02StreetView - Street-Level Imagery Lab

User manual, version 0.1.1 (2026-11-16). QGIS 3.28 to 4.x. GPL-2.0-or-later. Yusuf Eminoğlu (Geophilo).

1. What the plugin does

02StreetView carries a street-level imagery study from a street network to mapped, aggregated indicators inside QGIS. It samples points along the streets, matches open images from Panoramax and Mapillary or your own geotagged photos to them, keeps the licence and attribution of every image, segments the images locally, computes the green view index, sky view factor, tree and building view factors, sky share, visual enclosure and element shares with solid-angle weighting for 360° panoramas, aggregates them to street segments and areas, checks photographic SVF against a DSM, and writes a self-contained HTML report in which every formula is typeset with KaTeX.

The chain is available in the lab dock (page by page, with an image viewer, charts and styled layers) and in the Processing provider (one algorithm per step plus the whole chain). Only qgis.core, GDAL, NumPy and PyQt are required; SciPy is used for speed when present (same results without it) and onnxruntime is optional, for the scene-parsing model.

Why it exists: inside QGIS the street-level imagery tools stop at viewing (the Panoramax plugin, MapillaryClickPreview, mapillary2QGIS and several panorama viewers); the one analysis plugin, Green View Index, needs a paid Google Street View key and computes a single colour-rule GVI; the research packages that run the full chain (ZenSVI, streetscape, Treepedia) work outside QGIS and need PyTorch. Related Geophilo plugins compute SVF rasters from a DSM (PlanX, 02Geo3D) or a geometric visible-greenness index vgvi (02GreenRule); here the DSM only checks the photographs, and gvi is photographic.

2. Install

  1. QGIS: Plugins > Manage and Install Plugins, search 02StreetView, install. Or Install from ZIP with a release zip.
  2. A toolbar button and Plugins > 02StreetView open the lab; the Processing toolbox lists the provider 02StreetView.
  3. Optional, for the ONNX segmenter: the Segment page shows whether onnxruntime (1.16 or later, CPU build) is installed and, if not, the exact pip command for the Python of the running QGIS with a Copy command button. Run it in the OSGeo4W Shell (Windows) or a terminal and restart QGIS. The plugin never runs pip itself.
  4. The model file (238 434 293 bytes, UperNet + ConvNeXt-T, ADE20K, MIT weights) is not in the zip. Tick the licence notice on the Segment page and press Download model; the file goes to the plugin data folder and is checked against a pinned size and SHA-256.

Run time of the ONNX model on a CPU with 4 threads: about 1.47 s per 512 × 512 tile, three tiles per 1024 × 512 panorama plus resampling, so roughly 7 to 9 s per panorama. The Segment page shows an estimate before the run, and a run is capped at 300 images unless you raise the cap. The colour rules take well under 0.2 s per panorama.

3. Units, angles and working CRS

Sampling interval, matching radius, segment lengths and aggregation are computed in a projected CRS in metres. If the street layer is geographic, not in metres, or Web Mercator, every input is reprojected to the UTM zone of the layer's extent centre (EPSG:326xx north, 327xx south); otherwise its CRS is kept. Image services are queried in WGS84 and image positions are transformed into the working CRS before any distance is computed. The Network page badge, the run log and the report say which CRS was used and why.

Angles: azimuth aa in degrees clockwise from true north; elevation ee in degrees above the horizon; zenith angle θ=90∘−e\theta = 90^\circ - e. The centre column of a panorama looks along its heading (Panoramax view:azimuth, Mapillary computed_compass_angle), so the left edge is at a0=heading−180∘a_0 = \text{heading} - 180^\circ.

4. The lab dock, page by page

Network

Pick the street line layer, an optional area polygon layer (with a name field) and an optional DSM. Set the sampling interval (default 20 m), the offset of the first point (0 m) and the minimum gap (5 m: a point closer than this to an earlier point, as at junctions, is dropped). Sample points runs the sampling; Load demo street adds the synthetic canyons of section 13; Run the whole chain runs every step with the settings of every page. KPI cards: network length, segments, samples.

Sources

Source cards with switches. Panoramax: endpoint (default the federated API https://api.panoramax.xyz/api), no account needed. Mapillary: paste your client token (from the Mapillary developer page) and press Save token; it is stored in the QGIS authentication database as an API-header configuration and never in project files or logs. Local folder: JPEGs with EXIF GPS tags. Then the date range, the capture months (for example 5-9 to keep leaf-on pictures in the northern hemisphere, 11-3 in the southern; winter pictures of deciduous trees lower GVI and TVF; empty keeps every month), matching radius (15 m), panorama preference and heading tolerance (45°). Search and match runs as a background task; KPI cards show images found, matched samples, coverage, the share of panoramas and the median capture year, with a bar per source. A layer of the matched images, styled by capture year, is added.

Images

Download images fetches the matched images into the local cache (progress, cancel); the cache size and Clear cache are shown. The viewer shows the image of a sample (step with the arrows, or press the pick tool and click a sample on the map), the segmentation overlay in the group colours, a fisheye view of panoramas (equidistant projection of the upper hemisphere, the view behind SVF) and, under the picture, the attribution line and a link to the source page.

Segment

Chips choose the segmenter: Colour, ONNX ADE20K or Custom ONNX. The runtime card shows the onnxruntime status and install command; the model card shows size, licence and source, and Download model is enabled once the licence notice is ticked. Threads (default min⁡(4,cores−2)\min(4, \text{cores} - 2)) and the run-time estimate are shown. Run segmentation runs as a background task; label files are reused when the same segmenter revision labelled the image before. KPI cards: images segmented, seconds per image.

Indicators

A checklist of the indicators of section 8 with label, formula, unit and reference as tooltips; items the chosen segmenter cannot produce are disabled with the reason. The elevation band (default [−45∘,45∘][-45^\circ, 45^\circ], a pedestrian eye-level band) applies to the share indicators on panoramas. Compute indicators fills the sample fields; KPI gauges show mean GVI, SVF and enclosure, and the sample layer is styled by the chosen indicator (quantile classes).

Aggregate

Targets (street segments, areas), statistics (mean, median, n), minimum coverage (0.5) and street-length weighting for areas. Aggregate writes the layers and fills a ranked table with small bars. Compare with DSM computes svf_dsm at every sample and draws a scatter of photo SVF against DSM SVF with rr, MAE and bias.

Report

Writes the HTML report, CSV files and a GeoPackage with every layer (attribution and licence fields included; the attribution lines are also written to the layer metadata) into a chosen folder; a link opens the report.

Help

A short guide, the method references, the data licences, the privacy note and links to this manual and the repository.

Every long step runs as a QGIS background task with a progress bar and a Cancel button; the dock stays responsive. Results go to a 02StreetView layer group. A cancelled run keeps the steps finished before the cancel.

