agribound

Agricultural field boundary delineation from satellite imagery

Overview

Agribound runs published field-boundary models, geospatial foundation models and embedding-based methods on satellite and aerial imagery through one configuration and one pipeline: composite → optional fine-tuning → delineation → optional SAM refinement → study-area selection → post-processing → LULC crop filter → export. It supports ten sources (Landsat, Sentinel-2, HLS, NAIP and SPOT 6/7 composites built on Google Earth Engine; USGS NAIP Plus without Earth Engine; local GeoTIFFs; Google Satellite Embedding and TESSERA embeddings) and seven engines (Delineate-Anything, Fields of The World, GeoAI Mask R-CNN, DINOv3, Prithvi-EO-2.0, embedding clustering and ensembles).

Every run is seeded, caches its intermediates under content-addressed names, and writes a provenance record (configuration and hash, package versions, device, step timings, model weights, counts, warnings) next to its output. The package also provides object-level evaluation against reference boundaries, tiling of large regions for HPC clusters, a query helper for the published Fields of The World polygons, and an optional agent layer that proposes one run for a human to approve.

Upgrading from 0.1.x? Version 1.0.0 fixes defects that affected results produced with agribound 0.1.x, among them FTW season windows that were copies of the annual composite, Landsat/HLS inputs on the wrong radiometric scale, a silent Delineate-Anything fallback with swapped red/blue channels, and caches that ignored the study area and year. See the affected results and the migration guide.

How It Works

The agribound 1.0 workflow (select the image for full resolution): a six-stage pipeline from ten imagery and embedding sources (0.3–30 m, 1984–present) to field boundaries, with a deterministic entry point and an optional, human-confirmed agent layer above it.

Composite. Earth Engine builds a median or greenest-pixel (max-NDVI) composite for a year or a date window and exports it on a UTM grid. NAIP is mosaicked, and only Landsat, Sentinel-2 and HLS are cloud-masked and scaled to reflectance ×10 000. USGS NAIP Plus, TESSERA and local GeoTIFF inputs are read without Earth Engine.
Fine-tuning (optional). Full (Delineate-Anything, GeoAI, DINOv3, Prithvi) or LoRA (DINOv3, Prithvi) fine-tuning on reference boundaries, validated by default on a spatially blocked split (5 km blocks). GeoAI, DINOv3 and Prithvi's UPerNet mode need a checkpoint, from fine-tuning or supplied by the user.
Delineation. One of seven engines, coloured by family: task-specific segmentation, geospatial foundation model, label-free embedding clustering and multi-engine ensemble.
Refine and post-process. Optional SAM refinement (SAM 2, 2.1 or 3; the SAM 3 backends are untested), then study-area selection, merging, minimum-area filtering, smoothing and simplification.
LULC crop filter. Removes polygons whose crop fraction is below 0.3, computed on Earth Engine or locally on a downloaded crop raster. Annual NLCD, Dynamic World or C3S Land Cover is selected by coverage and year; CDL (CONUS only) is used on request.
Export. GeoParquet (fiboa-style columns), GeoPackage or GeoJSON, with per-field area, perimeter, compactness and crop fraction, plus a provenance.json record.

Around the pipeline:

Entry point. delineate() and agribound delineate --config run the six stages directly. Every run is seeded and uses a content-addressed cache, and provenance.json is written by default.
Agent layer (optional). A language model, reached through the Claude API or an MCP host (or a local Anthropic-compatible server via base_url), investigates with typed read-only tools and proposes one configuration. It runs only after you confirm that exact plan at the human gate, with an approval bound to the plan's hash and used once. At most one plan runs per session, and the session then stops (see Agent layer).
Scale out and evaluate. agribound tiles make, run and merge split a large study area into tiles that run as Slurm array jobs (see HPC and large areas). evaluate() scores results against reference boundaries with object-level and area-weighted metrics (see Evaluation).

Results

From the agribound 1.0.1 example runs (the San Juan County map shows 1.0.0 outputs, which 1.0.1 reuses unchanged). Each map is drawn on a composite from the run, named under the map with the model and its version: usually the engine's input; for FTW, its window A; for the SAM-refined embedding panels (Pampas, top), the Sentinel-2 composite SAM 2 read. Select an image for the full-resolution file. See the gallery for more regions and engines.

From 30 m to 1 m: Delineate-Anything v2 against a reference registry (San Juan County, New Mexico, USA)

Example 20: Delineate Anything v2, used as released, on Landsat, Sentinel-2, SPOT 6/7 and NAIP of 2018, each evaluated against the 944 NMOSE WUCB polygons (cyan; not used for training or fine-tuning in these runs). Object F1 (IoU ≥ 0.5) is 0.15 at 30 m, 0.34 at 10 m, 0.33 at 6 m and 0.43 at 1 m; recall rises from 0.08 to 0.44.

