Classification challenges

Classification challenges — and how pedotri helps

Anyone who's tried to use soil-texture classifications across projects, countries, or time periods runs into the same recurring problems. They're not specific to one country or one decade — they show up wherever soil scientists have classified soils, which is everywhere. This page lays out those problems and points to the pedotri features that address each.

This isn't a "look how good pedotri is" pitch. Most of these problems aren't fully solvable by a library — they're rooted in how classifications were authored. But several of them are at least mitigable, and that's where pedotri tries to be useful.

1. The same name does not mean the same soil

"Clay loam" in USDA (clay 27-40 %, sand 20-45 %) is not the same material as "clay loam" in Avery (UK), which is not the same as a "clay loam" entry in older textbooks following the now-superseded International system.

Even within one system, the underlying particle-size cutoff drifts:

System	Sand boundary	Silt boundary	"Loam" meaning
USDA	0.05 mm	0.002 mm	a specific polygon
ISSS	0.02 mm	0.002 mm	a different polygon — looks similar on the triangle but the input numbers don't match
KA5	0.063 mm	0.002 mm	another polygon
Russian (Качинский, 1965)	—	particles <0.01 mm	a 1-D classification by "physical clay" — not a triangle at all

Consequence: a sample reported as "loam" in one document and used as "loam" in another may classify into a wholly different polygon once the analyst tries to plot it on the receiving system's triangle.

What pedotri does:

• Ships 15 classifications as built-ins, so you don't have to guess what someone else's "loam" means — you can re-classify the sample under the convention you actually want.
• Provides pedotri.psd.convert(..., source=, target=) to convert the (sand, silt, clay) triple between the four sand-cutoff conventions (USDA / FAO / ISSS / KA5). The conversion is a log-linear approximation, not exact, and we explain the assumptions on the Soil-conversions page so you know when to trust it.

import pedotri
import pedotri.psd as psd

# Sample described under ISSS — re-classify under USDA
sand_usda, silt_usda, clay = psd.convert(
    40.0, 40.0, 20.0, source="ISSS", target="USDA"
)
pedotri.classify(float(sand_usda[0]), clay, "USDA")

2. Different classifications use different axes

Most international systems classify on the sand-silt-clay triangle (a 2-D simplex). The Russian Качинский / Полевой определитель почв России (2008) uses a single axis: physical clay, the mass fraction of particles <0.01 mm. The two cannot be mapped 1-to-1 from a three-number lab report alone — the 0.002-0.01 mm range that Качинский folds into "physical clay" cuts across the USDA silt/clay boundary, so you'd need a full sieve PSD to translate losslessly.

This isn't a bug — it's a deliberate scientific choice. The Russian community decided that a single, water-retention-correlated cutoff is more useful for the soils they work with than the triangle.

What pedotri does:

• Treats axes as first-class metadata on every classification. pedotri.classify adapts its calling convention to the axes the classification declares:

  pedotri.classify(60, 20, "USDA")          # two axes: sand, clay
  pedotri.classify(35, "KACHINSKY")          # one axis: physical_clay
  pedotri.classify(35, "RUS2004")            # same axis, six-class official label

• Provides pedotri.psd.interpolate_psd(sieves, passing, target) so you can derive physical clay from a richer PSD when you have one — the right way to translate to Качинский, instead of pretending the triangle is enough.

• Documents on the Soil-conversions page exactly why the triangle ↔ physical-clay conversion is lossy and which measurements you'd need to do it properly.

3. Classifications drift over time

A "GEPPA" classification (France, 1995) is a refinement of "Jamagne" (1967), which was itself a refinement of older French soil-survey work. Across the Russian classifications, the 1977 "Классификация и диагностика почв СССР", the 1997 revision, and the 2004 "Классификация и диагностика почв России" each renamed taxa, merged or split classes, and changed diagnostic horizons. The 2004 work even introduces a wholly new family of taxa for anthropogenically transformed soils and technogenic surface formations (ТПО) that simply did not exist in 1977.

The Russian Soil Institute's infosoil.ru maintains explicit synonymy tables — one for 1977 ↔ 1997, one for 1997 ↔ 2004, and one for the Russian 1988 soil map ↔ WRB/FAO. The mere fact that those tables need to be authored at all is the problem: a "клееземы тундровые" of 1988 maps onto multiple distinct WRB / FAO entries depending on which sub-property you key on.

The French / Russian cases are not exotic. Australian Northcote has gone through similar refinements; the German KA5 expanded to its current 31 subclasses from a smaller predecessor; ISSS lost ground to USDA partly because it was harder to re-key against on new measurements.

What pedotri does:

• Ships both legacy and modern classifications side-by-side: GEPPA and JAMAGNE (France), KACHINSKY 9-class and RUS2004 6-class (Russia), ISSS and USDA (international). You can re-classify the same sample under each one and see what changes — which is often more informative than picking a single one and hoping it's the right answer.

• Treats class keys as stable identifiers separately from class names, so renaming a class (or translating it) never breaks downstream code. This is the same pattern that lets us add the 6-class Russian 2004 variety scheme alongside Качинский without shadowing or renaming anything.

• Doesn't pretend to ship synonymy tables. Those are domain knowledge that belongs in a citable reference (and we link to one for the Russian case in Soil-conversions). Pedotri's job is to let you classify in either system and compare — not to invent a cross-reference.