5. Image sources, licences and privacy

Panoramax is the federated open street-level imagery commons. The plugin searches the federated STAC API anonymously over the sample extent grown by the matching radius, with at most four requests per second and back-off on HTTP 429 and 5xx. For every picture it reads the licence (properties.license, for example CC-BY-SA-4.0), the producer, heading, capture time, field of view (360 for panoramas), position accuracy and the 2048 px sd image asset. The same capture can be listed by two instances; duplicates are removed by producer, original file name and capture time, and by the same capture time within 1 m.

Mapillary needs a client token: every user registers their own application on the Mapillary developer page. The token is sent as Authorization: OAuth <token>; it never appears in logs, error messages or cache sidecars. Searches are split into cells smaller than 0.01 degree² (an API rule); the API returns at most 2000 images per query without paging, so a cell whose answer is full is split in four and asked again (up to six times), and the log names any cell that is still full. Image URLs are signed and expire, so images are downloaded in the same session as the search. Mapillary content is under CC BY-SA 4.0 and the Mapillary terms, which ask apps to supplement the service rather than redistribute the images: the plugin keeps images in a local cache and has no image export.

Local folder: JPEGs with EXIF GPS latitude and longitude (read through GDAL); the heading from EXIF when present; 2:1 images are treated as equirectangular panoramas.

Privacy. Panoramax and Mapillary blur faces and number plates. The plugin does not attempt any de-blurring or person identification, and the segmentation groups (person, vehicle) are used only as shares of the view.

6. Image geometry and weights

An equirectangular panorama of H×WH \times W pixels maps row vv and column uu to elevation and azimuth:

Pixel elevation and azimuth of an equirectangular panorama:

ev=90∘−180∘ v+0.5H,au=a0+360∘(u+0.5W−0.5)(1)e_v = 90^\circ - 180^\circ\,\frac{v + 0.5}{H}, \qquad a_u = a_0 + 360^\circ \left(\frac{u + 0.5}{W} - 0.5\right) \tag{1}

Exact solid angle of a pixel:

Ωv=2πW(sin⁡evtop−sin⁡evbot),∑v,uΩv=4π(2)\Omega_{v} = \frac{2\pi}{W}\left(\sin e_v^{\mathrm{top}} - \sin e_v^{\mathrm{bot}}\right), \qquad \sum_{v,u} \Omega_{v} = 4\pi \tag{2}

Exact cosine-weighted view-factor share of a pixel:

wv=1π 2πW sin⁡2θvbot−sin⁡2θvtop2,θ=90∘−e,∑v,uwv=1(3)w_{v} = \frac{1}{\pi}\,\frac{2\pi}{W}\,\frac{\sin^2\theta_v^{\mathrm{bot}} - \sin^2\theta_v^{\mathrm{top}}}{2}, \qquad \theta = 90^\circ - e, \qquad \sum_{v,u} w_v = 1 \tag{3}

Both weights are exact integrals over each pixel row (not centre-point approximations), so they sum to 4π4\pi and 1 to rounding error. Shares weighted by Ωv\Omega_v are equal-area: a raw pixel count over-weights the rows near the zenith and nadir. The view-factor weight wvw_v is the cosine (Lambertian) weight of a horizontal surface at camera height. The mapping holds only for a full sphere, which is a 2:1 picture: a picture tagged 360° that is not 2:1 (within one pixel or 1 %) is a cropped or partial panorama, and all its indices are NULL with a note in the log rather than computed on invented elevations.

Two closed forms are used as checks in the tests and the demo:

Sky view factor at the centre of an infinite canyon:

SVF=cos⁡ ⁣(arctan⁡2HW)(6)\mathrm{SVF} = \cos\!\left(\arctan \frac{2H}{W}\right) \tag{6}

Sky view factor under a uniform horizon at elevation β\beta:

SVF=cos⁡2β(7)\mathrm{SVF} = \cos^2 \beta \tag{7}

The fisheye view renders the upper hemisphere in equidistant projection, r=R θ/(π/2)r = R\,\theta / (\pi/2); svf_steyn applies the Steyn annulus formula to it as an independent cross-check of svf.

7. Segmenters

Every segmenter labels each pixel with one of twelve groups: sky, vegetation, building, road, sidewalk, vehicle, person, wall/fence, pole/sign, terrain, water, other. Label files are written next to the cached image as <id>.<segmenter>.png with a JSON note of the segmenter revision and the groups it can produce. An indicator needing a group the segmenter does not produce is NULL, never 0.

Excess-green vegetation rule:

ExG=2g−r−b,g=GR+G+B,veg  ⟺  ExG>τ  ∧  G>R  ∧  G>B(4)\mathrm{ExG} = 2g - r - b, \quad g = \frac{G}{R + G + B}, \qquad \mathrm{veg} \iff \mathrm{ExG} > \tau \;\wedge\; G > R \;\wedge\; G > B \tag{4}

Colour sky rule:

sky  ⟺  R+G+B3⋅255≥βmin⁡  ∧  B≥R  ∧  e>0  ∧  4-connected to the top edge(5)\mathrm{sky} \iff \tfrac{R + G + B}{3 \cdot 255} \ge \beta_{\min} \;\wedge\; B \ge R \;\wedge\; e > 0 \;\wedge\; \text{4-connected to the top edge} \tag{5}

Colour rules (built in, NumPy). Vegetation: excess green on the chromatic coordinates above τ=0.05\tau = 0.05 with G>RG > R and G>BG > B (Woebbecke et al. 1995; the spectral rule of Li et al. 2015). Sky: brightness at least βmin⁡=0.55\beta_{\min} = 0.55, B≥RB \ge R, above the horizon, and 4-connected to the top edge (so a white wall separated from the sky by a dark roof line is not sky). Produces sky, vegetation and other. Known weaknesses: white or blue façades touching the sky, overcast sky, dark foliage. The colour rules are the offline baseline, not the reference method.

ONNX ADE20K: UperNet (Xiao et al. 2018) with a ConvNeXt-T backbone (Liu et al. 2022) trained on ADE20K (Zhou et al. 2017), 150 classes mapped to the groups (for example building, house and skyscraper to building; tree, grass and plant to vegetation). Panoramas are resized to 1024 × 512 (an option up to 4096 px wide), split into 512 × 512 tiles with 64 px overlap, normalised with the ImageNet mean and standard deviation; logits are averaged in the overlaps, upsampled bilinearly and the arg-max class is taken. Perspective images are resized so the short side is 512. CPU session with min⁡(4,cores−2)\min(4, \text{cores} - 2) threads by default. The weights are MIT; ADE20K images are under the dataset's research terms.

Custom ONNX: any model with an NCHW float input and per-class logits, plus a JSON class map {"input_size": 512, "mean": [...], "std": [...], "labels": {"0": "sky", ...}, "groups": {...}}. You are responsible for the model's licence.

8. Indicators

Share indicators on panoramas use the solid-angle weight Ωp\Omega_p of eq. (2) within the elevation band B=[elo,ehi]B = [e_{\mathrm{lo}}, e_{\mathrm{hi}}]; on perspective images they are pixel shares. View factors use the weight of eq. (3) over the upper hemisphere and need a 360° panorama. [ ⋅ ][\,\cdot\,] is 1 when the condition holds and 0 otherwise; gpg_p is the group of pixel pp.