Supervised: DINOv3 fine-tuned + SAM 2 on four sources (eastern Lea County, New Mexico, USA)

Example 14: DINOv3 (SAT-493M weights) fine-tuned on the NMOSE polygons for each source and refined with SAM 2. In-sample F1 (the polygons are also the training labels): 0.06 on Landsat, 0.38 on Sentinel-2, 0.45 on SPOT and 0.59 on NAIP; at 30 m the box also holds only 4 training chips, against 2,014 at 1 m.

Label-free: embeddings + SAM 2 vs Delineate-Anything v2 (Pampas, Argentina)

Example 15, no reference data or training: Google Satellite Embedding and TESSERA v1 clusters of 2024, crop-filtered and refined with SAM 2 on Sentinel-2 (top; parts over 50 ha kept unrefined), against Delineate Anything v2 on the same Sentinel-2 composite and on SPOT 6/7 2023 (bottom); centre pivots near Pergamino. Orange = refined by SAM 2. The embedding panels come from the agribound 1.0.1 run of 2026-09-29; the Delineate-Anything panels are the 1.0.0 outputs, which that run reused. The gallery adds the whole study area and three zoomed windows.

Satellite Sources

From agribound.registry (agribound list-sources); facts checked against the Earth Engine catalogue and the providers on 2026-09-26 to 2026-09-28.

Source	Key	Export resolution	Years	Coverage	Value scale	Earth Engine
Sentinel-2 L2A (harmonized)	`sentinel2`	10 m	2017-present	global (2017-2018 L2A not global)	reflectance × 10000	yes
Landsat 5/7/8/9 C2 L2	`landsat`	30 m	1984-present	global	reflectance × 10000	yes
HLS v2.0 (L30 + S30)	`hls`	30 m	2013-present (S30 2015-)	global land	reflectance × 10000	yes
NAIP	`naip`	1 m (`naip_resolution_m`; native 0.6 m in most states since 2018)	2002-2023	conterminous US	uint8	yes
USGS NAIP Plus ImageServer	`usgs-naip-plus`	0.3-0.6 m (finest selected footprint)	2012-2023, latest vintage per state only	US states and territories	uint8	no
SPOT 6/7 multispectral	`spot`	6 m	2012-2023	global, restricted	uncalibrated DN	yes
SPOT 6/7 panchromatic	`spot-pan`	1.5 m	2012-2023	global, restricted	uncalibrated DN	yes
Local GeoTIFF	`local`	the file's	any	user-provided	unknown	no
Google Satellite Embedding (AlphaEarth)	`google-embedding`	10 m, 64-D	2017-2025	global land	embedding	default backend; `source_coop` backend needs none
TESSERA	`tessera-embedding`	10 m, 128-D	v1 2017-2025 (near-global 2024), v1.1 2015-2025, v2 beta	depends on version	embedding	no

Composites cover the study area's bounding box in the UTM zone of its centroid (export_crs) without polygon masking. The LULC crop filter reads Earth Engine for every source. Details: Satellite sources.

Delineation Engines

Engine	Key	Approach	Label-free	Fine-tunable	Default weights / model	GPU
Delineate-Anything	`delineate-anything`	YOLO11-seg instance segmentation	yes	yes	`large_v2` (Delineate Anything v2) from `MykolaL/DelineateAnything`, pinned revision, SHA-256 checked	recommended
Fields of The World	`ftw`	semantic segmentation with ftw-tools checkpoints, polygonised	yes	no	the ftw-tools registry default, `FTW_PRUE_EFNET_B5` in ftw-tools 2.0.0b5 (two crop-calendar season windows)	recommended
GeoAI	`geoai`	Mask R-CNN instance segmentation (geoai-py)	no: no published field weights; needs fine-tuning or `checkpoint_path`	yes	COCO Mask R-CNN as the fine-tuning start	recommended (CPU on Apple MPS)
DINOv3	`dinov3`	DINOv3 ViT + DPT head (geoai-py)	no: needs fine-tuning or `checkpoint_path`	yes (full by default, LoRA optional)	SAT-493M ViT-L/16 backbone	recommended
Prithvi-EO-2.0	`prithvi`	Prithvi-EO-2.0 ViT (terratorch): `embed` clustering, `pca` baseline, or fine-tuned `segment`	only `embed`/`pca`	yes	`Prithvi-EO-2.0-300M-TL`	recommended
Embedding	`embedding`	K-means clustering of pre-computed embeddings	yes	no	-	no (CPU)
Ensemble	`ensemble`	intersection, union or pixel vote of several engines/models	depends on members	no	members `delineate-anything` + `ftw`	depends

Engine/source support is checked when the configuration is created; there are no silent fallbacks to another model, band set or engine. GeoAI fine-tuning sizes its chips from the reference fields by default, and GeoAI joins the instances that a field was split into at the edges of its inference windows (see Fine-tuning). Optional SAM refinement (sam_refine=True) runs after any engine except embedding (which refines its own polygons, given engine_params["sam_rgb_bands"]), with backends sam2 (default), sam2.1, sam3 (Meta; CUDA + triton, Linux) and sam3-hf (transformers); both SAM 3 backends are currently untested (they have not been run end to end, because the facebook/sam3 weights are gated; agribound logs a WARNING when one is loaded). Fields whose padded bounding box is under sam_min_crop_px (64 px) are not refined. Details: Engines, SAM refinement.

