Road maps extracted from satellite images break easily. A single pixel lost across a road cuts a route in two. Breaks are most common where a model has not seen the local landscape before, and where tall buildings, viewed at a slant, lean over the street and hide it. Our proposal rests on one physical fact: a road lies on the ground. Once several satellite views of the same place are registered to ground coordinates, the true road network is the same in every view, while building lean, shadows, and occlusion change from view to view. We propose to use this invariance in two ways. First, as free supervision: in an unlabeled new region, disagreement between the road topologies predicted from different look angles signals an error, and we penalize it with a matching loss that compares connected pieces and loops only where they physically overlap. Second, as an inference procedure: the model reads the views one at a time and conditions each prediction on the road network accumulated so far, so a segment hidden at one angle can be recovered from another angle where it is visible.
The opportunity comes from three bodies of work that have not met. Topology-preserving road segmentation works on single top-down tiles. Multi-angle satellite benchmarks exist but measure each view separately. Methods that build up a scene one view at a time, conditioning each new view on earlier ones, exist only for generating natural images. Our claim is concrete and testable. When a road model trained in one set of cities is adapted to an unlabeled multi-angle target city, cross-view topological agreement should improve route-level connectivity (the Average Path Length Similarity score, APLS, which compares shortest-path lengths between the predicted and true graphs) beyond single-view topology-aware adaptation. The largest gains should fall on road segments that some look angles hide.
What the literature gives us, and what it leaves open
Three groups of papers motivate the gap. The first shows how badly road topology fails under geographic shift. Leveraging Topology for Domain Adaptive Road Segmentation in Satellite and Aerial Imagery adapts across cities using two output heads (road mask and centerline), a loss forcing the mask to contain the centerline, and pseudo-labels repaired by an 8-neighborhood gap-filling search. It reports that generic feature alignment fails because overhead imagery has no fixed scene layout. Dey et al. extract road graphs by tracing ridges in the probability map with discrete Morse theory. They note that one global persistence threshold cannot suit both paved Las Vegas and unpaved Khartoum. Both methods see a single top-down image. When a road is hidden, they can only guess from local 2D context. That is the gap we target.
The second group shows that viewing geometry, and not resolution, is the main obstacle. SpaceNet MVOI captures 27 look angles over Atlanta in one satellite pass. Building-segmentation F1 for TernausNet falls from 0.62 at nadir (looking straight down) to 0.22 beyond 40° off-nadir. A control that downsampled nadir images to off-nadir resolution "barely dropped." This rules out resolution augmentation as a fix and shows that lean and occlusion are the culprits. MVOI evaluates each angle on its own and labels only buildings, so it supplies our multi-view target domain but no road result.
The third group supplies methods. TopoMatch matches topological features (connected components and loops, tracked across probability thresholds by persistent homology) between a prediction and a target. It builds the comparison in a shared space, $\max(L, G)$, so features pair only if they overlap. It also shows that the older Wasserstein matching pairs features with spatially wrong partners up to 99.4% of the time. We reuse induced matching, but between two predictions from different views rather than between a prediction and ground truth. That use is new, and it is why spatial correctness matters: matching a road in view A to an unrelated road in view B would reward hallucination. clDice supplies a cheap differentiable centerline overlap that we use as a lighter companion loss. For sequential conditioning we borrow from novel view synthesis. Novel View Synthesis with Diffusion Models conditions each step on one randomly chosen earlier view (stochastic conditioning), which avoids attention over all past views and acts as a regularizer. ViewFusion combines per-view predictions with learned, softmax-normalized per-pixel weights. Those weights are exactly what is needed when one view says "occluded, no evidence" and another says "road."
One constraint shapes the design. Generative pipelines that condition on their own outputs accumulate drift: early errors compound over long sequences in Xiang et al. and Morpheus. We therefore do not generate pixels or synthesize new views. Every view we condition on is a real observation. The only thing carried forward is a ground-plane road-probability state, which the next real view can overwrite.
Core Claim
Road topology in ground coordinates is invariant across look angles, while occlusion and lean are not. Disagreement in topology between views is therefore an observable error signal that needs no labels. In an unlabeled target geography it is a more reliable self-training signal than single-view connectivity heuristics. A model that reads views in sequence, with learned per-pixel trust in each view, can turn that signal into recovered connectivity at test time.