FieldIndicatorUnitNeeds groupsReference
gviGreen view indexratiovegetationLi et al. 2015
gvi_pixelGreen view index (pixel share)ratiovegetationYang et al. 2009; Zhang & Dong 2018
svfSky view factorratiosky panoramas onlyJohnson & Watson 1984; Oke 1987; Gong et al. 2018
tvfTree view factorratiovegetation panoramas onlyGong et al. 2018
bvfBuilding view factorratiobuilding panoramas onlyGong et al. 2018
ovfOther view factorratiobuilding panoramas onlyGong et al. 2018
sky_shareSky share (openness)ratioskyGong et al. 2018
enclosureVisual enclosureratiobuilding, wall_fence, vegetationMa et al. 2021; Ewing & Handy 2009
bld_shareBuilding shareratiobuildingZhang et al. 2018
road_shareRoad shareratioroadZhang et al. 2018
walk_shareSidewalk shareratiosidewalkZhang et al. 2018
veh_shareVehicle shareratiovehicleZhang et al. 2018
svf_steynSky view factor (Steyn annuli)ratiosky panoramas only, off by defaultSteyn 1980
gvi_liGreen view index (Li et al. 18 views)ratiovegetation panoramas only, off by defaultLi et al. 2015

Green view index gvi

Panorama: solid-angle share of vegetation within the elevation band; perspective: vegetation pixels / all pixels.

GVI=∑p∈B[ gp=veg ] Ωp∑p∈BΩp,Ωp=Δφ (sin⁡eptop−sin⁡epbot)(11)\mathrm{GVI} = \frac{\sum_{p \in B} [\,g_p = \mathrm{veg}\,]\, \Omega_p}{\sum_{p \in B} \Omega_p}, \qquad \Omega_p = \Delta\varphi\,(\sin e_{p}^{\mathrm{top}} - \sin e_{p}^{\mathrm{bot}}) \tag{11}

Green view index (pixel share) gvi_pixel

Vegetation pixels / pixels whose centre lies in the elevation band (unweighted).

GVIpx=Nveg(B)N(B)(12)\mathrm{GVI}_{\mathrm{px}} = \frac{N_{\mathrm{veg}}(B)}{N(B)} \tag{12}

Sky view factor svf

Cosine-weighted share of the upper hemisphere, with S(θ,φ)=1S(\theta, \varphi) = 1 for sky.

SVF=1π∫02π ⁣ ⁣∫0π/2S(θ,φ) cos⁡θsin⁡θ  dθ dφ(13)\mathrm{SVF} = \frac{1}{\pi} \int_{0}^{2\pi}\!\!\int_{0}^{\pi/2} S(\theta,\varphi)\, \cos\theta \sin\theta \;d\theta\, d\varphi \tag{13}

Tree view factor tvf

The SVF integral with S=1S = 1 for vegetation.

TVF=1π∬V(θ,φ) cos⁡θsin⁡θ  dθ dφ(14)\mathrm{TVF} = \frac{1}{\pi} \iint V(\theta,\varphi)\, \cos\theta \sin\theta \; d\theta\, d\varphi \tag{14}

Building view factor bvf

The SVF integral with S=1S = 1 for buildings.

BVF=1π∬β(θ,φ) cos⁡θsin⁡θ  dθ dφ(15)\mathrm{BVF} = \frac{1}{\pi} \iint \beta(\theta,\varphi)\, \cos\theta \sin\theta \; d\theta\, d\varphi \tag{15}

Other view factor ovf

The SVF integral with S=1S = 1 for every group other than sky, vegetation and building (SVF+TVF+BVF+OVF=1\mathrm{SVF} + \mathrm{TVF} + \mathrm{BVF} + \mathrm{OVF} = 1).

OVF=1π∬O(θ,φ) cos⁡θsin⁡θ  dθ dφ,SVF+TVF+BVF+OVF=1(16)\mathrm{OVF} = \frac{1}{\pi} \iint O(\theta,\varphi)\, \cos\theta \sin\theta \; d\theta\, d\varphi, \qquad \mathrm{SVF} + \mathrm{TVF} + \mathrm{BVF} + \mathrm{OVF} = 1 \tag{16}

Sky share (openness) sky_share

Panorama: solid-angle share of sky in the full sphere; perspective: sky pixels / all pixels.

ssky=14π∑p[ gp=sky ] Ωp(17)s_{\mathrm{sky}} = \frac{1}{4\pi} \sum_{p} [\,g_p = \mathrm{sky}\,]\, \Omega_p \tag{17}

Visual enclosure enclosure

Solid-angle share of building, wall/fence and vegetation within elevations −10∘-10^\circ to +45∘+45^\circ.

E=∑p∈BE[ gp∈{bld,wall,veg} ] Ωp∑p∈BEΩp,BE=[−10∘,45∘](18)E = \frac{\sum_{p \in B_E} [\,g_p \in \{\mathrm{bld}, \mathrm{wall}, \mathrm{veg}\}\,]\, \Omega_p}{\sum_{p \in B_E} \Omega_p}, \qquad B_E = [-10^\circ, 45^\circ] \tag{18}

Building share bld_share

Solid-angle share of buildings within the elevation band.

sg=∑p∈B[ gp=g ] Ωp∑p∈BΩp(19)s_{g} = \frac{\sum_{p \in B} [\,g_p = g\,]\, \Omega_p}{\sum_{p \in B} \Omega_p} \tag{19}

Road share road_share

Solid-angle share of road within the elevation band.

sg=∑p∈B[ gp=g ] Ωp∑p∈BΩp(20)s_{g} = \frac{\sum_{p \in B} [\,g_p = g\,]\, \Omega_p}{\sum_{p \in B} \Omega_p} \tag{20}

Sidewalk share walk_share

Solid-angle share of sidewalk within the elevation band.

sg=∑p∈B[ gp=g ] Ωp∑p∈BΩp(21)s_{g} = \frac{\sum_{p \in B} [\,g_p = g\,]\, \Omega_p}{\sum_{p \in B} \Omega_p} \tag{21}

Vehicle share veh_share

Solid-angle share of vehicles within the elevation band.

sg=∑p∈B[ gp=g ] Ωp∑p∈BΩp(22)s_{g} = \frac{\sum_{p \in B} [\,g_p = g\,]\, \Omega_p}{\sum_{p \in B} \Omega_p} \tag{22}

Sky view factor (Steyn annuli) svf_steyn

Annulus sum on an equidistant fisheye with n=36n = 36 rings, pip_i sky pixels and tit_i all pixels of ring ii.

SVFSteyn=π2n∑i=1nsin⁡ ⁣(π(2i−1)2n)piti(23)\mathrm{SVF}_{\mathrm{Steyn}} = \frac{\pi}{2n} \sum_{i=1}^{n} \sin\!\left(\frac{\pi (2i-1)}{2n}\right) \frac{p_i}{t_i} \tag{23}

Green view index (Li et al. 18 views) gvi_li

Mean vegetation pixel share of 18 perspective views cut from the panorama: headings 0∘,60∘,…,300∘0^\circ, 60^\circ, \dots, 300^\circ from the panorama heading, pitches −45∘,0∘,+45∘-45^\circ, 0^\circ, +45^\circ, 60∘60^\circ field of view (the Google Street View view set of Li et al. 2015), for comparison with studies that use it.