Installation

Python >= 3.12. terratorch (Prithvi) requires lightning>=2.6 and ftw-tools 2.0.0b5 (FTW) requires lightning<2.6, so the full stack needs two environments:

Environment	File	Extra	Includes	Excludes
core	`environment.yml`	`agribound[all]`	GEE, Delineate-Anything, FTW, GeoAI, DINOv3, SAM 2, TESSERA, agent	Prithvi, SAM 3
GFM	`environment-gfm.yml`	`agribound[all-gfm]`	GEE, Delineate-Anything, GeoAI, DINOv3, Prithvi, SAM 2, TESSERA, agent	FTW, SAM 3

git clone https://github.com/montimaj/agribound.git && cd agribound
conda env create -f environment.yml          # or environment-gfm.yml; both install -e .
conda activate agribound                     # (agribound-gfm)

or with pip, in a fresh Python 3.12 environment:

pip install "agribound[all]"        # core
pip install "agribound[all-gfm]"    # separate environment, for Prithvi
pip install "agribound[gee,delineate-anything]"   # or only what you need

Extra	For
`gee`	Earth Engine sources, the default Google-embedding backend, the LULC filter
`delineate-anything`	Delineate-Anything (`numba` is used only by its reference backend)
`ftw`	FTW (`ftw-tools>=2.0.0b5,<3`, a pre-release on PyPI)
`geoai`, `dinov3`	GeoAI, DINOv3 (`geoai-py>=0.43.1`)
`prithvi`	Prithvi (`terratorch[peft]`; conflicts with `ftw`)
`samgeo`	SAM 2 / 2.1 refinement
`sam3`	SAM 3 Meta backend (CUDA; `triton-windows` on Windows); untested (see SAM refinement)
`tessera`, `embedding`	TESSERA (`geotessera>=0.10.2,<0.11`); both embedding sources
`agent`	the agent layer and MCP server (`anthropic`, `mcp`)
`docs`, `dev`	documentation; tests and linting

The GDAL Python bindings (osgeo, conda-forge gdal) are needed only by the Delineate-Anything reference backend. See Installation.

Quick Start (Python)

import agribound

gdf = agribound.delineate(
    study_area="my_region.geojson",  # or "bbox:minx,miny,maxx,maxy", WKT, GEE asset
    source="sentinel2",
    year=2024,
    engine="delineate-anything",
    gee_project="my-gee-project",
    output_path="fields.gpkg",
)
print(len(gdf), gdf.attrs["run_id"])  # fields.gpkg + fields.gpkg.provenance.json

Quick Start (CLI)

agribound delineate \
    --study-area my_region.geojson \
    --source sentinel2 \
    --year 2024 \
    --engine delineate-anything \
    --gee-project my-gee-project \
    --output fields.gpkg

Configuration

Every option is a field of AgriboundConfig; YAML files use the same names, and unknown keys are rejected.

# config.yml
study_area: my_region.geojson
source: sentinel2
year: 2024
engine: delineate-anything
gee_project: my-gee-project

composite_method: median        # median | greenest (max_ndvi is an alias)
cloud_cover_max: 20
export_crs: utm                 # UTM zone of the study-area centroid

aoi_selection: representative_point
min_field_area_m2: 2500         # m²
simplify_tolerance: 2.0         # metres
lulc_filter: true
lulc_crop_threshold: 0.3
lulc_on_error: raise            # raise | warn

engine_params:
  da_model: large_v2
  conf_threshold: 0.15          # the old name 'confidence' now raises

output_path: fields.gpkg
seed: 42

agribound delineate --config config.yml                      # YAML supplies every value
agribound delineate --config config.yml --year 2023 -o fields_2023.gpkg   # explicit flags override it
agribound delineate --config config.yml --dry-run            # print the resolved YAML

Reference: Configuration, CLI.

Reproducibility

Seed: seed (default 42) seeds Python, NumPy, torch and Lightning; the fine-tuning split and every sample are derived from it.
Cache keys: intermediates are named by a key over the study-area geometry, source, year, date range and compositing settings, so runs with different settings never reuse each other's files, even in one cache_dir.
Provenance: <output>.provenance.json holds the configuration and its hash, versions, platform, device, step timings, peak memory, engine metadata (for example weight revisions and SHA-256), stage counts and warnings.
Output reuse: an existing output is returned only if its provenance record reports success with the same configuration hash, study-area fingerprint and results versions; otherwise delineate() raises FileExistsError (overwrite=True re-runs). A record of agribound 1.0.0 or earlier has no study-area fingerprint; when nothing else differs, its output is still reused.