Research Program
The model is a road segmentation network with a parameter-efficient adapter on a geospatial foundation backbone. We follow Fine-tune Smarter, Not Harder: LoRA (small low-rank weight updates) beat full fine-tuning on a geographic hold-out (87.57% vs 82.07% mIoU on Sen1Floods11). Starting from that finding keeps our adaptation baseline strong. The network takes the current view $x_t$, its look-angle metadata, and a ground-registered state $S_{t-1}$, the fused road-probability map from earlier views. It outputs a per-view road map $P_t$ and a per-pixel trust map $w_t$. The new state is the trust-weighted combination of the current and earlier maps, in the spirit of ViewFusion's weighting. During training we use stochastic conditioning: the state is built from a random subset, sometimes a single earlier view, in random order. This keeps the model from depending on a fixed camera path, the trajectory dependence that limits Morpheus.
Adaptation uses labeled source cities (SpaceNet 3: Las Vegas, Paris, Shanghai, Khartoum) and the unlabeled multi-angle target (MVOI Atlanta). On the target the objective is
$$\mathcal{L}_{\text{target}} = \sum_{t} \mathcal{L}_{\text{IM}}\big(P_t,\ \mathrm{sg}[\hat{S}_{-t}]\big) + \lambda\, \mathcal{L}_{\text{clDice}}\big(P_t,\ \mathrm{sg}[\hat{S}_{-t}]\big),$$
where $P_t$ is the ground-registered road prediction from view $t$, $\hat{S}_{-t}$ is the fused state built from all views except $t$, $\mathrm{sg}$ stops gradients through that target, $\mathcal{L}_{\text{IM}}$ is TopoMatch's induced-matching loss, $\mathcal{L}_{\text{clDice}}$ is the soft centerline overlap, and $\lambda$ weights the two. Leaving view $t$ out of its own target is what forces information to flow from views where a segment is visible to views where it is hidden. Persistent homology costs roughly cubic time in pixel count, so the matching loss runs on 256-pixel ground tiles, with clDice covering longer-range continuity.
One assumption needs checking before anything else. MVOI views must agree at ground level after registration, with only elevated structures displaced. We will measure the residual offset of road-level features, such as building footprint bases and intersections, across angles. If misregistration exceeds a few pixels, we add a small learned alignment step before fusion.
Data construction
MVOI has no road labels. For evaluation only, we will hand-annotate road centerline graphs for a set of Atlanta tiles held out spatially, with a buffer, from any tiles used in adaptation. This follows the spatial-blocking practice argued for in Position: Mission Critical. Using MVOI's footprint polygons and the known lean direction of each look angle, we tag every annotated road segment as "occluded in some views" or "open in all views." Every comparison below is reported separately for the two strata.
Baselines
The naive reference is a single-view model trained on the SpaceNet 3 cities, applied to each Atlanta look angle independently with no adaptation, then fused by per-pixel averaging in ground coordinates. It tells us how much of any later gain comes from adaptation as opposed to simply having many views.
The strongest existing competitor is the topology-aware adaptation method of Leveraging Topology for Domain Adaptive Road Segmentation. For SpaceNet→DeepGlobe it reports beating competing methods by at least 6.6% IoU, 6.7% F1, and 9.8% APLS. Those are margins over other methods, not absolute values. No reported road result found on Lacuna for any method on MVOI. We will rerun it on our splits with the same LoRA backbone, the same source data, the same unlabeled target tiles (treated as a pool of single images), and the same tuning budget in GPU-hours and hyperparameter trials. We will also give it the same averaging fusion at test time, so both systems get multi-view input at inference and differ only in how they are trained and how they fuse. Beating it means a higher APLS on held-out Atlanta tiles, with confidence intervals from resampling tiles, and a clearly larger gain on the occluded-segment stratum than on the open stratum.
Two further reference points calibrate the graph-extraction stage. The first is a TopoMatch-trained single-view model with Dey et al.'s Morse graph extraction. On SpaceNet AOI_2, Dey et al. report APLS 0.796 semi-automatic and 0.592 label-free. The second is TD-Road's direct graph prediction, which reports APLS 65.74 on City-Scale against Sat2Graph's 63.14. We use these numbers only to confirm our reimplementations behave sensibly in-domain, not as targets on Atlanta.