GVILi=118∑i=16∑j=13Nveg(ψi,ϑj)N(ψi,ϑj),ψi=60∘(i−1),ϑj∈{−45∘,0∘,45∘},FOV=60∘(24)\mathrm{GVI}_{\mathrm{Li}} = \frac{1}{18} \sum_{i=1}^{6} \sum_{j=1}^{3} \frac{N_{\mathrm{veg}}(\psi_i, \vartheta_j)}{N(\psi_i, \vartheta_j)}, \qquad \psi_i = 60^\circ (i - 1), \quad \vartheta_j \in \{-45^\circ, 0^\circ, 45^\circ\}, \quad \mathrm{FOV} = 60^\circ \tag{24}

The ratios are dimensionless in [0,1][0, 1]. With the four view factors computed from their own groups, SVF+TVF+BVF+OVF=1\mathrm{SVF} + \mathrm{TVF} + \mathrm{BVF} + \mathrm{OVF} = 1 holds as a check, not by construction. The Steyn formula uses n=36n = 36 rings, pip_i sky pixels and tit_i all pixels of ring ii on the equidistant fisheye.

9. Matching, aggregation and coverage

Samples are placed every interval metres along each line from offset, both ends included, with the bearing of the local line direction. For each sample the candidate images are those within the radius after de-duplication and the age and accuracy filters. With prefer panorama the nearest panorama wins over a nearer perspective image; otherwise the nearest image. A perspective image must look along the street: ∣a−b∣≤δ\lvert a - b \rvert \le \delta or ∣a−b−180∘∣≤δ\lvert a - b - 180^\circ \rvert \le \delta for image azimuth aa, street bearing bb and tolerance δ\delta. Each image is used by one sample only, the nearest. Results do not depend on the input order.

Aggregation joins samples to their street through seg_id (an id field or the line feature id) and to areas by location (a sample on a shared boundary counts once, for the polygon with the lowest feature id). Statistics leave NULL values out. Coverage:

c=nimagesnsamples(10)c = \frac{n_{\mathrm{images}}}{n_{\mathrm{samples}}} \tag{10}

and low_coverage = 1 when cc is below the minimum coverage (0.5). With length weighting each sample in an area is weighted by Ls/nsL_s / n_s, its street length over the number of samples on that street.

10. DSM check

With a DSM (metres), the horizon elevation γi\gamma_i is searched along nn azimuths (default 36) up to the radius (100 m) from the DSM height plus the eye height z0z_0 (1.6 m), following Zakšek et al. (2011):

SVFDSM=1−1n∑i=1nsin⁡2γi,γi=max⁡ ⁣(0,max⁡darctan⁡zd−z0d)(8)\mathrm{SVF}_{\mathrm{DSM}} = 1 - \frac{1}{n} \sum_{i=1}^{n} \sin^2 \gamma_i, \qquad \gamma_i = \max\!\left(0, \max_{d} \arctan \frac{z_d - z_0}{d}\right) \tag{8}

This cosine-weighted form equals cos⁡2β\cos^2\beta for a uniform horizon at β\beta, the same quantity as the photographic svf. The agreement is reported as

MAE=1n∑k∣sk−tk∣,bias=1n∑k(sk−tk)(9)\mathrm{MAE} = \frac{1}{n}\sum_{k} \lvert s_k - t_k \rvert, \qquad \mathrm{bias} = \frac{1}{n}\sum_{k} (s_k - t_k) \tag{9}

with sks_k the photo SVF and tkt_k the DSM SVF, plus Pearson's rr. The DSM only checks the photographs at the same points; it does not produce an SVF raster (PlanX and 02Geo3D do that). Distances and the radius are metres: a DSM in degrees (EPSG:4326), in feet or in Web Mercator is first resampled (nearest neighbour) onto a metric grid in the working CRS around the samples, at about its own ground resolution; heights are read as metres. The DSM code is the vendored skyview SDK (MIT).

11. Processing algorithms

Provider zero2streetview. The steps exchange two layers: the sample points (match and indicator fields) and the image points, which carry every field of an image record (source, id, capture time, projection, field of view, size, download URL, licence, attribution, page link, sequence, accuracy) and, after the later steps, image_path and label_path. A model built from the step algorithms therefore gives the same values as the whole-chain algorithm (the test suite checks this to 10−910^{-9}). Inputs are copied; metric steps run in the working CRS; outputs keep real field types (doubles for indicators, NULL where a value is undefined) and are not styled. Parameters marked advanced are under Advanced parameters. From Python: processing.run("zero2streetview:<id>", {...}). Help in the Processing dialog gives the formulas in plain text; this manual typesets them.

Create demo street Demo

Id: zero2streetview:demo_street

Writes the synthetic test site of 02StreetView: four straight 200 m streets in EPSG:32635, each the centre line of an infinite street canyon of width W=20W = 20 m with aspect ratios H/W=0.2,0.5,0.9H/W = 0.2, 0.5, 0.9 and 1.51.5, two area polygons (West, East) and a DSM of the same canyons (1 m cells, roofs at H+1.6H + 1.6 m because the panoramas are taken 1.6 m above the street). The panoramas are rendered from this geometry and served by the Demo image source of the other algorithms, so the whole chain runs offline.

The street layer carries the planted truth in closed form: svf_true, SVF=cos⁡(arctan⁡(2H/W))\mathrm{SVF} = \cos(\arctan(2H/W)) (Sky view factor at the centre of an infinite canyon) and gvi_true, GVI=(sin⁡b−sin⁡a)/(sin⁡ehi−sin⁡elo)\mathrm{GVI} = (\sin b - \sin a)/(\sin e_{\mathrm{hi}} - \sin e_{\mathrm{lo}}), the solid-angle share of the planted vegetation band [a,b][a, b] within the elevation band [elo,ehi]=[−45∘,45∘][e_{\mathrm{lo}}, e_{\mathrm{hi}}] = [-45^\circ, 45^\circ].

ParameterDescriptionTypeDefaultNotes
NETWORKDemo streets Feature sink (output)-
AREASDemo areas Feature sink (output)-
DSMDemo DSM Raster (output)-
STREETSNumber of streetsOutput number-Returned in the result dictionary.

Sample street network Sampling

Id: zero2streetview:sample_network

Places sample points every interval metres along each street line, from offset, both ends included, with the local street bearing (degrees clockwise from north). A point closer than the minimum gap to an earlier point (junction overlap) is dropped. Geographic or Web Mercator input is reprojected to the UTM zone of the layer centre and the output is written in that CRS.

Output fields: sample_id, seg_id (the id field value, or the feature id of the line), dist_along (m), bearing (degrees) and seg_len (m, the line length, used for length weighting). Give an id field when the samples are aggregated back to the lines later.