See Reproducibility.

Evaluation

from agribound.evaluate import evaluate

metrics = evaluate(
    pred_gdf,
    ref_gdf,
    iou_threshold=0.5,
    boundary_tolerance_m=10,
    strata="county",
    size_bins="auto",
    bootstrap=1000,
)

One-to-one IoU matching (default; matching="many_to_one" uses the 0.1.x matching rule, although geometry repair and other 1.0.0 changes can still change the numbers slightly) gives precision, recall and F1; also mean IoU, area-weighted precision/recall, over- and under-segmentation (Persello & Bruzzone, 2010), Hausdorff and mean boundary distances, boundary precision/recall/F1 and coverage within a tolerance, per-stratum and per-size-class metrics, and percentile bootstrap intervals. agribound evaluate -p pred.gpkg -r ref.gpkg does the same from the command line. Definitions: Evaluation.

Large Areas and HPC

agribound tiles cuts a region into tiles with halos and runs each tile as an independent, idempotent job; the composite stage can run on nodes with internet access and the delineation stage on offline GPU nodes; tiles merge keeps each polygon in the tile that owns its representative point. examples/hpc/ has Slurm scripts, profiles for NSF ACCESS systems and Earth Engine throttling rules; examples/regions/ defines 16 regions. The region files name no Earth Engine project: pass your own with --gee-project (or GEE_PROJECT, gcloud, or a service-account key); the scripts check it with agribound tiles gee-project before they run or submit anything. See HPC.

Agent Layer (optional)

pip install "agribound[agent]" adds a planning assistant with a deliberately low level of autonomy: a language model investigates with read-only tools and proposes one configuration; a human confirms the exact plan (typed yes, bound to the plan's SHA-256 hash, single use); at most one approved plan runs; and the session stops after the run or a denial. There is no automatic re-run or re-tuning. Every turn and tool call is written to a JSON transcript.

import agribound

result = agribound.agent(
    "Delineate fields in this AOI for 2024 with a label-free approach",
    study_area="fields.geojson",
    gee_project="my-gee-project",
    dry_run=True,  # propose only; run the plan YAML with `agribound delineate --config`
)
print(result.report)

agribound agent "Delineate fields in this AOI for 2024" --study-area fields.geojson --dry-run
agribound mcp serve                   # the same tools for MCP hosts (Claude Desktop, Claude Code, ...)
agribound mcp serve --allow-execute   # also execute_plan, confirmed by an MCP elicitation
claude mcp add agribound -- /path/to/env/bin/agribound mcp serve     # Claude Code

The default model is claude-opus-5 (--model or AGRIBOUND_AGENT_MODEL changes it); --base-url points the backend at an Anthropic-compatible local endpoint such as Ollama or vLLM (not tested with agribound). The streamable-http MCP transport has no authentication, so it is refused with --allow-execute or a non-loopback --host unless --allow-unauthenticated-http is given. See Agent layer.

Query Published FTW Polygons

query_ftw retrieves the already-published Fields of The World polygons for an area of interest (it does not run inference; the polygons are model predictions, not ground truth). The default layout (by-admin-conf) holds 2024 and 2025 with a confidence column, which is null for all of New Mexico and 99.7 % of New South Wales in the published files, so min_confidence cannot select reliable polygons there.

import agribound as ab

ftw = ab.query_ftw(
    study_area="bbox:-106.80,34.60,-106.75,34.65",
    year=2024,
    clip=True,
    output_path="ftw_2024.parquet",
)

agribound query-ftw --study-area "bbox:-106.80,34.60,-106.75,34.65" --year 2024 -o ftw_2024.parquet

With clip=True (the default), polygons that cross the AOI boundary are clipped, their metrics:area and metrics:perimeter are recomputed from the clipped geometry, and the column agribound:clipped marks them; clip=False returns the whole published polygons with their published metrics. With an output path, the query parameters, backend and counts are also written to <output>.provenance.json.

See FTW polygon query.

