-Wide Color From Tile-Local Detail

Satellite images come in strips far too large to fit in GPU memory. So any neural network that removes haze or thin cloud has to process them in tiles. Today's best remote-sensing dehazers are "kernel-free": they map a hazy image straight to a clean one, without first estimating a physical model of light scattering. These networks are trained on small crops, and most contain components that look at the whole input, such as attention layers, long directional scans, and color-calibration modules. Run tile by tile, each tile forms its own opinion of how bright and what color the haze is. The mosaic then shows blocky shifts in brightness and color at tile borders. Blending overlapping tiles hides these shifts by averaging them, which smears exactly the fine field edges and building outlines that make the imagery useful. The same scene-wide haze color is also what changes most from one region to another, so it is where geographic domain shift does its damage.

We propose splitting the restoration into two parts that tile differently. A small global branch runs once on a heavily downsampled copy of the whole strip. It predicts a smooth, full-resolution "correction field" that removes the haze veil and color cast. A local detail branch has a strictly bounded receptive field, meaning each output pixel depends only on input pixels within a fixed distance. It reads the hazy tile together with the matching window of the shared correction field and restores high-frequency structure. Every tile reads the same field, and the local branch cannot see past its known radius. Tiled output is therefore identical to whole-image output, with no blending at all. The design also isolates where regional adaptation should happen: only the small global branch needs to change for a new region.

Core Claim

In remote-sensing dehazing, tile seams and geographic fragility come mainly from the low-frequency part of the answer: airlight and color cast. The high-frequency detail can be restored locally and transfers well across regions. Our claim is that factoring the network along this line gives three things. It makes tiled inference exact rather than blended. It keeps thin structures sharper than overlap-and-average stitching. And it allows adaptation to an unseen region by retuning only a few hundred thousand global-branch parameters. Written out, the restored tile is

$$\hat J_T = L_\theta\big(I_T,\; [\,U(G_\phi(D(I)))\,]_T\big),$$

where $I$ is the full hazy strip and $T$ indexes a tile. $D$ is fixed downsampling, and $G_\phi$ is the global branch, which outputs a coarse per-pixel affine color transform. $U$ is smooth bicubic upsampling of that transform to full resolution, and $[\cdot]_T$ crops it to tile $T$. $L_\theta$ is the local branch, whose receptive-field radius is $r$. If each tile is processed with a margin of at least $r$ pixels that is then discarded, $\hat J_T$ equals the corresponding crop of whole-strip inference exactly. The only coupling between tiles is the shared field, and every tile sees it identically.

Relation to Prior Work

The closest work is SGLC. It also pairs a global stage, which subsamples the whole 4K image on a grid, with a local enhancer run on overlapping windows. It reaches 25.43 dB PSNR on HD-NH-HAZE. SGLC motivates our split but differs in the part that matters: its windows are stitched with spline-weighted blending (the "MOPS" smoother), so tile disagreements are averaged rather than prevented. It takes about 9 minutes per 4K frame and was only tested on ground-level photos. We remove blending entirely by bounding the local receptive field, and we test on satellite imagery under region shift.

Continuous Patch Stitching supplies the method for exactness. For learned compression, it shows that seams come from padding and truncated receptive fields, and it computes a minimum overlap that makes block-wise processing match full-image processing. It does not address restoration, skip-connected encoder–decoders, or components with unbounded context. Our contribution is to show that a dehazer can keep scene-wide context and still tile exactly, by routing all unbounded context through one shared low-resolution field.

Current remote-sensing dehazers set the accuracy bar and motivate the geographic question. RSDehamba is a 1.80M-parameter U-Net built on Mamba, a state-space sequence model, with scans in four directions. It reaches 25.91 dB / 0.9157 SSIM on synthetic Sate1K. MCAF-Net adds axis-oriented attention and self-supervised color calibration, and reaches 24.11 dB / 0.6680 SSIM on the real RRSHID benchmark. Both use long-range or scene-level components, and both are trained and tested on fixed patches from narrow geographies. Neither addresses tiling or out-of-region deployment. We use both as the local-detail backbone, restricted to bounded context, and as baselines.

The geographic side constrains the design. Mission Critical reports drops of up to 30% in unseen regions. Trusting Small Training Dataset finds that a geographically diverse training set beats a local one on a French test city (Kappa 0.32 vs 0.12). Fine-tune Smarter finds that LoRA beats full fine-tuning on the Sen1Floods11 geographic hold-out (87.57% vs 82.07% mIoU). LoRA is low-rank adaptation, which trains small added weight matrices while freezing the original model. These results are all for classification or segmentation. They suggest that touching fewer parameters helps under region shift, and our global-branch-only adaptation tests whether that holds for pixel restoration. STAR branches networks into region-specific parts over discrete spatial units. We deliberately avoid discrete partitions, because haze varies continuously along a swath.

Finally, the global field-boundary map stitches overlapping U-Net predictions with Gaussian weighting and reports over-fragmentation of small fields. This is evidence that blended stitching harms fine boundaries, and its FTW data give us a downstream test of detail.

Research Program

We will build the local branch in two versions: a padding-free convolutional U-Net, and RSDehamba with its scans confined to fixed windows. For each version we compute the receptive-field radius $r$ analytically, including skip connections. Within the local branch, skip connections carry encoder features to the decoder at matching resolution. Those features are first corrected by the global field, which multiplies and shifts them. This is a direct test of the "haze leakage" concern: raw early features still contain the haze veil and may pass it straight to the output. The global branch is a lightweight network run at 1/16 resolution. It predicts the affine color field and is supervised through the full-resolution reconstruction loss. It also has a smoothness penalty, so that the field carries only low-frequency content.