ParameterDescriptionTypeDefaultNotes
NETWORKStreet lines Vector features-
ID_FIELDStreet id field (empty: feature id) optionalField-
INTERVALSampling interval (m) Number20.0min 0.5
OFFSETOffset of the first point (m) Number0.0min 0
MIN_GAPMinimum gap between points (m) Number5.0min 0
OUTPUTSample points Feature sink (output)-
SAMPLESNumber of samplesOutput number-Returned in the result dictionary.

Search and match images Imagery

Id: zero2streetview:search_images

Queries the chosen sources over the sample extent (grown by the radius) and matches at most one image to each sample: images within the radius, duplicates across Panoramax instances removed (same producer, original file and capture time, or the same capture within 1 m), images older than the age limit or with a worse position accuracy dropped. With Nearest panorama a 360 degree panorama is preferred over a perspective image; a perspective image must look along the street (∣a−b∣≤δ\lvert a - b \rvert \le \delta or ∣a−b−180∘∣≤δ\lvert a - b - 180^\circ \rvert \le \delta for image azimuth aa, street bearing bb and heading tolerance δ\delta). An image is used by one sample only (the nearest).

Sources. Panoramax is the federated open street-level imagery commons (no account; the licence of every picture is read from the service). Mapillary needs your own free client token, stored as a QGIS authentication configuration of the API header type holding Authorization: OAuth <token>; Mapillary download URLs are signed and expire after some hours, so download soon after the search. A local folder is searched for JPEGs with EXIF GPS tags. Demo serves the synthetic panoramas that belong to Create demo street and never uses the network.

Outputs: the samples with image_id, source, dist_m, captured, projection and licence (NULL when unmatched), and the image points layer.

The image points layer holds one point per matched image with every field of the image record (source, id, capture time, projection, field of view, size, download URL, licence, attribution, page link, sequence, accuracy) and, after the later steps, image_path and label_path. Keep the attribution with any map or figure.

ParameterDescriptionTypeDefaultNotes
SAMPLESSample points (from 'Sample street network') Vector features-
SOURCESImage sources Enumeration00: Panoramax (open, no account); 1: Mapillary (client token); 2: Local folder of geotagged JPEGs; 3: Demo (synthetic panoramas, offline)
ENDPOINTPanoramax API endpoint advancedStringhttps://api.panoramax.xyz/api
AUTHCFGMapillary token (authentication configuration) optionalAuthentication configuration-
FOLDERLocal folder of geotagged JPEGs optionalFile or folder-
SINCECaptured on or after (ISO date, optional) optionalString-
UNTILCaptured on or before (ISO date, optional) optionalString-
RADIUSMatching radius (m) Number15.0min 0.1
MAX_AGEMaximum image age (years, 0 = no limit) Number8.0min 0
PREFERPreference within the radius Enumeration00: Nearest panorama; 1: Nearest image (any)
HEADING_TOLHeading tolerance for perspective images (degrees) advancedNumber45.0min 0, max 90
MIN_ACCURACYDrop images with a position accuracy worse than (m, 0 = keep all) advancedNumber10.0min 0
MAX_IMAGESImage cap per run Number2000min 1
REFERENCE_DATEReference date for the age limit (ISO date, empty = today) optional advancedString-
MONTHSCapture months kept (e.g. 5-9 for leaf-on in the north, 11-3 in the south; empty = all) optional advancedString-
OUTPUTMatched samples Feature sink (output)-
IMAGESImage points Feature sink (output)-
FOUNDImages found in the search areaOutput number-Returned in the result dictionary.
MATCHEDMatched samplesOutput number-Returned in the result dictionary.
COVERAGECoverage (matched / samples)Output number-Returned in the result dictionary.

Download images to the cache Imagery

Id: zero2streetview:download_images

Downloads the image of every point of the image layer into the local cache (<cache>/<source>/<id[:2]>/<id>.jpg, with a JSON sidecar holding the licence and attribution; a file already in the cache is reused). Local-folder images are used where they are. Requests go through the QGIS network settings (proxy, SSL) with at most four requests per second and back-off on HTTP 429 and 5xx. A failed download is reported and skipped. The cache stays on this computer; there is no image export.

Output: the image layer with image_path (NULL when the download failed).

The image points layer holds one point per matched image with every field of the image record (source, id, capture time, projection, field of view, size, download URL, licence, attribution, page link, sequence, accuracy) and, after the later steps, image_path and label_path. Keep the attribution with any map or figure.

ParameterDescriptionTypeDefaultNotes
IMAGESImage points (from 'Search and match images') Vector features-
CACHE_DIRCache folder (empty: the plugin data folder) optionalFile or folder-
AUTHCFGMapillary token (authentication configuration) optionalAuthentication configuration-
OUTPUTDownloaded images Feature sink (output)-
DOWNLOADEDImages in the cacheOutput number-Returned in the result dictionary.
FAILEDFailed downloadsOutput number-Returned in the result dictionary.

Segment images Segmentation

Id: zero2streetview:segment_images

Labels every downloaded image with one group per pixel (sky, vegetation, building, road, sidewalk, vehicle, person, wall/fence, pole/sign, terrain, water, other) and writes the label file next to the image as <id>.<segmenter>.png with a JSON note of the segmenter revision and the groups it can produce; a label file of the same revision is reused.

Colour rules (built in, numpy): vegetation where the excess green ExG=2g−r−b\mathrm{ExG} = 2g - r - b of the chromatic coordinates exceeds the threshold and G>RG \gt R, G>BG \gt B (eq. 4); sky where the brightness is at least the minimum, B≥RB \ge R, above the horizon and 4-connected to the top edge. Produces sky, vegetation and other only. ONNX ADE20K: UperNet + ConvNeXt-T scene parsing (150 classes, MIT weights, downloaded from the dock's Segment page and checked by SHA-256), panoramas resized to the panorama width and tiled in 512 px tiles with 64 px overlap, classes mapped to the groups; needs the optional onnxruntime package. Custom ONNX: any model with NCHW float input and per-class logits plus a JSON class map (input_size, mean, std, labels, groups).

Output: the image layer with label_path.

ParameterDescriptionTypeDefaultNotes
IMAGESDownloaded images (from 'Download images to the cache') Vector features-
SEGMENTERSegmenter Enumeration00: Colour rules (built in); 1: ONNX UperNet ConvNeXt-T ADE20K; 2: Custom ONNX model + JSON class map
MODELCustom ONNX model (custom segmenter) optional advancedFile or folder-
CLASS_MAPJSON class map of the custom model optional advancedFile or folder-
THREADSONNX threads (0 = automatic) advancedNumber0min 0, max 64
PANO_WIDTHPanorama width for the ONNX model (px) advancedNumber1024min 512, max 4096
ONNX_CAPONNX image cap advancedNumber300min 1
EXG_THRESHOLDColour rules: excess green threshold advancedNumber0.05min -1, max 2
SKY_BRIGHTNESSColour rules: minimum sky brightness (0-1) advancedNumber0.55min 0, max 1
OUTPUTSegmented images Feature sink (output)-
SEGMENTEDImages with a label fileOutput number-Returned in the result dictionary.
SECONDS_PER_IMAGESeconds per newly segmented imageOutput number-Returned in the result dictionary.

Compute view indicators Indicators

Id: zero2streetview:compute_indicators

Adds one field per indicator to the matched samples, computed from the label file of the sample's image. On a 360 degree panorama the pixel at row vv has elevation ev=90∘−180∘(v+0.5)/He_v = 90^\circ - 180^\circ (v + 0.5)/H (eq. 1); shares are solid-angle weighted with the exact pixel solid angle Ωv\Omega_v (eq. 2), so the poles are not over-weighted, and view factors use the cosine weight wvw_v (eq. 3) with zenith angle θ=90∘−e\theta = 90^\circ - e (∑Ωv=4π\sum \Omega_v = 4\pi, ∑wv=1\sum w_v = 1). Share indices use the elevation band; perspective images use pixel shares. An indicator whose groups the segmenter does not produce, or that needs a panorama, is NULL (never 0). Fields label_path and the band are also recorded.

ParameterDescriptionTypeDefaultNotes
SAMPLESMatched samples (from 'Search and match images') Vector features-
IMAGESSegmented images (from 'Segment images') Vector features-
INDICESIndicators (empty: the defaults the segmenter supports) optionalEnumeration-0: gvi - Green view index; 1: gvi_pixel - Green view index (pixel share); 2: svf - Sky view factor; 3: tvf - Tree view factor; 4: bvf - Building view factor; 5: ovf - Other view factor; 6: sky_share - Sky share (openness); 7: enclosure - Visual enclosure; 8: bld_share - Building share; 9: road_share - Road share; 10: walk_share - Sidewalk share; 11: veh_share - Vehicle share; 12: svf_steyn - Sky view factor (Steyn annuli); 13: gvi_li - Green view index (Li et al. 18 views)
BAND_LOWElevation band for shares, lower bound (degrees) Number-45.0min -90, max 90
BAND_HIGHElevation band for shares, upper bound (degrees) Number45.0min -90, max 90
OUTPUTSamples with indicators Feature sink (output)-
WITH_VALUESSamples with indicator valuesOutput number-Returned in the result dictionary.

Aggregate to street segments Aggregation

Id: zero2streetview:aggregate_segments

Joins the samples to the street lines they were sampled from through seg_id (the id field given to 'Sample street network', or the line feature id when none was given; use the same street layer in the same order) and writes a copy of the lines in the samples' CRS with the statistics.

Statistics of the chosen sample fields (empty: every indicator field found): mean, median and n (number of non-NULL values); optional sd (sample standard deviation) and ci95 (half-width of the 95 % confidence interval of the mean, Student t); written as <field>_<stat>, plus n_samples, n_images (samples with an image), coverage c=nimages/nsamplesc = n_{\mathrm{images}}/n_{\mathrm{samples}} and low_coverage (1 when cc is below the minimum coverage, NULL without samples). NULL values are left out of every statistic.

ParameterDescriptionTypeDefaultNotes
SAMPLESSamples with indicators Vector features-
NETWORKStreet lines Vector features-
ID_FIELDStreet id field used for sampling (empty: feature id) optionalField-
FIELDSFields to aggregate (empty: every indicator) optionalField-
STATSStatistics Enumeration0, 1, 20: mean; 1: median; 2: n; 3: sd; 4: ci95
MIN_COVERAGEMinimum coverage (below: low_coverage = 1) Number0.5min 0, max 1
OUTPUTStreet segments Feature sink (output)-

Aggregate to areas Aggregation

Id: zero2streetview:aggregate_areas

Statistics of the samples inside each polygon (a sample on a boundary shared by two polygons counts once, for the polygon with the lowest feature id). With weight by street length each sample is weighted by the length of its street divided by the number of samples on that street, so short and long streets count in proportion to their length.

Statistics of the chosen sample fields (empty: every indicator field found): mean, median and n (number of non-NULL values); optional sd (sample standard deviation) and ci95 (half-width of the 95 % confidence interval of the mean, Student t); written as <field>_<stat>, plus n_samples, n_images (samples with an image), coverage c=nimages/nsamplesc = n_{\mathrm{images}}/n_{\mathrm{samples}} and low_coverage (1 when cc is below the minimum coverage, NULL without samples). NULL values are left out of every statistic.

ParameterDescriptionTypeDefaultNotes
SAMPLESSamples with indicators Vector features-
AREASArea polygons Vector features-
FIELDSFields to aggregate (empty: every indicator) optionalField-
STATSStatistics Enumeration0, 1, 20: mean; 1: median; 2: n; 3: sd; 4: ci95
WEIGHT_BY_LENGTHWeight samples by street length BooleanFalse
MIN_COVERAGEMinimum coverage (below: low_coverage = 1) Number0.5min 0, max 1
OUTPUTAreas Feature sink (output)-

DSM check of photo sky view factor DSM check

Id: zero2streetview:dsm_svf_check

Computes the sky view factor from a DSM at every sample and compares it with the photographic svf. At each sample the horizon elevation γi\gamma_i is searched along nn azimuths up to the radius, from the DSM height plus the eye height: γi=max⁡(0,max⁡darctan⁡((zd−z0)/d))\gamma_i = \max(0, \max_d \arctan((z_d - z_0)/d)), and SVFDSM=1−1n∑isin⁡2γi\mathrm{SVF}_{\mathrm{DSM}} = 1 - \frac{1}{n}\sum_i \sin^2 \gamma_i (eq. 8) (Sky view factor from DSM horizon angles; cosine-weighted, equal to cos⁡2β\cos^2\beta for a uniform horizon at β\beta). Output field svf_dsm; outputs nn, Pearson's rr, MAE and bias (eq. 9, with sks_k the svf and tkt_k the svf_dsm of sample kk). The DSM only checks the photographs at the same points; it does not produce an SVF raster.

ParameterDescriptionTypeDefaultNotes
SAMPLESSamples with indicators Vector features-
DSMDigital surface model (metres) Raster layer-
RADIUSHorizon search radius (m) Number100.0min 1
DIRECTIONSNumber of azimuths Number36min 4, max 720
EYE_HEIGHTEye height above the DSM (m) Number1.6min 0
OUTPUTSamples with svf_dsm Feature sink (output)-
NPairs comparedOutput number-Returned in the result dictionary.
PEARSON_RPearson rOutput number-Returned in the result dictionary.
MAEMean absolute errorOutput number-Returned in the result dictionary.
BIASBias (photo - DSM)Output number-Returned in the result dictionary.

Street-level imagery lab (whole chain) Lab

Id: zero2streetview:street_imagery_lab

The whole chain in one run, as in the dock: sample the street network, search and match images, download them into the cache, segment them, compute the indicators, aggregate to street segments (and to areas when given), check photo SVF against a DSM when given, and write the HTML report (formulas typeset with KaTeX, offline, with every attribution line, licence and reference). Each step behaves as the algorithm of the same name. A cancelled run keeps the finished steps.

Sources. Panoramax is the federated open street-level imagery commons (no account; the licence of every picture is read from the service). Mapillary needs your own free client token, stored as a QGIS authentication configuration of the API header type holding Authorization: OAuth <token>; Mapillary download URLs are signed and expire after some hours, so download soon after the search. A local folder is searched for JPEGs with EXIF GPS tags. Demo serves the synthetic panoramas that belong to Create demo street and never uses the network.

The image points layer holds one point per matched image with every field of the image record (source, id, capture time, projection, field of view, size, download URL, licence, attribution, page link, sequence, accuracy) and, after the later steps, image_path and label_path. Keep the attribution with any map or figure.

Statistics of the chosen sample fields (empty: every indicator field found): mean, median and n (number of non-NULL values); optional sd (sample standard deviation) and ci95 (half-width of the 95 % confidence interval of the mean, Student t); written as <field>_<stat>, plus n_samples, n_images (samples with an image), coverage c=nimages/nsamplesc = n_{\mathrm{images}}/n_{\mathrm{samples}} and low_coverage (1 when cc is below the minimum coverage, NULL without samples). NULL values are left out of every statistic.

ParameterDescriptionTypeDefaultNotes
NETWORKStreet lines Vector features-
ID_FIELDStreet id field (empty: feature id) optionalField-
AREASArea polygons (optional) optionalVector features-
DSMDSM for the SVF check (optional) optionalRaster layer-
INTERVALSampling interval (m) Number20.0min 0.5
OFFSETOffset of the first point (m) advancedNumber0.0min 0
MIN_GAPMinimum gap between points (m) advancedNumber5.0min 0
SOURCESImage sources Enumeration00: Panoramax (open, no account); 1: Mapillary (client token); 2: Local folder of geotagged JPEGs; 3: Demo (synthetic panoramas, offline)
ENDPOINTPanoramax API endpoint advancedStringhttps://api.panoramax.xyz/api
AUTHCFGMapillary token (authentication configuration) optionalAuthentication configuration-
FOLDERLocal folder of geotagged JPEGs optionalFile or folder-
SINCECaptured on or after (ISO date, optional) optionalString-
UNTILCaptured on or before (ISO date, optional) optionalString-
RADIUSMatching radius (m) Number15.0min 0.1
MAX_AGEMaximum image age (years, 0 = no limit) Number8.0min 0
PREFERPreference within the radius Enumeration00: Nearest panorama; 1: Nearest image (any)
HEADING_TOLHeading tolerance for perspective images (degrees) advancedNumber45.0min 0, max 90
MIN_ACCURACYDrop images with a position accuracy worse than (m, 0 = keep all) advancedNumber10.0min 0
MAX_IMAGESImage cap per run Number2000min 1
REFERENCE_DATEReference date for the age limit (ISO date, empty = today) optional advancedString-
MONTHSCapture months kept (e.g. 5-9 for leaf-on in the north, 11-3 in the south; empty = all) optional advancedString-
CACHE_DIRCache folder (empty: the plugin data folder) optional advancedFile or folder-
SEGMENTERSegmenter Enumeration00: Colour rules (built in); 1: ONNX UperNet ConvNeXt-T ADE20K; 2: Custom ONNX model + JSON class map
MODELCustom ONNX model (custom segmenter) optional advancedFile or folder-
CLASS_MAPJSON class map of the custom model optional advancedFile or folder-
THREADSONNX threads (0 = automatic) advancedNumber0min 0, max 64
PANO_WIDTHPanorama width for the ONNX model (px) advancedNumber1024min 512, max 4096
ONNX_CAPONNX image cap advancedNumber300min 1
INDICESIndicators (empty: the defaults the segmenter supports) optionalEnumeration-0: gvi - Green view index; 1: gvi_pixel - Green view index (pixel share); 2: svf - Sky view factor; 3: tvf - Tree view factor; 4: bvf - Building view factor; 5: ovf - Other view factor; 6: sky_share - Sky share (openness); 7: enclosure - Visual enclosure; 8: bld_share - Building share; 9: road_share - Road share; 10: walk_share - Sidewalk share; 11: veh_share - Vehicle share; 12: svf_steyn - Sky view factor (Steyn annuli); 13: gvi_li - Green view index (Li et al. 18 views)
BAND_LOWElevation band for shares, lower bound (degrees) Number-45.0min -90, max 90
BAND_HIGHElevation band for shares, upper bound (degrees) Number45.0min -90, max 90
STATSStatistics advancedEnumeration0, 1, 20: mean; 1: median; 2: n; 3: sd; 4: ci95
MIN_COVERAGEMinimum coverage (below: low_coverage = 1) advancedNumber0.5min 0, max 1
WEIGHT_BY_LENGTHWeight area statistics by street length BooleanFalse
DSM_RADIUSDSM horizon search radius (m) advancedNumber100.0min 1
DSM_DIRECTIONSDSM azimuths advancedNumber36min 4, max 720
EYE_HEIGHTEye height above the DSM (m) advancedNumber1.6min 0
SAMPLESSamples Feature sink (output)-
IMAGESImages Feature sink (output)-
SEGMENTSStreet segments Feature sink (output)-
AREAS_OUTAreas optionalFeature sink (output)-
REPORTHTML report optionalFile (output)-
N_SAMPLESSamplesOutput number-Returned in the result dictionary.
MATCHEDMatched samplesOutput number-Returned in the result dictionary.
COVERAGECoverageOutput number-Returned in the result dictionary.
DSM_RPearson r of photo SVF and DSM SVFOutput number-Returned in the result dictionary.

12. Output layers and fields

LayerGeometryFields
SamplesPointsample_id, seg_id, dist_along (m), bearing (°), seg_len (m); match fields image_id, source, dist_m, captured, projection, licence (NULL when unmatched); image_path, label_path; one double per indicator; svf_dsm after the DSM check
ImagesPointimage_id, source, captured, year, projection, azimuth, licence, attribution, page_url, sequence, image_path (the Processing step layer also carries every other image record field)
Street segmentsMultiLineStringthe street fields, <field>_mean, <field>_median, <field>_n, n_samples, n_images, coverage, low_coverage
AreasPolygonthe area fields and the same statistics and coverage fields

The HTML report holds the KPI cards, the indicator means with their formulas, the method equations (1) to (10), the street table, the DSM check, the image sources with every attribution line and licence, the parameters, timings, log and references. It starts with <!DOCTYPE html> and embeds KaTeX and its fonts, so it renders offline.