Project Structure

agribound/
├── agribound/                  # Main package
│   ├── __init__.py             # Public API (delineate, build_composite, evaluate, list_*, query_ftw, agent)
│   ├── _version.py             # Version string
│   ├── _cache.py               # Content-addressed cache keys
│   ├── _repro.py               # Seeding, seeded generators, version capture, run IDs
│   ├── auth.py                 # Earth Engine authentication
│   ├── cli.py                  # Click CLI (delineate, composite, prefetch, evaluate, ...)
│   ├── config.py               # AgriboundConfig dataclass and validation
│   ├── evaluate.py             # Object-level evaluation
│   ├── ftw_arrow.py            # PyArrow reader of the published FTW polygons
│   ├── ftw_query.py            # query_ftw()
│   ├── pipeline.py             # delineate() and build_composite()
│   ├── provenance.py           # Provenance records and configuration hash
│   ├── registry.py             # Source and engine registries
│   ├── visualize.py            # Interactive maps (leafmap/folium)
│   ├── agent/                  # Optional agent layer: tools, plans, gate, session, MCP server, backends
│   ├── clients/                # USGS NAIP Plus ImageServer client
│   ├── composites/             # Composite builders (base, gee, usgs, local + embeddings, dynamic_world)
│   ├── engines/                # Engines, SAM refinement (samgeo_engine), finetune/ package
│   ├── hpc/                    # Tiling, regions, `agribound tiles`
│   ├── io/                     # Raster, vector and CRS helpers
│   └── postprocess/            # Polygonize, merge, filter, simplify/smooth, regularize, LULC filter
├── assets/                     # Figures linked by URL from the README and docs (see assets/README.md):
│   ├── gallery_1.0/            #   rendered from the 1.0.0 and 1.0.1 example runs (tools/make_gallery*.py); WebP previews in preview/
│   ├── gallery_0.1x/           #   archived 0.1.x screenshots
│   └── agribound_workflow_1.0.*  # workflow diagram (tools/make_workflow_diagram.py)
├── docs/                       # MkDocs documentation (user guide, API reference, gallery, blog)
├── examples/                   # Example scripts 01-22, notebooks/ (generated from the scripts), hpc/, regions/
├── paper/                      # Manuscript materials (not included in the PyPI distribution)
├── tests/                      # Pytest suite (unit/, integration/)
├── tools/                      # Maintainer scripts: make_gallery.py, make_gallery_pampas_0.1x.py, make_workflow_diagram.py, sync_notebooks.py
├── CHANGELOG.md
├── CITATION.cff                # Citation metadata
├── CONTRIBUTING.md             # Developer guide
├── DISCLAIMER.md               # Software disclaimer
├── LICENSE                     # Apache 2.0
├── MANIFEST.in                 # Source distribution inclusions/exclusions
├── environment.yml             # Core conda environment
├── environment-gfm.yml         # GFM (Prithvi) conda environment
├── mkdocs.yml                  # MkDocs site configuration
├── pyproject.toml              # Build configuration, dependencies, extras
└── README.md

Examples

Example scripts and notebooks are in examples/; see the examples README.

Script	Notebook	Description
01_new_mexico_landsat_timeseries.py	notebook	Landsat time series over New Mexico with a Delineate-Anything model fine-tuned on NMOSE polygons
02_india_ganges_sentinel2.py	notebook	Four label-free approaches (FTW, Google and TESSERA embeddings, SPOT pan with Delineate-Anything) in Nadia, West Bengal
03_australia_murray_darling_hls.py	notebook	Prithvi `embed` and `pca` modes on HLS in the Murray-Darling Basin, compared with Delineate-Anything v2 on SPOT
04_france_beauce_sentinel2.py	notebook	FTW with crop-calendar season windows in the Beauce
05_pampas_embeddings.py	notebook	CPU-only embedding clustering (Google, TESSERA) in the Pampas
06_kenya_smallholder_ftw.py	notebook	FTW on smallholder fields in Kakamega with four minimum-area thresholds
07_usa_naip_high_res.py	notebook	Delineate-Anything on 1 m NAIP in the Central Valley
08_china_north_plain_spot.py	notebook	Delineate-Anything on SPOT 6/7 in the North China Plain (restricted source)
09_ensemble_comparison.py	notebook	Intersection, union and vote ensembles of Delineate-Anything and FTW (Andalusia)
10_local_tif_quickstart.py	notebook	Local GeoTIFF without Earth Engine
11_mississippi_alluvial_plain_spot.py	notebook	SPOT 6/7 series 2021-2023 with year-to-year agreement (restricted source)
12_new_mexico_ensemble_timeseries.py	notebook	Per-source multi-model ensembles with SAM 2, eastern Lea County, 2022
13_sam2_refine_dinov3.py	notebook	Stand-alone SAM refinement of existing DINOv3 boundaries
14_dinov3_sam2_ensemble.py	notebook	Fine-tuned DINOv3 with and without SAM 2 on five sources
15_pampas_semi_supervised.py	notebook	Label-free chain: embeddings → LULC filter → SAM 2, compared with Delineate-Anything v2 on Sentinel-2 and SPOT
16_usa_usgs_naip_plus.py	notebook	USGS NAIP Plus without Earth Engine (contributed by Jeremy Rapp)
17_query_published_ftw_polygons.py	notebook	Offline `query_ftw` demo with a local tile store (contributed by Jeremy Rapp)
18_agent_orchestration.py	notebook	Agent tools and a dry-run plan with the human confirmation gate
19_hpc_tiling.py	notebook	Tiling, two-phase runs and merging with `agribound.hpc`
20_stratified_evaluation.py	notebook	Stratified, size-class and boundary evaluation against NMOSE with bootstrap intervals; overall object and boundary metrics on Landsat, Sentinel-2, SPOT and NAIP of 2018 and with and without the crop filter
21_published_ftw_audit.py	notebook	Published FTW polygons evaluated against NMOSE
22_global_south_spot_pan.py	notebook	Delineate-Anything v2 on SPOT 6/7 panchromatic (1.5 m, restricted) in six farming landscapes of the Global South

