EthnoHarvest by Ecoroots

EthnoHarvest puts evidence before extraction, using AI and ecological science to reveal harvesting risks, guide responsible collection and trace medicinal plants from forest to buyer.

Comment

EthnoHarvest AI

Evidence before extraction: a harvest copilot for wild medicinal plants
Team name Eco Roots
Project name EthnoHarvest AI
Title Evidence before extraction: a harvest copilot for wild medicinal plants
Team members Shruthi and Vishnu
Hackathon EthnoHack 2026
Track Track 3, Biodiversity and Sustainability
Presentation https://drive.google.com/file/d/1ZmND90TMYBvuDxy3funerlU6loNMi8vz/view?usp=sharing
Youtube https://youtu.be/2VO2N2NWDVI
Live frontend demo https://ethnoharvest.vercel.app

About the live demo. The Vercel link hosts the frontend only, so judges can explore the interface. The full backend runs the population models, the Kew and World Flora Online name checks, the 1.4 million name checklist index, the spatial layers and the evidence agent. It is too heavy for free hosting limits, so it runs locally. The full end to end system appears in our demo video, and the repository README sets it up with one command.

1. Inspiration: the problem we faced

The world relies on medicinal plants, and most of them are still taken from the wild with almost no local evidence about how much a population can survive losing.

Number What it means Source
88 percent of WHO member states (170 of 194) report that their people use traditional and herbal medicine WHO Global Report on Traditional and Complementary Medicine 2019
45 percent of flowering plants are threatened with extinction Royal Botanic Gardens Kew, State of the World's Plants and Fungi 2023
723 medicinal plant species are threatened with extinction, with overharvesting a major driver Kew, State of the World's Plants and Fungi 2020
About 3,000 medicinal and aromatic plant species are in international trade TRAFFIC and IUCN, Wild at Home, 2018
60 to 90 percent of those traded species are harvested from the wild TRAFFIC and IUCN, 2018
Only 7 percent of medicinal and aromatic plants with documented uses have been assessed for the IUCN Red List TRAFFIC and IUCN, 2018
3 times growth in international trade value between 1999 and 2015, to over US$3 billion TRAFFIC and IUCN, 2018
Over US$100 billion estimated global herbal medicine market (published estimates vary by definition) Industry market reports, 2024

The gap is not a lack of data. Population studies, plant checklists, occurrence records and trade listings all exist. The gap is that a harvesting cooperative cannot join them up. To answer one question honestly ("how much can we take here?") they would need a botanist, a demographer, a GIS analyst and a lawyer. Generic calculators fill the gap with a confident number built on too little data, and a wrong number is worse than no number.

2. Case studies: three plants, three plant parts, one pattern

We anchored the project on three signature plants. Each is harvested for a different part, and each shows the same failure: global demand, no local evidence.

Jatamansi (Nardostachys jatamansi) African cherry (Prunus africana) Frankincense (Boswellia papyrifera)
Part taken Rhizome; the whole plant dies Bark Resin, by tapping the trunk
Where Nepal Himalaya Montane Africa (Cameroon, Madagascar and others) Ethiopia, Eritrea, Sudan
Scale of harvest 100 to 500 tonnes of rhizomes traded each year; traded by some of the more than 300,000 Nepali households involved in medicinal plant trade 3,225 tonnes of dried bark exported in 1997, the peak year 21,786 trees studied in 23 populations
What is happening Listed as threatened by IUCN; CITES Appendix II; illegal export reported CITES Appendix II since 1995; trade suspensions for several countries; wild harvest judged no longer sustainable 18 of 23 populations had almost no young trees; production projected to halve within 20 years
Source TRAFFIC; Paudel et al. 2023, PNAS Nexus CITES Plants Committee PC25 Doc 27; Stewart 2003, CIFOR Bongers et al. 2019, Nature Sustainability 2:602

What our pipeline found for Jatamansi. We reproduced a real harvest experiment (Ghimire et al. 2008, Journal of Applied Ecology 45:41) from the COMPADRE population models. Our reproduction ranks every harvest level in the same order as the paper, a perfect rank match.

Harvest intensity Population growth rate, our reproduction Population growth rate, published paper
0 percent 1.028 1.014
10 percent 0.958 0.928
25 percent 0.891 0.838
50 percent 0.831 0.779
75 percent 0.737 0.637

A growth rate below 1 means the population shrinks. Even 10 percent harvest pushes Jatamansi below that line. Each extra 10 percent of harvest costs about 0.036 in growth rate. Our decomposition (LTRE) shows why: the decline comes from growth (38 percent), shrinkage (31 percent) and survival (29 percent), and only 2 percent from seeds. So "leave the seeds" is the wrong advice for this plant; the harvest kills the plant itself.

3. Why this project is needed

  • No tool joins the evidence up. Name checklists, population models, maps and trade rules live in separate places with separate formats.
  • Confident guesses cause harm. Our own experiment E05 showed that generic removal models miss the real harvest effect by 0.068 in growth rate. That is the difference between a growing population and a collapsing one.
  • Small budgets need guidance. A cooperative or field team can afford only a few measurements per season and needs to know which ones unlock a decision.
  • Traceability is on paper. Buyers and regulators increasingly ask where a batch came from, and documents with the wrong plant part, an expired authorisation or a synonym name slip through.

4. What EthnoHarvest AI does, end to end

EthnoHarvest AI is a six step pipeline. Each step can stop the next one, so the app never produces an answer it cannot defend.

Step What happens Main data or method
1. Name Checks the plant name against two world checklists and shows its native range Kew WCVP (the data behind Plants of the World Online) and World Flora Online
2. Demography Runs published population models to see how harvest changes growth COMPADRE matrix models, growth rate, LTRE, harvest dose response
3. Evidence tier Grades what we know: A (backed by a harvest study), B (model only), C (insufficient) Audited harvest studies and quality control of matrices
4. Gate Asks "can a safe local harvest rate be estimated?" and shows which of six local evidence items exist Rule based gate; answer is "not yet" until all six exist
5. Plan Picks which measurements to collect next within a budget, explaining why for this species Integer linear program; an AI evidence agent that quotes papers word for word
6. Operate Records each harvest batch and checks its documents from field to buyer Harvest Passport, document checks, stock and recall

Our core rule: unknown is never zero. If something is missing, the app says it is missing, explains why it matters and shows how to get it. It never issues a quota, a permit or a "safe rate".

5. Features

Explore: the globe

  • A 3D globe (CesiumJS) with GBIF occurrence points, cleaned per region.
  • Native and introduced range from Kew WCVP, drawn over 369 world regions (TDWG level 3). Native is green, introduced is amber.
  • A region panel with climate (WorldClim, NASA POWER), tree cover loss (Global Forest Watch), active fires (NASA FIRMS) and land cover (ESA WorldCover).
  • A species dossier with photos (open licence, credited), CITES status, documented medicinal uses and a Name Check card showing whether Kew and World Flora Online agree.
  • The safe harvest question box, which answers "not yet" and lists what local evidence is missing.
  • A case library of ready made harvest scenarios, such as Jatamansi rhizome harvest in the Nepal Himalaya.

Simulate: the 3D Living Ecosystem Studio

  • A 3D field (Three.js) where every plant's size shows its life stage (seedling, young, adult), taken directly from the population model.
  • Harvest sliders and a year by year playback, so a collector can see what a harvest level does over time.
  • 33 plant models, species specific where available, with generic models by growth form.
  • A Science view with the life cycle board, the growth rate chart and the LTRE breakdown, all from the same simulation as the 3D view.

Plan: deciding what to measure

  • Evidence planner. An integer linear program chooses the measurements that unlock the most decisions within a budget, such as 20 person days. Each item explains why it matters for this species (for example, large tree survival for bark harvest, recruitment for fruit or resin harvest).
  • Harvest methods. Illustrated guides to how each plant part is harvested, with numbered markers and caveats. Illustrations annotated with AI are labelled "AI annotated illustration".
  • HarvestOps and the Harvest Passport. Records harvest area, plant part, fresh and dry weight and drying. Uploaded documents are checked automatically for problems such as wrong plant part, expired authorisation, wrong species or synonym names. Every page states that a Harvest Passport is not a permit.

Evidence: the AI agent and the science record

  • Evidence agent. Searches the literature and extracts evidence. Every claim must quote its source word for word; a second model check runs and any failures are shown.
  • Run graph. Shows each step the agent took, so judges and users can audit it.
  • Workflow builder. 19 agent components (CITES lookup, regulation cards, eligibility gate, harvest comparison, recovery rotation, evidence gaps and more) that can be chained into custom workflows.
  • Experiments tab. All 15 preregistered experiments, including the ones that failed.

6. A closer look at the product

The features above are the map. This section walks through the parts a user actually touches, and why each one is built the way it is.

6.1 The 3D Living Ecosystem Studio: seeing a harvest before it happens

Most people cannot read a population matrix, but anyone can read a field of plants. The Studio turns the model into something a collector, a cooperative board or a buyer can understand in seconds.

What you see What it means Where it comes from
A field of individual plants The population being harvested The stage counts in the population model
Small, medium and large plants Seedlings, young plants and adults The life stage of each individual; the size key on screen explains it
Plants thinning or filling in as years play Whether the population shrinks or recovers The same yearly projection used in the growth rate chart
A harvest slider (for example 0, 10, 25 or 50 percent) How much of the harvested part is taken each year The harvest dose response applied to the plant part (rhizome, bark, resin, leaf, fruit)
Species specific 3D plants What the real plant looks like Compressed 3D models for the signature species, generic models by growth form for the rest

How a user gets there. From the globe, a user picks a case such as Jatamansi in Nepal and presses "Simulate this". The Studio opens with that species, that plant part and that region already loaded, so there is no setup.

Why it matters for a product. The Studio is the conversation starter. A cooperative leader can show members what 25 percent harvest does over ten years without a single equation. A buyer can see why a supplier is choosing a lower harvest. Because the picture and the numbers come from the same simulation, the 3D view never tells a different story from the charts.

What it refuses to do. When the species only has a generic model (tier B or C), the Studio says so on screen. It is a way to understand a model, not a way to set a quota.

6.2 The globe: one screen instead of six websites

Without EthnoHarvest AI, checking one plant in one valley means opening a checklist site, a GBIF map, a CITES page, a climate dataset, a forest loss map and a fire map. On the globe these are layers on one view.

  • Click a region to see its climate, tree cover loss, recent fires and land cover in one panel.
  • Turn on the native range to see whether the plant is native, introduced or not recorded in that region. The app adds that "not recorded" is not proof of absence.
  • Open the dossier for photos, documented uses, CITES status and the Name Check, which shows whether the two world checklists agree on the name.
  • Read the answer box, which says plainly whether a safe local harvest rate can be estimated yet, and what is missing.

6.3 The Evidence Planner: a shopping list for the next field season

The planner answers the question every small team has: "We have 20 person days. What should we measure?" It does not just rank items; it picks the combination that unlocks the most decisions, because some evidence is only useful together. Each item carries a plain reason, such as "bark harvest removes large trees, so measure large tree survival first". The output is a list a field team can take into the season.

6.4 Harvest Methods: how the part is taken

Each species has a page showing how its harvested part is collected, with an illustration, numbered markers on the relevant parts and the caveats that matter for that method (for example, tapping depth for resin, or the fact that digging a rhizome kills the plant). Every illustration annotated with AI is labelled as such, and the names and facts drawn over it come from our own data.

6.5 HarvestOps and the Harvest Passport: from the field to the buyer

This is where EthnoHarvest AI stops being a research tool and becomes an operations tool.

Stage What the user does What the app does
Collect Records the area, the species, the plant part and the fresh weight Creates a Harvest Passport with its own ID
Dry and store Records the dry weight and drying method Tracks stock and the ratio of fresh to dry weight
Attach documents Uploads the collection authorisation, the transport slip and similar papers Reads each document and checks it against the passport
Check Reviews the flags Flags problems such as an authorisation that covers leaves while the passport covers bark, an expired date, a synonym or a wrong species
Sell or recall Shares the passport with a buyer, or traces a batch back if there is a problem Keeps the full record, so any batch can be traced to where and when it was collected

Every passport page states clearly that a Harvest Passport is not a permit. In the demo, all documents are synthetic and watermarked.

6.6 The Workflow Builder: make your own checks

The 19 agent components are building blocks. An organisation can chain them into its own routine, for example: check the name, look up CITES status, apply the regulation card for the country, run the eligibility gate, then list evidence gaps. A certifier could run the same workflow on every supplier; a researcher could run it on a list of species. Every run shows its run graph, so the result can be audited.

7. More than a research tool: EthnoHarvest AI as a product

EthnoHarvest AI is built on research, but it is designed to be used every season by people who are not scientists.

7.1 Product principles

Principle What it means in the app
One question per screen Each page answers one plain question: is the name right, where does it grow, what happens if we harvest, what should we measure, is this batch in order
Show, then explain The 3D field and the globe come first; charts and numbers are one click away
Honest by default Missing data is shown as missing; "not yet" is a valid answer, with the next step attached
Ready made starting points A case library and demo passports mean a new user sees a working example before entering any data
Every claim traceable Each fact links to its dataset, licence and version; each AI claim quotes its source
Runs on a laptop No special hardware or paid cloud services are needed to run the full stack

7.2 One product, many jobs

User The job they hire EthnoHarvest AI for Main screens
Cooperative manager Plan the season, explain harvest choices to members, keep batches in order Studio, Planner, HarvestOps
Collector See what different harvest levels do, learn the right method Studio, Harvest Methods
Field researcher or NGO Decide what to measure on a small budget, gather published evidence Planner, Evidence agent
Buyer or herbal product company Check suppliers' batches before purchase, keep a recall record HarvestOps, Workflow Builder
Certifier or auditor Run the same checks on every supplier and keep an audit trail Workflow Builder, run graph, HarvestOps
Regulator or CITES authority See the evidence behind a harvest request in one place Globe, dossier, evidence tiers

7.3 Why someone would keep using it

  • It saves work. Name, range, trade status, site conditions and published evidence are gathered in one place instead of many.
  • It reduces risk for buyers. A wrong plant part or an expired authorisation found before purchase avoids a rejected shipment or a legal problem.
  • It helps cooperatives sell. A clean, traceable Harvest Passport is a stronger offer to a buyer who must prove responsible sourcing.
  • It gets better every season. Each measurement a cooperative collects fills a gap in the gate, moving a species from "not yet" toward a locally supported answer.

7.4 How it could be sustained (proposed)

These are plans, not current customers or revenue.

Model Who What they get
Free Harvesting cooperatives, community groups, students and researchers The full app, open source, running locally
Paid hosted service Buyers, herbal product companies, certifiers Hosted backend, supplier dashboards, shared passports and workflow runs
Grants and partnerships Conservation NGOs, development agencies, botanical gardens Pilots in specific regions and support for new species

7.5 Path to adoption

  1. Pilot with one Jatamansi cooperative in Nepal (planned), using the Studio and Planner before the season.
  2. Add HarvestOps for that cooperative's batches and one partner buyer.
  3. Expand to the other signature plants, Prunus africana and Boswellia, with local partners.
  4. Open the Workflow Builder to certifiers so the same checks run across many suppliers.

8. Data at a glance

Number What
1,448,984 plant names scanned in Kew's World Checklist of Vascular Plants (WCVP)
365,813 accepted vascular plant species in it
1,995,338 native range records read and mapped to 369 world regions
9,178 population matrix models checked (COMPADRE v6.26.3.0)
812 plant taxa from 546 studies in 64 countries
200 medicinal taxa with population models
752 of 812 names on which Kew and World Flora Online agree; 17 disagreements shown to the user
3,447 documented medicinal use records from 870 references (USDA Dr Duke, Herbolario)
13 open data sources, each with its licence and version recorded
Question the user asks Dataset
Does harvest change population growth? COMPADRE Plant Matrix Database
Is the name right? Where is it native? Kew WCVP and World Flora Online
Where has it been seen? GBIF (including iNaturalist research grade records)
Is it traded internationally? CITES Species+
What is the climate? WorldClim 2.1 and NASA POWER
What threatens the site? Global Forest Watch, NASA FIRMS, ESA WorldCover
When does it flower? PhenObs (Zenodo)
Is it medicinal? USDA Dr Duke database and Herbolario (documented use only)

Not used yet, with the honest reason: the IUCN Red List API (token pending) and TRY plant traits (requires a reviewed data request). Both slots are already in the layer list, marked pending, and are never filled with defaults.

9. Scientific rigour: what worked and what did not

We registered 15 experiments before running them, with a fixed random seed. Every result is saved as a JSON file with the exact command that produced it. Six were negative or inconclusive, and we published them all and changed the app because of them.

Experiment Finding What we changed in the app
E03 Jatamansi harvest reproduction Same ordering as the published paper; 0.036 loss in growth rate per extra 10 percent harvest Used as the reference case
E04 Does harvest lower growth across 8 studies? 6 of 8 lower, but not significant (p = 0.29) No transfer of a harvest effect between species
E05 Can generic removal models predict the experiment? No; error of 0.068 in growth rate Model only results carry a warning and never become harvest advice
E09 Does an independent demography archive replicate? No clear answer Not used for decisions
E10 Habitat model accuracy Spatial AUC 0.786; random validation would have overstated accuracy by up to 0.19 Only spatial cross validation is used
E11 Does AI triage papers better than keyword rules? No; macro F1 0.768 versus 0.773 AI stays advisory; rules decide
E13 Does our planner beat a greedy choice? Equal or better in 24 of 24 budget cases, strictly better in 15 Planner used for evidence plans
E14 Does harvest hurt Khaya bark in a stochastic model? No detectable penalty Shown as inconclusive, not used
E15 Do poor quality matrices change answers? Yes; 15 percent fail quality control and 9 species shift growth rate by more than 0.05 Failing matrices are excluded

Document checks. On 13 synthetic document packets (37 PDFs written by us and watermarked), rules alone found 14 of 15 planted problems with zero false alarms. With the grounded AI, all 15 were found. A held out layout the rules had never seen was flagged as "fields not found" rather than passed.

Testing. 140 Python tests and 35 browser tests, all passing.

10. Who uses it and how they benefit

User Their question How EthnoHarvest AI helps
Harvesting cooperatives and collector groups (for example, rhizome collectors in western Nepal; this is an illustrative persona, not a current customer) How much can we take, and can we prove where our batch came from? Name and CITES check before the season, a clear "not yet" with the missing evidence listed, and a Harvest Passport that catches document problems before the buyer does
Field researchers and conservation NGOs What should we measure first with a small budget? A budgeted evidence plan with reasons for each species, built on published population models
Forest and wildlife authorities, CITES scientific authorities Is this harvest supported by evidence? Transparent evidence tiers, native range checks and regulation cards, with every source cited
Buyers, herbal product companies and certifiers Is this batch correctly named, legally collected and traceable? Document checks for wrong plant part, wrong species, synonyms and expired authorisations, plus stock and recall records
Pharmacognosy and ethnobotany researchers Which medicinal plants are at risk, and where is the evidence weakest? One place combining names, ranges, demography and documented uses, with every gap shown

A season with EthnoHarvest AI, for a cooperative:

  1. Before the season: check the name, the native range and the CITES status, and see what evidence is missing.
  2. Plan: spend the measurement budget on the items that unlock the most decisions.
  3. Harvest: a Harvest Passport records the area, the part taken, the weights and the drying.
  4. Sell: documents are checked automatically, so problems are caught before the buyer sees them.

11. Responsible research

Principle What EthnoHarvest AI does
Licences and attribution Licence and credit stored for every dataset and every photo; COMPADRE raw data is never redistributed
Privacy Coordinates of sensitive species are generalised to 0.1 degrees; uploads stay private
Indigenous and traditional knowledge No traditional knowledge is collected or stored; only published, licensed records are used
Scientific integrity Unknown is never zero; negative results are published
No medical claims Medicinal use is shown as "documented use", never as efficacy
No false authority No quotas, permits or safe rates are ever issued
Responsible AI AI is advisory, quotes its sources word for word, is clearly labelled and was benchmarked against rules (E11)
Synthetic data All demo documents are synthetic and watermarked

12. How we built it

Layer Tools
Backend Python, FastAPI, SQLite
Science NumPy, SciPy, matrix population models, LTRE, PuLP (integer linear programming), Shapely for spatial range overlap
Frontend React 19, TypeScript, Vite
Visuals CesiumJS globe, Three.js 3D studio
AI Evidence agent with verbatim quote checks and a second model verification step; AI annotated illustrations reviewed by eye and labelled
Testing pytest (140 tests) and Playwright (35 browser tests)
Hosting Frontend demo on Vercel; full stack runs locally

Security. API keys stay on the server and are never committed or sent to the browser.

13. Challenges we ran into

  • Joining two world checklists. Kew and World Flora Online sometimes disagree on names. Rather than silently choosing one, we show the 17 disagreements to the user.
  • Scale. Reading 1.4 million names and 2 million range records and mapping them to 369 world regions had to stay fast enough to rebuild on a laptop. The full checklist build reruns in about 35 seconds.
  • Making science readable. Population models are hard to explain. The 3D Studio was our answer: plant size shows life stage, so the model becomes a picture.
  • Our own AI did not win. Experiment E11 showed the AI was no better than keyword rules at triaging papers, so we kept rules in charge and made the AI an advisor.
  • Hosting. The backend is too heavy for free hosting tiers, so we deployed the frontend on Vercel and documented the local setup.

14. Accomplishments we are proud of

  • A tool that is honest by design: it says "not yet" instead of guessing.
  • A reproduction of a real Jatamansi harvest experiment from published population models.
  • A 3D Studio that lets anyone see what a harvest does, driven by the same model as the science.
  • 15 preregistered experiments, with every failure published and acted on.
  • A full chain from the globe to the warehouse in one working app.

15. What we learned

  • A missing value must never be treated as zero, in code or in conservation.
  • Spatial validation matters; random validation would have overstated our habitat model accuracy.
  • Negative results are useful; four of our app's rules exist because an experiment failed.
  • AI is most useful when it must show its sources and is tested against simple baselines.
  • People trust what they can see. Showing a model as a living field made it far easier to explain than any chart.

16. Limits and future directions

Honest limits today Next six months Why it scales
Habitat models for 2 species only Pilot with one Jatamansi cooperative in Nepal (planned) Any of the 200 COMPADRE medicinal taxa works today; new plants enter as tier C, never with invented data
Jurisdiction rules for 4 countries Add IUCN Red List and TRY plant traits Open data and open source; runs on a laptop
AI illustrations not yet reviewed by experts Expert review by pharmacognosists and ethnobotanists Rules are versioned, so each country can add its own
Demo documents are synthetic A real document benchmark with a cooperative Name and range checks already cover 365,813 species
Backend not publicly hosted Hosted backend with caching for the heavy layers

Further directions:

  • Offline mobile mode for collectors in areas without internet.
  • Multilingual interface, starting with Nepali, Amharic and French.
  • Links to certification schemes such as FairWild.
  • Community controlled data sharing, so cooperatives decide what leaves their hands.
  • Shared Harvest Passports that a buyer can open directly from a QR code on the batch.

17. Key references

  • WHO Global Report on Traditional and Complementary Medicine, 2019.
  • Royal Botanic Gardens Kew, State of the World's Plants and Fungi, 2020 and 2023.
  • TRAFFIC and IUCN, Wild at Home, 2018.
  • TRAFFIC, Jatamansi trade from Nepal.
  • Paudel et al. 2023, PNAS Nexus.
  • CITES Plants Committee, PC25 Doc 27 (Prunus africana).
  • Stewart 2003, CIFOR (Prunus africana).
  • Bongers et al. 2019, Frankincense in peril, Nature Sustainability 2:602.
  • Ghimire et al. 2008, Journal of Applied Ecology 45:41.
  • Govaerts et al. 2021, The World Checklist of Vascular Plants, Scientific Data 8:215.
  • COMPADRE Plant Matrix Database, version 6.26.3.0.
  • World Flora Online.

Evidence before extraction.

Eco Roots: Shruthi and Vishnu

Built With

Share this project:

Updates

Submission history