Swiss Landscape Signatures · a computational atlas

Landscape Signatures in Switzerland

Landscape Signatures are the patterned subset of landscape structure that is selected through attention, stabilized, through repeated experience, and made socially meaningful through memory, use, and practice.

Explore Switzerland's perceptual structure Explore the Research

0images

0model backbones

0parameter settings

0archipelago readings

0persistent density modes

0communities — cuts, not types

What is the visual structure of the Swiss landscape?

The question. Does the everyday visual landscape consist of discrete types — an archipelago of islands with genuine gaps between them — or does it form a single continuum on which every typology is a human-drawn set of contour lines?

Clustering seven million images always returns clusters. The scientific work is telling “clusters because the data is discrete” from “clusters because we cut a gradient”. Four mathematically independent instruments — a dip test, persistent homology, density-based clustering, and a certified density-mode persistence gap — were validated on known answers first, then pointed at the full 7.2 million embeddings.

None reads archipelago. Not on any backbone, not under any of 546 parameter settings. The certified gap finds zero persistent density modes.

hypothesis — discrete types

The raw material — 7,215,142 Mapillary frames · DINOv3 ViT-7B/16 · ViT-H/16+ · V-JEPA 2.1

Five independent layers of evidence

Each layer is methodologically independent of the ones below it; the verdict survives all five.

What varies along the continuum

Thirty-two shared axes, extracted from the three backbones’ common geometry and named by two independent vision-language models from the pole photographs themselves. Per-axis cross-model agreement 0.76–0.98, always reported.

The continuum, mapped

A continuum still supports operational cartography — gradients, distinctiveness and thresholds instead of hard type boundaries.

The signature field — all 32 axes in color

Plate IThe signature field All 32 shared axes at once: each place colored by its position in the full signature space — similar color, similar overall visual character.

Plate IILocalizability How geographically identifiable each place looks. Beats a coverage-matched null on 88% of the grid.

Plate IIIWhere character changes fastest Nugget-corrected perceptual edges. Align with official typology boundaries at 2.4× chance.

Regions of the gradient, not types

Community detection on the 7.2 M-node similarity graph partitions the continuum into 116 visual communities. They are real, stable regions — 13 reproduce strongly and 23 moderately across all three backbones — but they are not density modes: move a single detection threshold and the count of “discrete signatures” slides from 54 to 3.

Each community is characterized by two independent vision-language models, profiled along the shared axes, and grounded against the official Swiss landscape typology.

Browse the 116 communities

Three models, one geometry

The gradient is a property of the landscape, not of a network: two DINOv3 scales and a different model family agree on the low-dimensional geometry while drawing different fine cut lines.

Live model Predict from a photograph Upload any image — the trained dual-encoder returns its position on the continuum and the implied geography. Rigor Why the result can be trusted Known-answer certification, retired instruments, pre-registration, and a complete forensic audit trail. Papers Read the full analysis The capstone synthesis, consolidated results, and methods — rendered from the living project documents.

Leiden identifies visual-embedding communities; communities stable across the robustness checks are treated as candidate landscape signatures — and after the full battery, almost none qualify as discrete, because the landscape itself is a continuum.

Plate —

The 116 communities

Regions of the gradient found by Leiden on the raw-embedding similarity graph — named by their dominant axes, characterized by two vision-language models, grounded against the official typology. None survives the discreteness gate robustly: they are cuts through the continuum.

sort with photographs

The evidence

Every instrument was validated on known answers before being believed on real data; every major correction is on the record.

Direct instruments — full 7.2 M, three neighborhood scales

clusterability battery · manifold null

“How many discrete signatures?” is a threshold artifact

modal-separation gate sweep · native4096

Communities passing the modal-separation gate as the responsibility threshold ρ sweeps. There is no plateau — no natural count of discrete types. At the pre-registered operating point the full gate passes 3 of 116.

Partition stability under perturbation

size-weighted mean · reference config

Fraction of images keeping their community when the pipeline is re-run with a different seed, neighborhood size, graph type, or resolution. Seed-stable, parameter-sensitive — the signature of cuts through a gradient.

Across three backbones

transcribed from FINAL_ANALYSIS.md §3

Against the official typology

Swiss Landschaftstypologie · chance-aware

Provenance

how the result earned trust

Where does this look belong?

The trained continuous field regresses any photograph to its most likely locations — the mixture modes of p(location | image). Several modes are the honest answer on a continuum: a look that exists in three valleys yields three candidates. Median modal error ~20 km on the leak-free spatial split — a coarse landscape field, not a geolocator.

predicted location — mixture modes

upload an image to predict its location

What is the visual structure of the Swiss landscape?

The question these documents answer, each at a different depth. Seven million street-level views, three independent vision models, one finding: a single connected visual continuum. Written for the general reader — start with the Summary, or read the full account.