Google Earth Engine Authentication

Earth Engine is needed for the Landsat, Sentinel-2, HLS, NAIP and SPOT composites, for Google embeddings with the default backend, for GEE-asset study areas, and for the LULC crop filter (on by default, for every source). local, usgs-naip-plus and tessera-embedding runs need no Earth Engine only with lulc_filter=False.

earthengine authenticate                       # once
agribound auth --project YOUR_GEE_PROJECT      # check

Credentials are tried in this order: gee_service_account_key (--gee-service-account-key), AGRIBOUND_GEE_SERVICE_ACCOUNT_KEY, stored earthengine authenticate credentials, Application Default Credentials (GOOGLE_APPLICATION_CREDENTIALS). Batch jobs (Slurm, no TTY) never fall back to an interactive browser prompt; they raise with instructions. See GEE setup.

SPOT Access

The SPOT 6/7 collection in Google Earth Engine (AIRBUS/SPOT6_7) is restricted and not publicly available; access is limited to select Earth Engine users. In agribound this source is for internal use at the Desert Research Institute (DRI). External users who need SPOT-based field boundaries can contact the package author to request processing.

Apple Silicon (MPS)

Observed with torch 2.10 on Apple MPS during the 1.0.0 checks: Delineate-Anything (FP16), FTW, DINOv3 and Prithvi embed mode ran on MPS. GeoAI's Mask R-CNN always runs on CPU (on MPS it reported Metal command-buffer errors and its detections differed from CPU). Prithvi + UPerNet runs on MPS only for compatible sizes (for example 192 px chips and tiles) and otherwise falls back to CPU with a WARNING. SAM masks differ between MPS and CPU. The Meta SAM 3 backend needs CUDA; use sam_backend="sam3-hf" on macOS (both SAM 3 backends are untested). Scripts that run FTW need an if __name__ == "__main__": guard (spawned data-loader workers).

Citation

If you use agribound in your research, please cite:

Majumdar, S., Rapp, J., Huntington, J. L., ReVelle, P., Nozari, S., Smith, R. G., Hasan, M. F., Bromley, M., Atkin, J., Jensen, E. R., Ketchum, D., & Roy, S. (2026). Agribound: Unified agricultural field boundary delineation from satellite imagery using geospatial foundation models, pre-trained segmentation, and embeddings [Software]. Zenodo. https://doi.org/10.5281/zenodo.19229665

Majumdar, S., Rapp, J., Huntington, J. L., ReVelle, P., Nozari, S., Smith, R. G., Hasan, M. F., Bromley, M., Atkin, J., Jensen, E. R., Ketchum, D., & Roy, S. (2026). Measuring what geospatial AI delivers for policy-grade agricultural field boundaries. In prep. for Remote Sensing of Environment.

Please also cite the underlying engines, models and data as appropriate:

