End-to-end pipeline: Gemma 4 + spatial + spectral + key + GIS export • soilKey

This vignette walks the complete soilKey pipeline on a real Brazilian soil profile, hitting every public entry point in canonical order:

Spatial guide — soil_classes_at_location() returns ranked likely classes at the field GPS coordinate before any pedon data is collected.
Multimodal extraction — classify_from_documents() runs Gemma 4 (local Ollama) on a soil-description PDF and a profile-wall photograph, extracts horizons + Munsell + site metadata, and feeds everything into a PedonRecord.
Spectral analogy — classify_by_spectral_neighbours() consumes a Vis-NIR scan of the surface horizon, finds the K most similar OSSL profiles within a regional radius, and returns a probabilistic class prediction.
Deterministic classification — classify_wrb2022(), classify_sibcs(include_familia = TRUE), classify_usda() walk the canonical YAML rules and produce the final names with full key trace + provenance + evidence grade.
Reports — report() writes a self-contained HTML pedologist report.
GIS export — report_to_qgis() produces a multi-layer GeoPackage that QGIS opens natively.

The whole pipeline runs offline once the Ollama Gemma 4 model is pulled; the only network hit is the optional SoilGrids fetch in step 1.

1. Set the scene

We use a canonical Latossolo Vermelho Distrocoeso from the Mata Atlântica around Seropédica, RJ, parent material gneiss. The fixture mimics a real Embrapa survey profile.

# Field GPS coordinates of the planned profile pit.
field_lat <- -22.7
field_lon <- -43.7

2. Spatial guide – before any pedon data

soil_classes_at_location() queries SoilGrids 2.0 (or any WRB-coded raster the user provides) and returns a ranked list of likely classes plus the canonical attribute thresholds that distinguish them.

guide <- soil_classes_at_location(
  lat        = field_lat,
  lon        = field_lon,
  system     = "wrb2022",
  source_url = "https://files.isric.org/soilgrids/latest/data/wrb/MostProbable.vrt"
)

guide$distribution
#> # Ranked candidate classes:
#> # rsg_code  rsg_name      probability
#> # FR        Ferralsols    0.62
#> # AC        Acrisols      0.21
#> # NT        Nitisols      0.12
#> # CM        Cambisols     0.05
guide$typical_attributes
#> # Per-class diagnostic thresholds to confirm in the field.

The function does not classify – it tells the pedologist “you are most likely standing on a Ferralsol; here is what to look for to confirm”.

3. Multimodal extraction with local Gemma 4

The pedologist arrives at the pit, photographs the wall against a Munsell chart, scans the field sheet, and exports the survey report PDF. classify_from_documents() chains the entire downstream pipeline – VLM extraction, all three classifications, optional report rendering – in a single call.

The default provider is local Gemma 4 edge (gemma4:e4b, ~3 GB, multimodal text + image + audio) via Ollama – no API key, no data leaving the laptop. Pull the model once:

ollama pull gemma4:e4b
ollama serve

res <- classify_from_documents(
  pdf      = "perfil_042_descricao.pdf",
  image    = "perfil_042_parede.jpg",
  report   = "perfil_042.html",
  provider = "ollama"  # default; uses gemma4:e4b
)

res$classifications$wrb$name
#> [1] "Geric Ferric Rhodic Chromic Ferralsol (Clayic, Humic, Dystric, Ochric, Rubic)"
res$classifications$sibcs$name
#> [1] "Latossolos Vermelhos Distroficos tipicos, argilosa, moderado"
res$classifications$usda$name
#> [1] "Rhodic Hapludox"

Every extracted attribute is stamped source = "extracted_vlm" in the PedonRecord’s provenance log; the deterministic key is consumed by the PedonRecord unaware of how each value got there. The architectural invariant – the key is never delegated to a model – holds.

For the rest of the vignette we keep working with the populated pedon res$pedon.

# For a runnable demo without Ollama / a real PDF, reuse the
# canonical Ferralsol fixture -- the downstream code is the same.
pedon <- make_ferralsol_canonical()

4. Spectral analogy

If a Vis-NIR scan is available for the surface horizon, classify_by_spectral_neighbours() adds another evidence layer. It finds the K most spectrally similar OSSL profiles within a regional radius and returns a probabilistic class prediction.