Three ablations isolate the key ingredients. Replacing induced matching with a pixel-level cross-view consistency loss tests whether topology matters, or whether plain agreement across views does the work. Replacing it with Wasserstein matching tests whether spatially correct pairing matters. Removing the sequential state, so that each view is predicted independently under the cross-view loss and fused by averaging, tests whether conditioning adds anything beyond the training signal.
Why not just…?
Why not just segment each view and fuse the probability maps, without any learning across views? This is the most serious threat, and the design treats it as a baseline, not a strawman. Averaging has a known failure: a view in which the road is hidden votes "no road" and dilutes the views that see it. Taking the maximum instead admits every view's false positives. Whether learned trust weights beat both is an empirical question. The ablation without the sequential state answers it directly. If averaging plus our cross-view training matches the full model, we will report the contribution as the training signal alone and drop sequential inference from the claim.
Why not just generate the hidden road segments with a multi-view diffusion model, as in view-synthesis work? Filling in roads that no sensor observed is the wrong behavior for a map used for routing. The view-synthesis literature also documents compounding drift when models condition on their own outputs (Xiang et al., Morpheus). Real multi-angle passes already contain the hidden pixels, so we condition only on real views.
Why not just fine-tune a geospatial foundation model with LoRA and pseudo-labels? LoRA is a real improvement for geographic hold-outs (Escofet et al.), which is why every arm of our comparison uses it. Single-view pseudo-labels in an occluded street are still broken, however, and the same paper found that adding coordinate metadata gave negligible gains, so geography alone will not repair connectivity. The comparison against LoRA plus single-view topology-aware pseudo-labels measures what cross-view agreement adds on top.
Why not just use cross-view pixel consistency, a standard self-training trick, and skip persistent homology? This could plausibly capture most of the gain, since roads are thin and pixel agreement partly implies connectivity. The pixel-consistency ablation settles it. Our prediction rests on the observation that Dice-style overlap can be high while routes are severed (Stucki et al., clDice). The difference should therefore show up in APLS and Betti error (the count of wrongly added or missing components and loops), not in IoU. If it does not, the topological loss is not earning its cost.
Decisive Experiments
All metrics are computed on the spatially held-out, hand-annotated Atlanta tiles: APLS, TOPO precision and recall, Betti error, and IoU. Every result is reported for the occluded and open strata and as a function of look angle. Three outcomes would support the core claim. First, the full method beats the topology-aware adaptation baseline on APLS. Second, the gain is concentrated in the occluded stratum and at the steepest angles. A uniform gain would point to a generic regularization effect rather than recovery of hidden segments. Third, both topology ablations lose most of the APLS gain while keeping similar IoU.
A geographic control separates the two kinds of shift. We repeat the comparison with labeled source data that includes Atlanta-like nadir imagery, so that viewing geometry is the only remaining shift. If cross-view agreement helps equally with and without the change of city, its value lies in handling occlusion. If it helps more across cities, it is also acting as a cure for geographic shift in pseudo-labels, which is the stronger result.
A final check guards against invented roads. We count predicted graph edges with no annotated counterpart and require that the full method does not increase them relative to averaging fusion. Gains in connectivity must come from finding roads that exist, not from adding ones that do not.
References
- Leveraging Topology for Domain Adaptive Road Segmentation in Satellite and Aerial Imagery
- Road Network Reconstruction from satellite images with Machine Learning Supported by Topological Methods
- SpaceNet MVOI: A Multi-View Overhead Imagery Dataset
- Topologically Faithful Image Segmentation via Induced Matching of Persistence Barcodes
- clDice - A Novel Topology-Preserving Loss Function for Tubular Structure Segmentation
- Novel View Synthesis with Diffusion Models
- ViewFusion: Learning Composable Diffusion Models for Novel View Synthesis
- 3D-aware Image Generation using 2D Diffusion Models
- Morpheus: Text-Driven 3D Gaussian Splat Shape and Color Stylization
- Fine-Tune Smarter, Not Harder: Parameter-Efficient Fine-Tuning for Geospatial Foundation Models
- Position: Mission Critical – Satellite Data is a Distinct Modality in Machine Learning