Delineate-Anything: Lavreniuk, M., Kussul, N., Shelestov, A., Yailymov, B., Salii, Y., Kuzin, V., & Szantoi, Z. (2025). Delineate Anything: Resolution-agnostic field boundary delineation on satellite imagery. European Conference on Artificial Intelligence (ECAI 2025). arXiv:2504.02534. https://doi.org/10.48550/arXiv.2504.02534
Delineate-Anything v2 (default model): Lavreniuk, M., Kussul, N., Shelestov, A., Salii, Y., Kuzin, V., Wang, C. J. L.-X., & Szantoi, Z. (2026). Delineate Anything v2: A global foundation model for field delineation. European Conference on Computer Vision Workshops (ECCVW 2026), GAIA workshop. arXiv:2607.19069. https://doi.org/10.48550/arXiv.2607.19069
Fields of The World (FTW): Kerner, H., Chaudhari, S., Ghosh, A., Robinson, C., Ahmad, A., Choi, E., Jacobs, N., Holmes, C., Mohr, M., Dodhia, R., Lavista Ferres, J. M., & Marcus, J. (2025). Fields of The World: A machine learning benchmark dataset for global agricultural field boundary segmentation. Proceedings of the AAAI Conference on Artificial Intelligence, 39(27), 28151–28159. https://doi.org/10.1609/aaai.v39i27.35034
FTW PRUE models (default FTW model): Muhawenayo, G., Robinson, C., Khanal, S., Fang, Z., Corley, I., Wollam, A., Gao, T., Strnad, L., Avery, R., Estes, L., Tárano, A. M., Jacobs, N., & Kerner, H. (2026). PRUE: A practical recipe for field boundary segmentation at scale. arXiv:2603.27101. https://doi.org/10.48550/arXiv.2603.27101
Published FTW polygons: Robinson, C., et al. (2026). The first global agricultural field boundary map at 10m resolution. arXiv:2605.11055 (preprint). https://doi.org/10.48550/arXiv.2605.11055
GeoAI: Wu, Q. (2026). GeoAI: A Python package for integrating artificial intelligence with geospatial data analysis and visualization. Journal of Open Source Software, 11(118), 9605. https://doi.org/10.21105/joss.09605
DINOv3: Siméoni, O., Vo, H. V., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., Khalidov, V., Szafraniec, M., Yi, S., Ramamonjisoa, M., Massa, F., Haziza, D., Wehrstedt, L., Wang, J., Darcet, T., Moutakanni, T., Sentana, L., Roberts, C., Vedaldi, A., ... Bojanowski, P. (2025). DINOv3. arXiv:2508.10104. https://doi.org/10.48550/arXiv.2508.10104
Prithvi-EO-2.0: Szwarcman, D., Roy, S., Fraccaro, P., et al. (2026). Prithvi-EO-2.0: A versatile multitemporal foundation model for Earth observation applications. IEEE Transactions on Geoscience and Remote Sensing, 64, 1–20. https://doi.org/10.1109/TGRS.2025.3642610
TerraTorch: Gomes, C., Blumenstiel, B., de Sousa Almeida, J. L., et al. (2025). TerraTorch: The geospatial foundation models toolkit. IGARSS 2025, 6364–6368. https://doi.org/10.1109/IGARSS55030.2025.11243570
TESSERA: Feng, Z., Atzberger, C., Jaffer, S., Knezevic, J., Sormunen, S., Young, R., Lisaius, M. C., Immitzer, M., Jackson, T., Ball, J., Coomes, D. A., Madhavapeddy, A., Blake, A., & Keshav, S. (2026). TESSERA: Temporal embeddings of surface spectra for Earth representation and analysis. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 34818–34831. arXiv:2506.20380
Google Satellite Embeddings (AlphaEarth): Brown, C. F., Kazmierski, M. R., Pasquarella, V. J., Rucklidge, W. J., Samsikova, M., Zhang, C., Shelhamer, E., Lahera, E., Wiles, O., Ilyushchenko, S., Gorelick, N., Zhang, L. L., Alj, S., Schechter, E., Askay, S., Guinan, O., Moore, R., Boukouvalas, A., & Kohli, P. (2025). AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data. arXiv:2507.22291. https://doi.org/10.48550/arXiv.2507.22291
SAM 2: Ravi, N., Gabeur, V., Hu, Y.-T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C., Gustafson, L., Mintun, E., Pan, J., Alwala, K. V., Carion, N., Wu, C.-Y., Girshick, R., Dollár, P., & Feichtenhofer, C. (2025). SAM 2: Segment anything in images and videos. ICLR 2025. arXiv:2408.00714
SAM 3: Carion, N., Gustafson, L., Hu, Y.-T., et al. (2026). SAM 3: Segment anything with concepts. ICLR 2026. arXiv:2511.16719
SamGeo: Wu, Q., & Osco, L. P. (2023). samgeo: A Python package for segmenting geospatial data with the Segment Anything Model (SAM). Journal of Open Source Software, 8(89), 5663. https://doi.org/10.21105/joss.05663
SAM for Remote Sensing: Osco, L. P., Wu, Q., de Lemos, E. L., Gonçalves, W. N., Ramos, A. P. M., Li, J., & Marcato Junior, J. (2023). The Segment Anything Model (SAM) for remote sensing applications: From zero to one shot. International Journal of Applied Earth Observation and Geoinformation, 124, 103540. https://doi.org/10.1016/j.jag.2023.103540
geemap: Wu, Q. (2020). geemap: A Python package for interactive mapping with Google Earth Engine. Journal of Open Source Software, 5(51), 2305. https://doi.org/10.21105/joss.02305
Google Earth Engine: Gorelick, N., Hancher, M., Dixon, M., Ilyushchenko, S., Thau, D., & Moore, R. (2017). Google Earth Engine: Planetary-scale geospatial analysis for everyone. Remote Sensing of Environment, 202, 18–27. https://doi.org/10.1016/j.rse.2017.06.031
Awesome GEE Community Catalog: Roy, S., Majumdar, S., & Swetnam, T. (2025). samapriya/awesome-gee-community-datasets: Community Catalog (3.9.0). Zenodo. https://doi.org/10.5281/zenodo.17641528

More references (HLS, Dynamic World, NLCD, C3S, Cloud Score+, evaluation methods, ACCESS) are listed in the documentation.

