kingmin/streetscape-transfer-audit

Cross-city streetscape transfer data and code

Evaluation data, spatial partitions, scientific code and executed results for cross-city transfer of tabular streetscape descriptors. The collection is organized as three archives, with a standalone technical protocol.

Files

The online viewer displays the archive file index, with one row per archive member. It does not display individual benchmark observations. Download and extract the archives below to access the scientific data and code.

File	Contents
reproducibility_package.zip	Processed primary and four additional spatial benchmarks, full-population partition manifests, model code, configurations and numerical results.
supplementary_tables.zip	Machine-readable task, input-ablation, calibration, spatial-sensitivity and external-case tables.
supplementary_external_validation_data.zip	Prepared Brazilian census-tract data, predictions, metrics, source records and income-analysis scripts.
PROTOCOL.md	Target definitions, information access, spatial evaluation, model settings, uncertainty and execution examples.
archive_file_descriptions.csv	Actual members of the three archives, with byte sizes, SHA-256 and descriptions.
dataset_metadata.json, source_provenance.json	Scientific dataset metadata and original-source attribution.
LICENSES.md, LICENSE	Component-specific source terms.
MANIFEST.json, SHA256SUMS	Payload sizes and SHA-256 checks.

The archives include numerical inputs and results. Original street-view images, all neural model weights and per-model neural prediction arrays are not included. The code recreates checkpoints and predictions when models are trained.

Data and statistical units

The underlying subset contains 1,321,985 image records from central sampling windows in 26 Global Streetscapes cities. The primary prepared benchmark has 129,998 records, seven source cities and 19 target cities. Its partition column identifies source training, source validation, target adaptation and target test. UUIDs, coordinates, capture sequences, geographic exclusions and exact scene mappings are recorded in the included manifests.

The primary targets are an exact-mapped five-class scene proxy and an image-derived green-view indicator. Vegetation and Terrain are excluded from primary green-view inputs. Perception dimensions are excluded from primary predictors and conditioning; a separate perception-composite diagnostic is not an independently observed settlement label. Missing targets are handled by task.

Spatial and same-sequence exclusions precede sampling caps. For the primary cross-city comparison, preprocessing and checkpoint selection use source labels only. Four further spatial settings use the same training rules and add 1,416 model fits and 2,280 city-level evaluation rows. Three seeds are averaged within each city, and the 19 cities remain the bootstrap clusters across configurations.

The independent IBGE case uses census tracts as observations. Its outcomes are 2022 area-level income variables, and its imagery coverage is confined to the Sao Paulo and Rio de Janeiro sampling windows. It is a separate target and model family from the image-derived neural benchmark.

Getting started

Verify SHA256SUMS and extract reproducibility_package.zip, preserving paths:

import pandas as pd
path = "project/study/analysis/experiments/prepared_v3/benchmark.parquet"
data = pd.read_parquet(path)
print(data.groupby(["city_std", "split"]).size())

Use PROTOCOL.md and the archive README for executable commands, scientific package versions and the distinction between prepared-data training and reconstructing inputs from upstream sources. These heterogeneous archives are intended for direct download and extraction. A generic load_dataset call is not the documented entry point.

Interpretation

The spatial comparisons support evaluating direct transfer alongside the tested adaptation procedures, while fixed-test experiments quantify target-label calibration. Report all configurations and paired intervals rather than selecting a setting by its target score. Individual cities and interval bounds vary.

The approximately 2 x 2 km central sampling windows are not whole-city samples. The visual targets are derived measurements and can share upstream model errors. A buffer change can alter training/adaptation samples even where test IDs remain fixed, so the comparisons measure protocol sensitivity rather than isolated causal buffer effects. City-bootstrap intervals are descriptive and unadjusted for multiple comparisons. Census and imagery dates differ; the two-city income case does not establish a global socioeconomic or settlement classifier.

Sources and terms

Global Streetscapes is provided by NUS Urban Analytics Lab: Hou et al. (2024), DOI 10.1016/j.isprsjprs.2024.06.023, upstream dataset. Those data and derivatives retain CC BY-SA 4.0. Official IBGE releases provide the independent census information; exact URLs and source descriptions are in source_provenance.json. The category vocabulary is attributed to the MIT CSAIL Places project. LICENSES.md records component-specific terms; no new blanket software license is granted.

Reuse

Cite the original Global Streetscapes and IBGE sources when reusing their data. Record the actual repository commit used for an analysis.