Satellite road-mapping systems turn imagery into road graphs that are used for routing and travel-time estimates. They break in a characteristic way: wherever a road passes under tree canopy, a building shadow or an overpass, the predicted road mask develops a gap. The gap is a few pixels wide and barely changes pixel accuracy, yet it disconnects a whole part of the routing graph. Current fixes are added from outside the network. Some redesign the architecture or add connectivity losses; others bridge gaps with geometric heuristics after segmentation. None of them asks the earlier question: does the network's internal representation already encode that the road continues through the occlusion, so that the information is dropped only at the final pixel classifier? Or is the continuity never computed at all?
We propose to answer that question directly with representation probing. Probing means training small, deliberately weak classifiers on a frozen network's intermediate activations to test what information those activations contain. Here the probes target two things that matter for road graphs: whether a thin road is present at a location, and whether two road points are connected along the network, including across occluded gaps. We then test with interventions whether the network actually uses what the probes find. The answer decides where engineering effort should go. If connectivity is decodable but discarded, a cheap readout should repair occluded gaps better than geometric stitching. If it is absent, decoder work is futile and the encoder or training signal must change.
Core Claim
The hypothesis is that, in standard road-extraction backbones, pairwise road connectivity across short occlusions stays decodable from intermediate features at a depth where the receptive field covers both sides of the gap, and that this signal is lost in the decoder's per-pixel thresholding, not in the encoder. This is the dense, continuous-geometry analogue of two results from other domains. In LLMs Know More Than They Show, linear probes recover correct facts from intermediate states while the language model's output is a hallucination. In Diagnosing Bottlenecks in Data Visualization Understanding by Vision-Language Models, the authors find that a vision encoder linearly encodes scatter-plot point coordinates with near-100% accuracy, that this information collapses at the handoff to the language model, and they confirm the cause with activation patching.
Both of those studies probe discrete targets: a fact, or the coordinates of an isolated point. Neither involves a dense segmentation network or a continuous thin structure whose value lies in its topology. The claim is falsifiable in both directions, and both outcomes are useful.
The thin-structure literature itself leaves the location of failure unsettled, which motivates the study. SimpleClick reports that even a plain vision transformer with a tiny head fails on very thin objects, which suggests an encoder-side loss. A Revisit to the Decoder for Camouflaged Object Detection argues that the decoder is the main bottleneck: its ENTO decoder recovers thin insect legs without touching the encoder. We treat this disagreement as an empirical question to settle for roads, not as a premise.
Research Program
Benchmarks and labels. The application regime is the standard road benchmarks: SpaceNet at 30 cm resolution with vector centerline labels, used by CRESIv2 and CoANet, and DeepGlobe road extraction at 50 cm (DeepGlobe 2018). The vector labels supply a ground-truth road graph per tile. From that graph we derive connectivity labels for any pair of road points.
Occlusion labels come from two sources. The first is natural occlusion: segments of the ground-truth centerline that fall on pixels flagged as vegetation (using SpaceNet's multispectral bands) or deep shadow (using a brightness threshold), with a small hand-verified subset to estimate how noisy the flags are. The second is synthetic occlusion: we paste real canopy and shadow patches over visible road segments. Synthetic occlusion gives exact counterfactual pairs, the same tile with and without a known gap, which the causal tests require.
Networks probed. We probe four networks that represent the main design families. The first is a ResNet-34 U-Net in the CRESIv2 style. The second is CoANet, which builds connectivity in through directional strip convolutions and a loss on each pixel's eight neighbours. The third is a plain vision transformer segmenter of the SimpleClick type. The fourth is the radar-optical foundation model CROMA, probed on Sentinel-2 imagery co-registered to the SpaceNet city extents. At Sentinel's 10 m resolution, roads are one to two pixels wide, which is the extreme thin regime. CROMA's own use of linear probing was a single whole-image land-cover score, with no layer-wise or geometric probing.
Probe for presence. The presence probe is a linear classifier applied to a single feature location at a time. It predicts whether a road centerline passes through the image region covered by that location, using no pooling over neighbouring locations. This keeps it from becoming a small segmentation head.