We will then measure three properties. First, exactness: the maximum absolute difference between tiled and whole-image outputs, which should be zero up to floating-point error. Second, detail: gradient-domain error on thin structures, and downstream field-boundary F1. Third, regional transfer: we freeze $L_\theta$ and adapt only $G_\phi$, using a few labeled pairs, or no pairs with a self-supervised color-consistency objective like the calibration module in MCAF-Net.

Baselines

The naive reference point is non-overlapping tiling of RSDehamba and MCAF-Net, each trained with its published recipe on the same data, with tile borders left as they are. This measures how large the seam problem actually is. If seams are small for realistic tile sizes, the proposal loses much of its motivation, and we will report that.

The strongest existing competitor for the full-scene setting is SGLC-style processing: a grid-subsampled global stage plus overlapping local windows merged with spline blending. We re-implement it on the same remote-sensing training data, with the same parameter budget and hyperparameter search budget. It is the strongest competitor because it already combines global context with local windows; the only difference is that it blends where we make tiling exact. No reported result found on Lacuna for SGLC or for any seam metric on satellite data. Its only reported number is 25.43 dB on HD-NH-HAZE. We will also run overlap-and-blend versions of RSDehamba and MCAF-Net with Gaussian weighting, the scheme used for the global field map, at 25% and 50% overlap.

For patch-level accuracy, the numbers to beat are MCAF-Net at 24.11 dB / 0.6680 on RRSHID and RSDehamba at 25.91 dB / 0.9157 on Sate1K, both as reported in their papers. Our method must stay within about 0.3 dB of these numbers on standard crops, since we give up some local context. It must also match or beat them on full scenes and on held-out regions.

Two ablations isolate the key ingredient. In "global-only", the local branch is replaced by the identity, so the output is just the upsampled color transform applied to the input. In "local-only", the shared field is removed, so each tile corrects its own color. Beating the SGLC-style baseline means three things: lower seam error and tile-vs-whole inconsistency, equal or higher PSNR on the full scene, and higher downstream boundary F1, all at a fraction of its runtime.

Why not just…?

Why not just overlap tiles heavily and blend with a Gaussian window? This is the most serious threat. With enough overlap, blending may hide low-frequency seams well enough for most users. The weakness is that blending averages two different color estimates, and averaging misaligned edges produces ghosting on thin structures. The field-boundary map reports small-field over-fragmentation under exactly this scheme. We will settle this by sweeping the overlap ratio and reporting seam error, boundary F1, and compute together. If 50% overlap with blending matches our boundary F1 at similar cost, we will say so, and our remaining contribution reduces to exactness and cheaper adaptation.

Why not just downsample, dehaze globally, and transfer the correction back up? Haze is mostly low-frequency, so a low-resolution correction applied to the full-resolution input may be enough. SGLC's evidence against pure downsampling (Uformer at 14.90 dB) involved restoring the downsampled pixels themselves, not transferring a correction, so it does not rule this out. Our global-only ablation is exactly this baseline. If it matches the full model, the local branch is unnecessary. We expect it to fall short on thick haze and cirrus, where texture is actually lost and not merely tinted. The split of RRSHID by haze density will show where the two diverge.

Why not just fine-tune the whole dehazer on imagery from the target region? Paired clean and hazy imagery is scarce in any new region. The geographic evidence we have also points the other way: diverse training beats local training for change detection, and LoRA beats full fine-tuning under region hold-out for segmentation. We test this directly for restoration by comparing full fine-tuning, LoRA on the whole network, and global-branch-only adaptation, with the same number of target pairs.

Why not use a physical scattering model, which generalizes by construction? MCAF-Net's authors report that uniform-airlight assumptions fail on real satellite haze, and the gated-context and FD-GAN lines find direct mapping consistently stronger. Our global branch keeps the useful part of the physical view, a smooth spatially varying veil, without the error-prone algebraic inversion. We will still include dark-channel dehazing as a floor.

Proposed Experiments

For patch-level accuracy, we train on RRSHID and Sate1K and report PSNR and SSIM on their test sets. For geographic transfer inside RRSHID, we hold out one landscape type at a time: urban, agricultural, or coastal. We also train on synthetic Sate1K and test on real RRSHID, which is a shift from synthetic to real data. HRSI100 is scored with no-reference image-quality metrics only, since it has no ground truth.

For full-scene seams, we synthesize haze with spatially varying airlight and transmission over large Sentinel-2 mosaics from FTW, spanning many countries. We then tile at 128–512 pixels and report three quantities: tile-vs-whole inconsistency, a boundary-gradient seam error along tile lines, and runtime per 10k×10k pixels.

For detail, we run a fixed field-boundary U-Net, trained on clean FTW, on each method's dehazed output and report boundary F1 per country. We also run the pix2pix-style setting from RICE-II on Landsat-8 cirrus as a real-degradation check.

The decisive check has three parts. On synthetic hazy FTW mosaics from held-out countries, the method must show near-zero tile inconsistency and higher boundary F1 than overlap-blended RSDehamba and the SGLC-style baseline at matched compute. Global-branch-only adaptation must recover most of the gap to full fine-tuning. The global-only ablation must fall measurably short on thick haze. If any of these three fails, the claim that seams and region shift live mainly in the low-frequency color field is wrong, and the note will say so.