abcd demo

[Track 3] 60% of South Asia's medicinal plants have never been assessed for extinction risk. RedRoot uses open data and validated ML to show assessors where to look first.

Comment

Inspiration

Almost every Indian household uses herbal medicine: giloy for fever, ashwagandha as a tonic, jatamansi in oils. Most of these plants are harvested from the wild. Some of the best-known ones, such as jatamansi, kuth and red sanders, are already threatened.

That made us ask how many medicinal plants nobody has checked at all. The answer surprised us. 60% of the 1,319 plants with recorded medicinal uses in South Asia have no usable IUCN Red List assessment. An unassessed plant is invisible to conservation policy, trade regulation and funding. Assessments take scarce expert time, so the people doing them need to know where to look first.

What it does

RedRoot is an extinction-risk radar for South Asia's medicinal plants. It produces:

  • A priority list of 103 unassessed species that our model screens as likely threatened (90 native to the subcontinent), each with an uncertainty range. These are candidates for an expert Red List assessment.
  • A harvest watchlist of 51 species whose range looks safe but whose profile matches the heavily harvested plants that range-based screening misses. These are candidates for trade monitoring.
  • A live website with a hotspot map, filterable lists you can download as CSV, and a species explorer for all 1,319 plants. Each species shows its IUCN status, model score, range and occurrence map.

Along the way we found two things:

  1. Synonyms hide threatened plants. GBIF attaches IUCN categories to the name IUCN used. When that name has become a synonym, the plant shows as "Not Evaluated". Kuth is Critically Endangered, but under its current name it looks unassessed. We recovered 45 hidden assessments, 6 of them threatened.
  2. Range-based screening has a blind spot. It catches narrow-range endemics but misses widespread, heavily harvested plants: jatamansi, kuth, sandalwood, guggul. The threatened plants it misses have a median of 15.5 recorded uses, against 4 for those it catches (p = 0.03).

How we built it

  • Medicinal plants: every plant with a South Asian record in Dr. Duke's Ethnobotanical Database (USDA, CC0), matched strictly against the GBIF Backbone Taxonomy.
  • IUCN status: taken from GBIF's mirror of the Red List, then reconciled against synonyms and the accepted names in Kew's World Checklist of Vascular Plants.
  • Occurrence data: GBIF records gridded at about 2.4 km through the Maps API vector tiles. Only records inside each species' native range (from Kew's checklist and TDWG regions) are kept. For Madagascar periwinkle, 99% of records lie outside its native range.
  • Features: IUCN criterion-B range metrics (extent of occurrence, area of occupancy), sampling effort and recency (to tell rare from rarely recorded), life form and climate. Medicinal use is deliberately not a model input, so risk is estimated independently of how often a plant is used.
  • Model: gradient-boosted trees, trained on 1,619 IUCN-assessed plants native to the Indian subcontinent. Cross-validation is grouped by plant family, so the model is always scored on families it never saw, and repeated over 5 seeds. A 25-model bootstrap ensemble gives each prediction an uncertainty range.
  • Website: a static site built with Leaflet and D3, hosted on GitHub Pages. It has no backend, and every number comes from the pipeline's JSON output.

Result: ROC-AUC is 0.912 ± 0.004, against 0.884 for the standard criterion-B rule. At the same 83% sensitivity, specificity rises from 0.27 to 0.86. The standard rule flags 866 of 1,187 non-threatened species as threatened; ours flags 108.

Challenges we ran into

  • Scale: paging through millions of individual GBIF records was far too slow, and we kept hitting rate limits. We switched to GBIF's map tiles and count queries and built a cached, resumable downloader that backs off when rate-limited.
  • Taxonomy almost fooled us: our first rule for recovering statuses from synonyms made ashwagandha Critically Endangered. It had borrowed the status of Withania chevalieri, a Cape Verde endemic that GBIF treats as a synonym. We rebuilt the rule to transfer status only from homotypic synonyms, which are the same species under a different genus. We rejected 22 wrong transfers and logged each one.
  • Too few threatened medicinal plants to train on: only 43 have usable labels. We trained on the wider regional flora and report medicinal-subset performance separately.
  • Accepting a negative result: we hoped use counts would predict threat. They don't: among range-safe species the AUC is 0.55, with a 95% confidence interval of 0.42–0.67. Rather than force use counts into the model, we kept harvest risk as a separate monitoring list and documented the limitation.

Accomplishments that we're proud of

  • A model that beats the standard method on plant families it never saw, with honest uncertainty ranges.
  • Recovering 45 hidden IUCN assessments, and catching the ashwagandha error before it shipped.
  • Turning a blind spot in the model into a finding, and a second list, instead of hiding it.
  • A pipeline anyone can rerun from public data with eight scripts, plus a live website.

What we learned

  • How a plant's name is recorded can decide whether it gets protected. Data cleaning was a scientific result in itself.
  • Validation design matters more than model choice. Grouping by family gives far more honest numbers than a random split.
  • A tool that is honest about what it can't see is more useful to decision-makers than one with an inflated score.

What's next

  • Add CITES trade data and market prices so the watchlist can detect decline directly.
  • Field-check the top priority species with regional botanists and the Botanical Survey of India.
  • Add climate-change exposure for Himalayan species.
  • Extend the pipeline to other regions; only the country list changes.
  • Co-design the watchlist with harvesting communities, under the Nagoya Protocol and India's Biological Diversity Act. Conservation priorities must not become a reason to take away communities' sustainable use rights.

Built With

Share this project:

Updates