Gunung Nyiut zoning explorer

This repository plans conservation zones for a community-forest landscape in West Kalimantan (Indonesian Borneo) and lets you explore the trade-offs between biodiversity, carbon and livelihoods. It is reproducible end to end and uses open data only.

The landscape is the Cagar Alam Gunung Nyiut nature reserve plus a 10 km belt of surrounding village land, 294 kha in total. For every target combination, the pipeline finds the zoning (protection / sustainable use & agroforestry / community use) that meets the biodiversity and carbon targets at the smallest loss of livelihood opportunity. A static web app then lets anyone move through 150 precomputed plans with sliders.

Open the explorer → · 2-page methods note · scope · build report

Demonstrator, not a recommendation. The real project landscape and field data are not public, so the habitat and livelihood layers are transparent proxies (see limitations). The pipeline is built so that the project's own layers can replace them.

What it shows

• Where the reserve falls short. The reserve already holds 50–56 % of each indicator species' modelled habitat and 84 % of montane forest. It holds only 17 % of lowland and 32 % of riparian forest, so those two features drive most of the zoning outside it.
• Carbon targets are the expensive constraint. Livelihood opportunity kept falls from 90 % (reserve only) to 83 % at 50 % biodiversity / 60 % carbon targets, and to 55 % at 70 % / 90 %.
• Agroforestry compatibility decides the high-ambition plans. With a 90 % carbon target, livelihood opportunity kept is 51 % if agroforests retain little biodiversity and 63 % if they retain a lot (then 123 kha go to the agroforestry zone). Measuring this is exactly the job of camera-trap and acoustic surveys in agroforests.
• Connectivity is cheap. The boundary penalty cuts zone edge length by 43 % on average while keeping most of the livelihood opportunity.

How it works

flowchart LR
  A[Open data<br/>STAC / COG range reads] --> B[100 m layers<br/>forest, carbon, BII, DEM,<br/>roads, settlements]
  B --> C[Derived layers<br/>connectivity, condition,<br/>travel time, streams,<br/>livelihood opportunity]
  C --> D[Features & costs<br/>4 forest types, 7 species,<br/>carbon; 11,730 x 25 ha PUs]
  D --> E[Multi-zone MILP<br/>min-set + boundary penalty<br/>HiGHS]
  E --> F[150 scenarios<br/>metrics + zonings]
  F --> G[Static explorer<br/>MapLibre, GitHub Pages]
  F --> H[Figures + methods PDF]

Formulation. This is the prioritizr / Marxan-with-Zones minimum-set problem. Each planning unit is assigned to one zone. Each feature f must reach its target, counting x_P + κ_f · x_A (protection secures all of a feature, agroforestry a share κ). The objective is the livelihood opportunity forgone plus a zone-boundary penalty; P|C edges cost twice as much as P|A or A|C edges, which rewards buffered protection cores. It is solved exactly as a mixed-integer programme with HiGHS (all 150 plans within 0.9 % of optimal, ~2.6 s each). Six scenarios re-solved in R with prioritizr 9.0.1 give identical objectives and identical zonings (results/prioritizr_crosscheck.csv); the cross-check script is R/prioritizr_crosscheck.R. Full details are in the methods note and src/zoning/solver.py.

Scenario axes:

• biodiversity target (30–70 %)
• carbon target (30–90 %)
• agroforestry compatibility κ (low / central / high)
• connectivity penalty (off / on)

Data (all open, no accounts, clipped on read)

Layer	Source	Licence
Tree cover 2000, loss 2001–2024	Hansen/UMD/Google/USGS/NASA Global Forest Change v1.12	CC BY 4.0
Land cover 2021, 10 m	ESA WorldCover v200 (AWS COG)	CC BY 4.0
Above-ground biomass 2024 + SD	ESA CCI Biomass v7.0 (CEDA)	ESA CCI data policy (free, attribution)
Elevation / slope / streams	Copernicus DEM GLO-30 (AWS COG)	Copernicus DEM licence
Biodiversity Intactness Index 2020	Impact Observatory & Vizzuality (Planetary Computer STAC)	CC BY 4.0
Roads, settlements, rivers, reserve boundary	© OpenStreetMap contributors (Overpass API)	ODbL
Regency boundaries	geoBoundaries gbOpen (BPS / OCHA)	open
Species occurrences (plausibility check only)	GBIF occurrence API	per record
Basemap (app only)	Sentinel-2 cloudless 2023 by EOX	CC BY-NC-SA 4.0

Downloads use HTTP range requests on Cloud-Optimised GeoTIFFs or tiles. Only the study window is transferred (about 55 MB in total). Every request is logged with URL, size and date in data/manifest.json.