13. Worked example: the demo street

The demo has four straight streets 200 m long in EPSG:32635, each the centre line of an infinite street canyon of width W=20W = 20 m with aspect ratio H/W=0.2,0.5,0.9,1.5H/W = 0.2, 0.5, 0.9, 1.5, and a 1024 × 512 panorama rendered every 20 m (positions jittered by up to 2 m, random headings). In each panorama the wall top at azimuth φ\varphi from the street axis lies at elevation arctan⁡ ⁣(H∣sin⁡φ∣/(W/2))\arctan\!\left(H \lvert \sin \varphi \rvert / (W/2)\right); above it is sky, below the horizon is road, and a band of vegetation is planted between elevations aa and bb below the horizon. The truth is written in closed form: SVF=cos⁡(arctan⁡(2H/W))\mathrm{SVF} = \cos(\arctan(2H/W)) (eq. 6) and GVI=(sin⁡b−sin⁡a)/(sin⁡ehi−sin⁡elo)\mathrm{GVI} = (\sin b - \sin a)/(\sin e_{\mathrm{hi}} - \sin e_{\mathrm{lo}}) for the band [elo,ehi]=[−45∘,45∘][e_{\mathrm{lo}}, e_{\mathrm{hi}}] = [-45^\circ, 45^\circ]. A DSM of the same canyons (buildings 40 m deep, 1 m cells) is written for the DSM check; the panoramas are taken 1.6 m above the street, so the DSM roofs stand at H+1.6H + 1.6 m and the DSM eye (1.6 m) sees the same horizon.

Steps

  1. Dock: Network > Load demo street, then Run the whole chain. Or Processing: Create demo street, then Street-level imagery lab (whole chain) with the image source Demo and the demo DSM.
  2. Open the Images page to step through the panoramas with the segmentation overlay and the fisheye view.
  3. Read the indicators on the Indicators page and the street means on the Aggregate page; run the DSM check there.
  4. Write the report on the Report page.

Result (colour segmenter, measured when this manual was built)

44 samples, 44 matched, coverage 100 %. Largest per-sample error against the truth: SVF 0.00011, GVI 0.00170. DSM check: n=44n = 44, r=0.9999r = 0.9999, MAE 0.0089, bias -0.0089 (photo minus DSM). The small remaining difference comes from the 1 m DSM cells (the horizon is read at cell centres, up to half a cell behind the wall face) and the finite DSM canyons (buildings 40 m deep, horizon searched to 100 m).

StreetH/WSVF truthmean svfmean svf_dsmGVI truthmean gvisamples
Canyon A0.20.92850.92840.93320.23080.231411
Canyon B0.50.70710.70710.71950.39290.394611
Canyon C0.90.48560.48570.49700.05880.057911
Canyon D1.50.31620.31620.32320.47890.477711

The colour segmenter produces sky, vegetation and other only, so bvf, ovf, enclosure and the element shares are left out (or NULL when requested); with the ONNX model they are computed as well.

14. Limitations

15. References

  1. Ewing, R., Handy, S. (2009). Measuring the unmeasurable: urban design qualities related to walkability. Journal of Urban Design 14(1), 65-84. https://doi.org/10.1080/13574800802451155
  2. GDAL/OGR contributors (2025). GDAL/OGR Geospatial Data Abstraction software Library. Open Source Geospatial Foundation. https://gdal.org
  3. Gong, F.-Y., Zeng, Z.-C., Zhang, F., Li, X., Ng, E., Norford, L. K. (2018). Mapping sky, tree, and building view factors of street canyons in a high-density urban environment. Building and Environment 134, 155-167. https://doi.org/10.1016/j.buildenv.2018.02.042
  4. Johnson, G. T., Watson, I. D. (1984). The determination of view-factors in urban canyons. Journal of Climate and Applied Meteorology 23(2), 329-335.
  5. Li, X., Zhang, C., Li, W., Ricard, R., Meng, Q., Zhang, W. (2015). Assessing street-level urban greenery using Google Street View and a modified green view index. Urban Forestry & Urban Greening 14(3), 675-685. https://doi.org/10.1016/j.ufug.2015.06.006
  6. Liu, Z., Mao, H., Wu, C.-Y., Feichtenhofer, C., Darrell, T., Xie, S. (2022). A ConvNet for the 2020s. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 11976-11986.
  7. Ma, X., Ma, C., Wu, C., Xi, Y., Yang, R., Peng, N., Zhang, C., Ren, F. (2021). Measuring human perceptions of streetscapes to better inform urban renewal: a perspective of scene semantic parsing. Cities 110, 103086. https://doi.org/10.1016/j.cities.2020.103086
  8. Mapillary (2026). Mapillary API v4 and Terms of Use. https://www.mapillary.com/developer (contributed imagery under CC BY-SA 4.0).
  9. Oke, T. R. (1981). Canyon geometry and the nocturnal urban heat island: comparison of scale model and field observations. Journal of Climatology 1(3), 237-254.
  10. Oke, T. R. (1987). Boundary Layer Climates, 2nd edition. Routledge, London.
  11. Panoramax contributors (2026). Panoramax, the open and federated street-level imagery commons. https://panoramax.fr (pictures under the licence stated for each picture).
  12. QGIS Development Team (2026). QGIS Geographic Information System. Open Source Geospatial Foundation Project. https://qgis.org
  13. Steyn, D. G. (1980). The calculation of view factors from fisheye-lens photographs: research note. Atmosphere-Ocean 18(3), 254-258. https://doi.org/10.1080/07055900.1980.9649091
  14. Woebbecke, D. M., Meyer, G. E., Von Bargen, K., Mortensen, D. A. (1995). Color indices for weed identification under various soil, residue, and lighting conditions. Transactions of the ASAE 38(1), 259-269.
  15. Xiao, T., Liu, Y., Zhou, B., Jiang, Y., Sun, J. (2018). Unified perceptual parsing for scene understanding. In: Computer Vision - ECCV 2018, Lecture Notes in Computer Science 11209, 432-448. Springer.
  16. Yang, J., Zhao, L., McBride, J., Gong, P. (2009). Can you see green? Assessing the visibility of urban forests in cities. Landscape and Urban Planning 91(2), 97-104. https://doi.org/10.1016/j.landurbplan.2008.12.004
  17. Zakšek, K., Oštir, K., Kokalj, Ž. (2011). Sky-view factor as a relief visualization technique. Remote Sensing 3(2), 398-415. https://doi.org/10.3390/rs3020398
  18. Zhang, Y., Dong, R. (2018). Impacts of street-visible greenery on housing prices: evidence from a hedonic price model and a massive street view image dataset in Beijing. ISPRS International Journal of Geo-Information 7(3), 104. https://doi.org/10.3390/ijgi7030104
  19. Zhang, F., Zhou, B., Liu, L., Liu, Y., Fung, H. H., Lin, H., Ratti, C. (2018). Measuring human perceptions of a large-scale urban region using machine learning. Landscape and Urban Planning 180, 148-160. https://doi.org/10.1016/j.landurbplan.2018.08.020
  20. Zhou, B., Zhao, H., Puig, X., Fidler, S., Barriuso, A., Torralba, A. (2017). Scene parsing through ADE20K dataset. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 633-641.

16. Citation, licence, links

Eminoğlu, Y. (2026). 02StreetView - Street-Level Imagery Lab (version 0.1.1) [QGIS plugin]. https://gitlab.com/geophilo1/zero2streetview. https://doi.org/10.5281/zenodo.24263662. The DOI 10.5281/zenodo.24263662 always resolves to the latest version; version 0.1.1 has DOI https://doi.org/10.5281/zenodo.24263663.

Licence GPL-2.0-or-later. Vendored skyview modules: MIT. KaTeX: MIT.