Probe for connectivity. The connectivity probe is the central instrument. Given features $f_i$ and $f_j$ at two road locations from the same layer, it predicts whether the two points are joined in the ground-truth road graph:
$$p(\text{connected}_{ij}) = \sigma\big(f_i^\top U V^\top f_j + b\big),$$
Here $\sigma$ is the logistic function, $U$ and $V$ are $d \times r$ matrices with small rank $r$ (we sweep $r$ from 1 to 8), $d$ is the feature width, and $b$ is a bias.
The negatives are what make the probe meaningful. Both points always lie on roads, and they are matched in straight-line distance to the positives. Typical negatives are the two carriageways of a divided highway, or parallel streets with no crossing inside the tile. A probe that only detects "both points are road" or "the points are close" therefore scores at chance. Selectivity is measured against two controls: the same probe trained on a randomly initialized copy of each backbone, and the same probe trained with shuffled graph labels.
The quantity of interest is a layer-wise curve of probe accuracy, split three ways: visible pairs, pairs bridged by an occlusion, and occluded pairs where the gap is shorter or longer than the layer's receptive field.
Causal tests. We run two interventions. In the first, activation patching, we copy layer-$\ell$ features at the gap locations from the unoccluded counterfactual into the occluded forward pass and measure whether the output mask closes the gap. In the second, we project out the probe's subspace at layer $\ell$ and measure whether output connectivity on visible roads degrades. Together these separate "decodable" from "used," which is the standard weakness of probing studies.
Readout. If the diagnosis finds a layer $\ell^\ast$ where connectivity is decodable but lost downstream, we turn the connectivity probe at $\ell^\ast$ into a gap-scoring function inside graph construction. Candidate edges between dangling endpoints of the skeleton are kept when the probe says the endpoints are connected. All other parts of the CRESIv2 vectorization pipeline stay unchanged, and the backbone stays frozen.
Baselines
Naive reference. The do-nothing reference is the unmodified CRESIv2-style pipeline: a ResNet-34 U-Net mask, then smoothing, Zhang-Suen skeletonization (thinning the mask to one-pixel lines) and the default heuristic gap stitching. CRESIv2 reports APLS_length of 0.68 ± 0.21 on SpaceNet test chips. APLS, average path length similarity, compares shortest-path lengths between the predicted and true graphs. For the deliberately simple end of the scale, DeepGlobe reports road IoU of 0.545 for its DeepLab ResNet-18 baseline. All variants are retrained on identical splits so that occlusion-restricted metrics are comparable.
Strongest existing method. The strongest competitor is CoANet, because it attacks the same failure, connectivity, inside the network. It reports F1 of 76.91%, IoU of 62.48% and APLS of 65.53% on SpaceNet, and IoU of 68.37% on DeepGlobe. The file gives no numeric DeepGlobe APLS for CoANet, so we will measure it. We retrain CoANet under our splits and the same vectorization so that APLS is computed identically. Beating it means one of two things. Either our frozen-backbone readout matches CoANet's overall APLS within noise while improving recall of occlusion-bridged connections, or, applied on top of CoANet's own features, the readout raises CoANet's occlusion-bridged recall. The file reports no occlusion-restricted result for any method, so these numbers will be new.
Domain-shift comparator. For the cross-geography setting, the comparator is the topology-based domain adaptation method of Leveraging Topology for Domain Adaptive Road Segmentation. It uses an auxiliary skeleton decoder, a loss forcing the mask to contain the skeleton, and pseudo-label repair. It reports margins of at least 9.8% APLS over competing methods on SpaceNet-to-DeepGlobe. The file gives no absolute value, so we rerun it.
Ablations. The key ablation compares two readouts with matched parameters and training data: one restricted to the probe-selected layer $\ell^\ast$, and one free to use all layers. If the free readout does as well, the diagnosis added nothing beyond "train a better gap scorer," and we will report exactly that. A second ablation replaces the learned score with a purely geometric score (endpoint distance and direction alignment), which isolates whether learned features matter at all. Every baseline and ablation receives the same hyperparameter-search budget.
Why not just…?
Why not just fine-tune CoANet or add a topology loss end to end? This is the most serious threat on raw APLS, and we do not claim the readout will beat a well-tuned end-to-end model overall. CoANet itself reports that it still fails under canopy, tunnels and overpasses, so architectural priors alone do not close occlusion gaps. Our contribution there is to show why. If probes on CoANet's features show that bridged connectivity is decodable, its failure is an extraction failure. If they show it is absent, no decoder change will help, and wider context or occlusion-augmented training is the lever. The CoANet-plus-readout experiment settles whether the diagnosis buys anything beyond CoANet.
Why not just bridge gaps with better post-processing? CRESIv2's heuristics already do this, and they are known to fail at overpasses and complex intersections, where a Euclidean gap rule cannot tell a real continuation from an adjacent parallel road. The geometric-score ablation tests this directly on the matched hard negatives. If it matches the learned score on occlusion-bridged pairs, the representation adds nothing and we will say so.
Why not just attach a strong decoder to the frozen features, as ENTO does? A big enough decoder on all layers may well recover gaps, which is the ENTO argument. In that case probing serves as an explanation, not a performance gain. The all-layer readout ablation is precisely this alternative at matched capacity. The probes remain informative, though, because a decoder that succeeds does not reveal whether it is reading stored continuity or hallucinating plausible roads. Only the patching and projection tests distinguish those two cases.
Why not just look at the output masks? A mask cannot distinguish "the encoder never saw the continuation" from "the encoder saw it and the threshold pruned it." The two call for opposite fixes. The scatter-plot study shows that output-only evaluation would have misplaced the bottleneck entirely.
Proposed Experiments
The first decisive check is the layer-wise connectivity curve on the synthetic-occlusion counterfactuals. The interesting result is high selectivity for occlusion-bridged pairs at some intermediate layer, well above the random-backbone control, coexisting with a broken output mask. A null result is equally informative: if bridged-pair accuracy stays at chance at every layer while visible-pair accuracy is high, encoders do not compute continuity across occlusion, and the decoder-bottleneck story fails for roads. We expect decodability to appear only once a layer's receptive field spans the gap, because the network cannot represent continuity it has no access to. Plotting accuracy against gap length divided by receptive-field size tests this prediction.
The second check is causal. Patching unoccluded features into gap locations at $\ell^\ast$ should close output gaps. Removing the probe subspace should fragment visible roads. If patching works but subspace removal does nothing, the network carries the signal without relying on it, and we report decodability without use.
The third check is practical, on SpaceNet and DeepGlobe. We report overall APLS, F1 and IoU against the baselines above, and recall of ground-truth connections that cross natural occlusions, using the manually verified subset to bound label noise. We also run the Las Vegas-to-Khartoum split, where DeepGlobe documents a building-detection F1 collapse from 0.885 to 0.544. There we test whether connectivity probes transfer across cities better than presence probes. That would support the claim, from the topology adaptation work, that road skeletons are more domain-invariant than road appearance, now measured inside the representation.
The fourth check uses CROMA on Sentinel-2. It asks whether a general-purpose remote-sensing foundation model encodes sub-pixel roads at all. Its answer tells practitioners whether adapting such models for road mapping can rely on a light head or needs the backbone retrained.
References
- LLMs Know More Than They Show: On the Intrinsic Representation of LLM Hallucinations
- Diagnosing Bottlenecks in Data Visualization Understanding by Vision-Language Models
- SimpleClick: Interactive Image Segmentation with Simple Vision Transformers
- A Revisit to the Decoder for Camouflaged Object Detection
- City-Scale Road Extraction from Satellite Imagery v2: Road Speeds and Travel Times
- CoANet: Connectivity Attention Network for Road Extraction From Satellite Imagery
- DeepGlobe 2018: A Challenge to Parse the Earth Through Satellite Images
- CROMA: Remote Sensing Representations with Contrastive Radar-Optical Masked Autoencoders
- Leveraging Topology for Domain Adaptive Road Segmentation in Satellite and Aerial Imagery