4. Classifications are not always texture classifications

Most national soil-survey schemes are profile-genetic: they classify soils by what genetic horizons formed, in what order, under what climate and parent material. Texture is one property used to diagnose a horizon — not the classification itself.

The Russian 2004 system is explicit about this: the variety ("разновидности") level — the only level on which texture appears — is outside the hierarchical taxonomic system. WRB does the same: texture appears as a qualifier (Arenic / Loamic / Clayic / Siltic), not as a primary axis.

Consequence: if you have a profile description and want to classify it under WRB or the Russian 2004 system end-to-end, you need a horizon analyzer, not a texture classifier.

What pedotri does not try to do:

• It does not classify by genetic horizons, parent material, water regime, or any non-textural property. It is a texture classifier, full stop. Profile-genetic classification (WRB, Soil Taxonomy, the Russian 2004 system at типа/подтипа level) is out of scope.

What pedotri does do:

• Provides texture classification as a clean, composable building block. If your pipeline starts with profile analysis and ends with "OK, what texture variety is this Ap horizon?", pedotri's classify is the right answer for the last step. It's not the right answer for "is this a Чернозём or a Кастанозём?" — that's a different (and much harder) classification problem.

• Returns enough metadata via detailed=True to be honest about how close a sample is to a class boundary:

  result = pedotri.classify(45, 15, "USDA", detailed=True)
  result.key       # 'loam'
  result.distance  # 2.3 — modest distance from the boundary

  A small distance value is your signal to not over-trust the label: the sample is borderline and a different system, or a small measurement error, would put it elsewhere.

5. There are always disturbed soils that nobody planned for

Every classification was designed against a population of "well- formed" soils. Urban soils, mine spoils, landfill caps, dredged sediments, ash deposits — these are real surface materials, classifiable in the trivial sense (they have a sand / silt / clay fraction you can measure), but in the rich sense they fall through every genetic-classification framework. The Russian community made this explicit: their 2004 scheme adds a wholly separate ствол ("trunk") for антропогенно-преобразованные почвы (anthropogenically transformed soils) and a parallel non-soil system of ТПО — techogenic surface formations (квазизёмы, артифабрикаты, литостраты, урбистраты, …) — that explicitly cannot be a genetic soil classification because they have no genetic horizons.

What pedotri does:

• Honors the input you give it. If you measure the texture of an urban litho­strat (urban dredge), pedotri will classify the texture fine — that's a meaningful, defensible measurement. It will not tell you "this is не почва" (not a soil) because that's not a textural judgment.

• Lets you ship a custom classification for your domain. If your workflow uses an urban-soils grouping like "compactible / friable / contaminated", you can author it as a TOML and load it through pedotri.register_classification. See Custom-classifications-in- practice.

6. Most classifications were never i18n'd

Russian soil literature uses Russian class names; French uses French; German uses German — and the receiving English-speaking pipeline often silently re-labels everything as USDA's English names, losing the connection to the source convention.

What pedotri does:

• Every built-in classification ships class names in six languages (en / fr / de / es / ru / pt), plus the source-region language as default (KACHINSKY defaults to ru, GEPPA to fr, EMBRAPA to pt). The class key ("sandy_clay_loam") stays stable across locales; only the display name changes.

  pedotri.classify(13, 50, "USDA")                     # 'clay' (key)
  pedotri.classify(13, 50, "USDA", locale="ru")        # 'глина'
  pedotri.classify(13, 50, "GEPPA", locale="ru")       # 'тяжёлая глина'

• Falls back through the locale chain (fr-FR → fr → en → class key) so a partial translation never silently drops to an empty string.

7. People don't have one classification — they have several

In practice, a single dataset gets classified under several systems over its lifetime: a French soil sample analyzed at INRA might be GEPPA-labeled in its original record, USDA-labeled when entered into an international database, FAO-labeled when summarized for a continental-scale model, and Качинский-labeled when discussed in collaboration with Russian colleagues.

Tools that handle one classification at a time impose a translation cost at every interface.

What pedotri does:

• Treats all classifications uniformly. The same classify() call works against every built-in and every user-registered TOML. Same DataFrame accessor (df.soil.classify("USDA"), df.soil.classify("GEPPA")), same vectorized performance, same error handling.

• Ships a units= keyword so the input convention (%, g/kg, g/g) can also drift between projects without forcing manual conversion at every boundary.

Where pedotri can't help

Some problems aren't ours to solve, and it's better to say so:

• No reference data conversion. If you have a sample classified as "глеезём арктический" in the Russian 1988 soil map legend, no amount of texture classification will give you the WRB equivalent — that's a synonymy / type-conversion problem in profile-genetic classification space. infosoil.ru/index.php?pageID=mapcorr publishes the table; pedotri does not duplicate it.
• No fitted PTFs outside Saxton-Rawls / Wösten. Russian PTFs (Vadyunina-Korchagina, Kachinsky bulk-density relations) are not shipped. If you need them, ship them via the same TOML extension mechanism a user can extend custom classifications with — or open an issue.
• No genetic / morphological classification. As above. WRB and Soil Taxonomy are out of scope.

Related

• Classifications-and-when-to-use-them — what each built-in is good for
• Soil-conversions — translating triples between cutoffs
• Custom-classifications-in-practice — when none of the built-ins fit