# Hypothetical: a real OSSL South-America library with WRB labels
# obtained via `download_ossl_subset_with_labels()`.
ossl_lib <- download_ossl_subset_with_labels(
  region          = "south_america",
  max_distance_km = 10
)

# Pull the surface-horizon Vis-NIR scan from the populated pedon.
query_spectrum <- pedon$spectra$vnir[1, ]

spectral <- classify_by_spectral_neighbours(
  spectrum     = query_spectrum,
  ossl_library = ossl_lib,
  k            = 25,
  region       = list(lat = field_lat, lon = field_lon,
                      radius_km = 500)
)
spectral$distribution
#> # class    n_neighbours  probability
#> # FR              22       0.88
#> # AC               2       0.08
#> # NT               1       0.04
spectral$neighbours
#> # The 25 closest OSSL profiles + their distances + labels.

The biome-aware regional filter prevents the analogy from drifting to non-tropical reference soils.

5. Deterministic classification

The canonical step. classify_wrb2022() / classify_sibcs() / classify_usda() walk the canonical YAML rules over the populated PedonRecord.

cls_wrb   <- classify_wrb2022(pedon, on_missing = "silent")
cls_sibcs <- classify_sibcs(pedon, include_familia = TRUE)
cls_usda  <- classify_usda(pedon)

cls_wrb$name
#> [1] "Geric Ferric Rhodic Chromic Ferralsol (Clayic, Humic, Dystric, Ochric, Rubic)"
cls_sibcs$name
#> [1] "Latossolos Vermelhos Distroficos tipicos, argilosa, moderado"
cls_usda$name
#> [1] "Rhodic Hapludox"

# Each ClassificationResult carries the full key trace, the per-
# attribute provenance, and an evidence grade A/B/C/D.
cls_wrb$evidence_grade
#> [1] "A"
length(cls_wrb$trace)         # number of RSGs tested before assignment
#> [1] 16

6. HTML report

report() writes a self-contained HTML one-pager with the cross-system summary, full key trace, evidence grade, qualifiers, ambiguities, missing-data hints, the horizons table, and the per-source provenance summary.

results <- list(wrb = cls_wrb, sibcs = cls_sibcs, usda = cls_usda)
report(results, file = "perfil_042.html", pedon = pedon)

The output is a single HTML file with inline CSS – no external network requests, suitable for emailing to a colleague or attaching to a laudo.

7. GIS export

report_to_qgis() produces a multi-layer GeoPackage (.gpkg) that QGIS reads natively.

results <- list(wrb = cls_wrb, sibcs = cls_sibcs, usda = cls_usda)
report_to_qgis(
  pedon           = pedon,
  classifications = results,
  file            = "perfil_042.gpkg",
  report_html     = "perfil_042.html"
)

The GeoPackage carries three layers:

pedon_point – POINT geometry at the profile coordinates with all classification metadata as attributes (WRB / SiBCS / USDA names, RSG / Ordem / Order codes, evidence grades, principal qualifiers, supplementary qualifiers, hyperlink to the rendered HTML report).
horizons_table – one row per horizon, with the canonical horizon-schema attributes. Joined to pedon_point by site_id.
provenance_log – per-(horizon, attribute, source) provenance rows for downstream auditing.

In QGIS: Layer → Add Layer → Add Vector Layer → perfil_042.gpkg. The point appears on the canvas with all classification metadata in the feature pop-up; styling rules can map symbol colour to the evidence grade or the assigned RSG.

8. The complete picture

# Pipeline summary:
#
#   field GPS      ->  soil_classes_at_location()         "what to expect"
#                                  |
#                                  v
#   PDF + photo    ->  classify_from_documents() (Gemma 4)  populates PedonRecord
#                                  |
#                                  v
#   Vis-NIR scan   ->  classify_by_spectral_neighbours()    spectral prior
#                                  |
#                                  v
#                  ->  classify_wrb2022()  + classify_sibcs() + classify_usda()
#                                  |       (the deterministic step -- canonical)
#                                  v
#                  ->  report() / report_to_qgis()         deliverables

Each step’s output carries explicit provenance into the next; the final evidence_grade reflects the worst-source rule applied to the attributes that were decisive in the assigned name. Two pedologists running this pipeline on the same documents get the same output bit-for-bit.

Summary

soilKey separates four distinct stages:

Spatial guides (soil_classes_at_location) – expectations from a soil-class raster.
Extraction (classify_from_documents, extract_*) – VLM populates a PedonRecord, never classifies.
Spectral analogy (classify_by_spectral_neighbours) – OSSL nearest-neighbour analogy as a prior.
Deterministic classification (classify_wrb2022 / classify_sibcs / classify_usda) – the canonical step.

Plus two delivery formats: HTML reports (report) and GeoPackage exports (report_to_qgis). All four stages preserve provenance and evidence grading; the deterministic key remains the only thing that assigns a class.