CrossViewGate

When Satellite and Street Views Disagree
Cross-view complementarity in disaster damage assessment, studied through conflict cases

Accepted as a lightning talk at the 5th ACM SIGSPATIAL Workshop on Searching and Mining Large Collections of Geospatial Data (GeoSearch 2026), Riverside, CA, 3 November 2026. Yifan Yang and Lei Zou, Department of Geography, Texas A&M University.

Satellite and street-view imagery see different parts of a damaged building: roofs from above, facades and debris from the ground. Fusion methods usually report that combining the two views raises average accuracy, but average accuracy hides whether the complementary evidence is actually used, because for most buildings the two views agree. CrossViewGate studies the conflict cases, the samples on which independently trained single-view models disagree, as a diagnostic subset. On conflicts, at least one view carries evidence the other lacks, so they are where fusion either uses complementary information or does not.

Research questions

The paper asks four questions on paired street/overhead collections from the 2025 Eaton wildfire and Hurricanes Ian and Milton (five seeds, one converged protocol):

1. How much complementary headroom do fusion methods leave on conflict cases? We compare each method with an oracle that selects whichever single view is correct.
2. Can a simple, readable gate recover part of it from cues about whether the ground image actually shows the building?
3. Does aiming the ground view at the building change the value of fusion? We vary framing within a dataset by controlled cropping.
4. Does the spatial density of conflicts, which needs no labels once the models are trained, track damage?

Findings

• The oracle gap. Conflicts make up 9.6–33.2% of the test data. Fusion improves conflict accuracy over either view, yet an oracle that picks the correct single view per conflict remains 0.37–0.48 above the better single view and 0.19–0.32 above the best evaluated fusion method. Current fusion leaves most view-specific evidence unused.
• The fusion gate. A linear mixture over 11 features (building visibility from a frozen segmenter, calibrated per-view confidence, cross-view disagreement) assigns a per-sample weight to the street, overhead, and end-to-end fusion predictions. Where ground photographs are aimed at the building (Eaton), it beats end-to-end fusion by +0.072 conflict accuracy and +0.018 on the full test set (both p < 10⁻⁴) and calibrated averaging by +0.051 (p = 0.0001). On the panoramic hurricane data, where the building is rarely prominent, it is statistically indistinguishable from end-to-end fusion. The two-expert variant reads as one rule: weight the street view when it is confident and the building is centered in the frame.
• Field-of-view intervention. Cropping hurricane panoramas to building-centered 90° windows raises the conflict gain of fusion (Ian 0.064 → 0.109, Milton 0.067 → 0.149); random crops of identical geometry do not, in 6/6 dataset-seed pairs. Building-oriented framing is a driver of which view carries the evidence.
• Label-free damage indicator. Tile-level conflict density correlates with wildfire damage without annotations (Spearman r = 0.615, 25 tiles; nominal p, damage is spatially autocorrelated). Single-view uncertainty anti-correlates. The hurricane signal is weaker (Milton r = 0.289, n.s.).
• A negative result. Once single-view models are converged and calibrated, probability averaging is as strong as end-to-end learned fusion. Learning pays off when it decides per sample which view carries the evidence.

Which view carries the evidence depends on hazard and capture setting. On property-centric wildfire photographs, visibility cues identify when the street view is decisive. On panoramic hurricane imagery, calibrated averaging or end-to-end fusion suffices, and framing ground photographs on structures raises the value of fusion. Mining conflicts turns "does fusion help?" into "which view carries the evidence, where, and why?".

Results

Conflict-case test accuracy, five-seed means.

Method	Eaton wildfire	Hurricane Ian	Hurricane Milton
Street only	0.486	0.529	0.557
Overhead only	0.475	0.367	0.407
End-to-end cross-view fusion	0.699	0.624	0.649
Calibrated probability averaging	0.732	0.617	0.626
Fusion gate (linear)	0.768	0.618	0.631
Oracle single view	0.961	0.896	0.964

Fusion gate. Accuracy by method and the coefficients of the two-expert (street vs. overhead) linear gate. Positive values shift weight toward the street view.	Field-of-view intervention. Same panorama, same crop geometry; the building-centered crop raises the fusion gain, the random crop does not.
Label-free damage map. Conflict density tracks labeled damage; uncertainty density does not.	The oracle gap. Each bar spans best single view to oracle; dots place each method.

Full tables, pooled paired tests, calibration decomposition, and ordinal metrics are in docs/results/ (_v2 = headline five-seed protocol).

Quick start

pip install -e .
bash scripts/run_multiseed_v2.sh                                   # converged 5-seed suite
python scripts/analyze_calibration_fusion.py --multiseed-root outputs/multiseed_v2 --seeds 42,123,456,789,1011
python scripts/train_reliability_gate.py     --multiseed-root outputs/multiseed_v2 --seeds 42,123,456,789,1011
python scripts/analyze_pooled_seed_tests.py  --multiseed-root outputs/multiseed_v2 --seeds 42,123,456,789,1011
python scripts/build_paper_figures.py                              # regenerate figures/

Imagery stays local; sources and split provenance are in docs/datasets.md. The portable georeferenced Hurricane Ian split lives in disaster-crossview-datasets.

crossview_conflict/   package: data, models, training, decision
scripts/              pipeline and analysis
tests/                protocol, provenance, and integrity tests
configs/              frozen experiment contracts
paper/                GeoSearch '26 short paper (LaTeX + PDF)
figures/              publication figures
docs/                 datasets, protocols, results, project website

Terminology. The paper calls the per-sample weighting model the fusion gate. The code, result logs, and tests keep their original identifier reliability_gate (for example scripts/train_reliability_gate.py and docs/results/reliability_gate_results_v2.md) so that every number in the paper still maps to the log it was checked against. The two names refer to the same model.

Beyond classification

A follow-up line asks when evidence is sufficient, which view to acquire next, and when to defer to a human. Development results so far are transparent No-Go or sensitivity-only findings, recorded with their consumed-data ledger in docs/results/.

Citation

@inproceedings{yang2026disagree,
  title     = {When Satellite and Street Views Disagree: Cross-View Complementarity
               in Disaster Damage Assessment},
  author    = {Yang, Yifan and Zou, Lei},
  booktitle = {Proceedings of the 5th ACM SIGSPATIAL International Workshop on
               Searching and Mining Large Collections of Geospatial Data (GeoSearch '26)},
  year      = {2026},
  doi       = {10.1145/3849732.3857333},
  eprint    = {2610.04327},
  archivePrefix = {arXiv},
  note      = {Lightning talk. Code: https://github.com/rayford295/CrossViewGate}
}

Acknowledgments

Supported by the U.S. National Science Foundation (Award 2318206) and the Gulf Research Program of the National Academies of Sciences, Engineering, and Medicine (SCON-10000653 and SCON-10001536).

Formerly CrossViewConflict (old URLs redirect; the package import name stays crossview_conflict). Licensed under Apache 2.0.