Reproduce

Requires uv (the Python 3.12 environment is locked in uv.lock).

uv sync --locked --all-groups
uv run zx all        # or: make all
make serve           # explorer at http://localhost:8000

zx all runs these stages, each cached and individually callable (uv run zx <stage>):

Stage	What it does
aoi	Study area from OSM + geoBoundaries
download	Clipped downloads
layers	100 m harmonised and derived layers
check	GBIF plausibility check
problem	Planning units, features, costs
solve	150 scenarios with HiGHS, ~2 min on 5 cores
export	App data
figures	Figures and docs/numbers.json
methods	Typst → docs/methods.pdf

Use --force to rebuild a stage. The optional zx crosscheck stage compares the solver with prioritizr; it needs R with prioritizr, highs and tibble. Rscript is found on PATH, via ZX_RSCRIPT, or via an uncommitted config/local.yaml with rscript: and r_libs: keys. The study area, zones, features and scenario grid are plain YAML in config/. Pointing config/study_area.yaml at another landscape re-runs everything there.

config/        study area, zones, features/species rules, scenario grid
R/             prioritizr cross-check (reference formulation in R)
src/zoning/    pipeline package (sources, layers, hydrology, accessibility, species,
               problem, solver, scenarios, metrics, export, figures, cli)
app/           static explorer (index.html, app.js, style.css) + generated data/
data/          manifest.json (provenance) + processed/ planning-unit table
results/       scenario metrics, zonings, GBIF check
docs/          methods.typ/.pdf, figures, screenshots
tests/         pytest suite

Tests and CI

make test runs ruff and the pytest suite, which covers:

• the HiGHS solution matching brute-force enumeration on small instances (with and without the boundary penalty)
• locked units, target attainment and capping of infeasible targets
• monotone cost in the targets
• travel-time geometry, road and barrier handling
• D8 flow routing
• habitat-rule bounds
• a re-solve of committed scenarios against the stored results
• consistency of the app data

GitHub Actions runs the same suite and builds the methods PDF. A second CI job installs R and prioritizr and requires the cross-check to match. A separate workflow deploys app/ to GitHub Pages.

Limitations

• Species layers are rule-based habitat models. They combine elevation, forest structure, connectivity and hunting pressure, with parameters taken from IUCN accounts as expert judgement. They are not occupancy models. GBIF illustrates the data gap: no helmeted hornbill record within 500 km, and no clouded leopard or sun bear record within 250 km.
• The livelihood opportunity index is a biophysical proxy (slope, elevation, travel time, current use). It does not measure income, tenure or customary use. Village boundaries and existing permits are not represented.
• ESA CCI biomass is uncertain. Per-pixel uncertainty is large, and the product overestimates low biomass. Hansen "loss" includes swidden and plantation rotation.
• The study area is an approximate stand-in for the studentship's 12-village landscape, whose boundaries are not public.

Intended use: supporting co-production

The explorer is meant for workshops. Communities and partners set targets and zone rules, see the consequences immediately, and discuss the protection-frequency map: units protected in almost every plan are robust, while units that switch between plans are where choices matter. Real zoning requires free, prior and informed consent, participatory mapping of customary land and resource use, and community-defined targets. The tool supports those processes; it does not replace them.

Next steps

See REPORT.md. In brief:

• occupancy layers from validated camera-trap and acoustic data
• empirical κ from surveys in agroforests
• SMART patrol threat layers
• village-level equity constraints
• a participatory serious-game mode
• coupling zonings to an agent-based model of forest use

Author and licence

Jan Hildebrand (@yanhildebrand).

• Code: MIT.
• Derived data in app/data, results and data/processed: CC BY 4.0, subject to the upstream licences above. OSM-derived extracts are ODbL.

Citation metadata is in CITATION.cff.