License

This project is licensed under the Apache License 2.0. The Delineate-Anything model code and weights, and Ultralytics, are AGPL-3.0. The DINOv3 weights used by the dinov3 engine are under the DINOv3 License (custom, not OSI-approved), whose clause 1.b.ii requires publications to acknowledge the use of the DINO Materials; the SAM 3 licence has a similar clause. Check the licences of the models and datasets you use.

Acknowledgments

Agribound builds on the work of many open-source projects and research teams:

The Ultralytics team for the YOLO ecosystem
Mykola Lavreniuk and co-authors for Delineate-Anything
Meta AI Research for the Segment Anything models and DINOv3
The Fields of The World consortium and Hannah Kerner's group at Arizona State University
Qiusheng Wu for the GeoAI and samgeo Python packages
NASA and IBM Research for the Prithvi geospatial foundation model and TerraTorch
Google DeepMind for AlphaEarth satellite embeddings
Feng et al. for the TESSERA foundation model embeddings
The Google Earth Engine team for planetary-scale geospatial computing
The fiboa community for the field boundary schema standard
The TorchGeo team for geospatial deep learning data loaders and utilities
The Desert Research Institute (DRI) for supporting this research

Funding

This work was supported by multiple funding sources. The New Mexico Office of the State Engineer (NMOSE) provided reference field boundary data and supported the development of agricultural water use mapping in New Mexico. The Google Satellite Embeddings Dataset Small Grants Program enabled the integration of pre-computed satellite embeddings for unsupervised field boundary delineation. Access to the SPOT 6 and 7 archive on Google Earth Engine was provided through the Google Trusted Tester opportunity. Additional support was provided by the U.S. Army Corps of Engineers and The U.S. Department of Treasury/State of Nevada. This work was also supported by the NASA Water Resources Applications Program, the United States Geological Survey (USGS) and NASA Landsat Science Team, the USGS Water Resources Research Institute, the Desert Research Institute Maki Endowment, and the Windward Fund.

AI Usage Disclosure

Portions of this software were developed with the assistance of AI coding tools, including Anthropic's Claude. AI was used to accelerate code scaffolding, documentation drafting, and test generation. All AI-generated code was reviewed, tested, and validated by the human authors. The scientific methodology, architectural decisions, algorithm selection, and domain-specific implementations reflect the expertise and judgment of the authors.

Navigation Menu

Repository files navigation

agribound

Overview

How It Works

Results

From 30 m to 1 m: Delineate-Anything v2 against a reference registry (San Juan County, New Mexico, USA)

Supervised: DINOv3 fine-tuned + SAM 2 on four sources (eastern Lea County, New Mexico, USA)

Label-free: embeddings + SAM 2 vs Delineate-Anything v2 (Pampas, Argentina)

Satellite Sources

Delineation Engines

Installation

Quick Start (Python)

Quick Start (CLI)

Configuration

Reproducibility

Evaluation

Large Areas and HPC

Agent Layer (optional)

Query Published FTW Polygons

Project Structure

Examples

Google Earth Engine Authentication

SPOT Access

Apple Silicon (MPS)

Citation

License

Acknowledgments

Funding

AI Usage Disclosure

About

Releases

Packages

Contributors

Languages

Name		Name	Last commit message	Last commit date
.github/workflows		.github/workflows
agribound		agribound
assets		assets
docs		docs
examples		examples
tests		tests
tools		tools
.gitattributes		.gitattributes
.gitignore		.gitignore
.lycheeignore		.lycheeignore
CHANGELOG.md		CHANGELOG.md
CITATION.cff		CITATION.cff
CONTRIBUTING.md		CONTRIBUTING.md
DISCLAIMER.md		DISCLAIMER.md
LICENSE		LICENSE
MANIFEST.in		MANIFEST.in
README.md		README.md
environment-gfm.yml		environment-gfm.yml
environment.yml		environment.yml
mkdocs.yml		mkdocs.yml
pyproject.toml		pyproject.toml

Latest commit

History

Folders and files

Repository files navigation

agribound

Overview

How It Works

Results

From 30 m to 1 m: Delineate-Anything v2 against a reference registry (San Juan County, New Mexico, USA)

Supervised: DINOv3 fine-tuned + SAM 2 on four sources (eastern Lea County, New Mexico, USA)

Label-free: embeddings + SAM 2 vs Delineate-Anything v2 (Pampas, Argentina)

Satellite Sources

Delineation Engines

Installation

Quick Start (Python)

Quick Start (CLI)

Configuration

Reproducibility

Evaluation

Large Areas and HPC

Agent Layer (optional)

Query Published FTW Polygons

Project Structure

Examples

Google Earth Engine Authentication

SPOT Access

Apple Silicon (MPS)

Citation

License

Acknowledgments

Funding

AI Usage Disclosure

About

Topics

Resources

Contributing

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages