Diversity across 439 photographs

What the data support

Feßbach has the most even detected family composition and the highest mean within-image DINO diversity. Heidesheim is dominated by Poaceae detections; Ribbeck by Asteraceae. Rettmer has seven detected families and the largest unresolved fraction. Equal-effort comparison at 74 images retains the order Feßbach, Ribbeck, Heidesheim, Rettmer.

Structural values are continuous feature-variance proxies in DINO space, with equal weight per photo. They are not counts of plant species or measured canopy layers.

1. Per-image diversity and rankings

The primary structural ranking uses exact within-photo Rao Q from all eligible native DINO patches. The primary taxonomic ranking uses effective detected families, q1 = exp(Shannon entropy), weighted by accepted family-region areas and conditioned on family-resolved area. A mixed-rank named-label score and the earlier structural composite remain available as sensitivity views.

Ranks are assigned across the full cohort. Equal family diversity receives a tied rank; unresolved images have no family-diversity rank. High scores can be driven by visible ground, senescence, viewpoint or unresolved regions, so the photo viewer shows assignment coverage alongside each score.

Highest within-image DINO diversity

#1 · plot_141Feßbach · 2025-08-20Q = 0.4587#2 · plot_08Feßbach · 2025-07-06Q = 0.4587#3 · plot_101Feßbach · 2025-07-06Q = 0.4557#4 · plot_131Feßbach · 2025-08-06Q = 0.4536#5 · plot_059Feßbach · 2025-06-21Q = 0.4529

Highest effective detected-family diversity

#1 · plot_346Ribbeck · 2025-06-17q1 = 2.951; 3 families#2 · plot_196Heidesheim · 2025-06-25q1 = 2.850; 3 families#3 · plot_405Ribbeck · 2025-08-20q1 = 2.680; 3 families#4 · plot_319Rettmer · 2025-07-29q1 = 2.628; 3 families#5 · plot_36Rettmer · 2025-06-17q1 = 2.364; 3 families

Click a photo to inspect its image, family hypotheses, configuration map, coordinates, original source path, metadata and duplicate aliases.

2. Cumulative family coverage

Curves count distinct detected families as image effort increases. The line is the exact expected richness over all possible subsets of each size. Shading spans the central 95% of 1,000 randomized image orders, describing order sensitivity within this observed collection. Every distinct photo counts as effort, including family-unresolved photos.

Only Rettmer includes spring (May); all other images are summer (June–August). Consequently, a four-site spring-versus-summer comparison cannot be estimated from this collection. Individual campaigns and month groupings preserve the finer sampling structure.

3. Site diversity, composition and configuration

¹ Family-resolved photos only, conditional family-area proportions, equal photo weights. γ1 is exp(entropy of pooled proportions); α1 is exp(mean within-photo entropy); β1 = γ1/α1. α0 is mean detected richness. ² Pairwise means between family-resolved photos from the same campaign; these describe spatial image heterogeneity at sampling time, including different views of a plot. Repeated plots and unknown physical footprints limit independence. See Jost (2007) for multiplicative diversity partitioning.

Turnover accounts for most detected compositional dissimilarity at every site; nestedness is smaller. Pooling campaigns also mixes temporal changes with spatial differences, which is why the same-campaign comparison is shown separately. The Sørensen turnover/nestedness decomposition follows Baselga (2010).

Configuration within the photographed footprint

³ Gower distance averages absolute differences in range-scaled same-family neighbor joins, component density and resolved grid fraction. Masks include only unambiguous, accepted single-family regions; ambiguous regions are unresolved. They are sampled on an aspect-preserving grid with longest side 256. Same-family joins are about 99% at all sites, partly reflecting the coarse, smooth region masks. Components and their densities depend on support coverage and grid resolution. These are exploratory image configuration descriptors, not mapped field patches.

Metadata join to 313 photos at Feßbach, Heidesheim and Rettmer. Ribbeck plot keys are inferred from filenames and marked accordingly. There are 359 between-campaign image comparisons sharing a recorded plot key; multiple orientations can generate several comparisons per plot. Those comparisons are exported for descriptive follow-up. Geographic coordinates and calibrated plot footprints were not recovered, so distance-decay and true geographic beta diversity are not estimated here.

4. Community NMDS and photo exploration

Family incidence uses Jaccard dissimilarity; conditional family-area composition uses Bray–Curtis. Both ordinations include 342 family-resolved photos. Identical profiles share coordinates: there are only 49 distinct incidence profiles and 75 distinct area profiles. Many points coincide because a photo often has one detected family. Thus an apparently separated group may reflect that family rather than a distinct whole community.

Click a point to inspect every coincident photo. The 3D view is an orthographic projection of the 3D fit; rotation changes the view, not the embedding.

Static NMDS figures and optimization diagnostics

The implementation uses weighted strong-tie nonmetric SMACOF: equal dissimilarities receive equal fitted disparities. Exact duplicate profiles are collapsed and weighted by photo multiplicities. Twenty-four starts are used in 2D and twelve in 3D; Stress1 is normalized residual distance error. This custom implementation is not a claim that vegan was executed. Strong-tie handling and the absence of step-across differ from some ecological defaults; see vegan’s NMDS guidance for the implications of high dissimilarities and tied distances. Start-wise diagnostics and coordinates are exported.

5. Shared DINO PCA and structural α, β, γ

One PCA is fitted to all 439 photo means in the same 768-dimensional DINOv3 coordinate system. Every eligible full-frame patch is unit-normalized before averaging; the resulting photo means retain their lengths. PCA is centered, without whitening. This is a between-photo PCA; within-photo variation is retained separately in Rao Q.

PC1 explains 21.9% and PC2 15.7% of between-photo variance (37.6% together). Inspection of representative photos suggests PC1 runs from dry vegetation or exposed ground toward green canopies. PC2 also reflects apparent canopy density, ground visibility and camera framing. These are visual interpretations of examples, not independently validated structural traits. Viewpoint, illumination, plant condition and taxonomic composition can all contribute.

An exact continuous-feature diversity partition

Here μᵢ is a photo’s mean patch feature and μ̄ the equal-photo group mean. Q equals the expected cosine dissimilarity of two independently sampled eligible patches, including self-pairs. It is the quadratic diversity of these visual features; its exact variance identity avoids choosing cluster counts. The calculation uses all 768 dimensions, rather than the first two plotted PCs. The general dissimilarity-based framework is described by Rao (1982).

Across all sites, α = 0.3730, β = 0.1306 and γ = 0.5035. The hierarchical partition assigns 74.1% of total feature diversity to variation within photos, 23.3% to variation between photos within sites, and 2.6% to differences between site centroids. Among between-photo variation alone, site centroids account for 10.2%. Most visual diversity therefore occurs within sites and photographs. This descriptive partition is not a significance test or an effect of management.

Evidence, limitations and reproducibility

Source identity is SHA-256 of original photo bytes. The five duplicated archive entries occur at Rettmer and are counted once; their aliases are retained. The source archive prefix is 2025/Annotation Indicator/Plot view/all_farms_plot/. All input IDs, 439 hashes, campaigns and metadata matches are in the dataset inventory. The source results and selection exports match the audited server versions.

Taxonomic support 0.7 combines the existing crop-quality/context support and the stored classifier adoption rules; it is not a calibrated 70% probability of correctness. Family, genus and species hypotheses are collapsed to the classifier’s recorded family lineage. Across the cohort this includes 263 direct family hypotheses, 155 genus hypotheses and 84 species hypotheses. For example, Rumex contributes to Polygonaceae and Malva moschata to Malvaceae. Promoted genus/species hypotheses account for 39.1% of the aggregated family-assigned pixel area. The aggregation was independently verified in all 439 photos. Order/class-only identifications cannot establish a family. Accepted mixed hypotheses share their source region area according to the earlier export. Relative area is an image-region proxy, not specimen abundance. The catalogue is preserved without a new botanical harmonization; labels such as Hydrophyllaceae and Boraginaceae remain distinct.

Raw DINOv3 ViT-B/16 patch features use facebook/dinov3-vitb16-pretrain-lvd1689m, pinned revision 5931719e67bbdb9737e363e781fb0c67687896bc. The working full frames were limited to 1,280 pixels on their longest side. Artifact-excluded and padded patches are omitted; there is no vegetation-only foreground mask. Some background therefore legitimately contributes to the reported visual diversity. Original thumbnails shown here still display annotation boards or plot frames that may be excluded from feature calculations.

The reproduction script reads the pinned cached run and metadata_join.json on the existing Linux analysis server and writes a separate report directory. It needs NumPy, SciPy, scikit-learn, Matplotlib and Pillow. Current full-cohort outputs do not train classifiers or regenerate taxonomic evidence. No ecological p-values are claimed because site, campaign, repeated plots, viewpoints and model-detection uncertainty are not independent controlled replicates.