Open source · Apache-2.0 · Sentinel-1 + Sentinel-2
An open, local-first engine for crop mapping and parcel verification. GeoCrop turns Copernicus time series into crop maps with probabilities, and those into parcel decisions that state their evidence.
quicklook of Semberija, August 2023 composite. Contains modified Copernicus Sentinel data 2023.Web workspace
A local workspace served by the engine itself. Choose an area and a period and look at real Sentinel-2 imagery, then go as deep as the command line allows.
The New analysis sheet takes an area by coordinates or drawn on the map, a period preset, and shows the exact command before it runs quicklook: monthly cloud-screened composites from windowed COG reads, as true colour, false colour or NDVI.
Live syntax colouring and completion of commands, flags, choices and workspace paths, with ghost-text hints, history search, streamed job output and Ctrl C to cancel.
Press ⌘ K or Ctrl K. Forms are generated from the real CLI parser, grouped, with a live command-line preview.
Follows the system appearance or your choice, respects reduced motion, and becomes a phone layout with a tab bar.
Actual captures of the running workspace; provenance and checksums in screenshots.json.
Validation rigour
Neighbouring pixels are not independent. GeoCrop evaluates with that in mind by default, and every number traces back to the split that produced it.
Class-stratified spatial blocks, optional buffers that drop training polygons near held-out pixels, leakage assertions and nested tuning.
Whole spatial blocks set aside before model selection and scored once. In holdout calibration mode the predictor stays frozen.
Leave-one-region-out, leave-one-year-out and region-year CV report per-group metrics and the gap to spatial-block CV.
Temperature-scaled probabilities and class-conditional conformal sets: an ambiguous parcel says which crops it could be.
Flags predictions in feature space the model never saw. Parcels below the AOA threshold go to review instead of being accepted.
Field points near training polygons are excluded automatically; bootstrap CIs, area-adjusted estimates and calibration error.
Public benchmark · protocol v1
Seven crop classes, Sentinel-2 L2A, a frozen protocol file and one command to reproduce. Parcel-level scores on a spatial-block holdout with 95 % block-bootstrap intervals.
A public transfer task on farmers’ declarations, not field-verified labels, and not Republic of Srpska data. Valid for this protocol only; not evidence of accuracy elsewhere.
| Model | OA % | Macro‑F1 % | Coverage % conformal, target 90 |
|---|---|---|---|
| rf | 86.9 [81.1, 92.2] | 85.5 [78.2, 91.7] | 91.7 |
| extra_trees | 88.1 [82.8, 93.1] | 86.3 [78.6, 92.6] | 92.9 |
| hist_gb | 87.8 [83.0, 92.5] | 86.1 [79.4, 91.6] | 89.1 |
| catboost | 88.5 [83.5, 93.2] | 87.2 [80.9, 92.7] | 92.0 |
EuroCrops labels: Schneider et al. (2023), CC-BY-SA-4.0. Full report in docs/benchmarks.
Parcel decisions
Pixel probabilities are aggregated inside an interior buffer, or parcels are classified directly from their signatures. Each parcel is accepted or sent to review with stable reason codes.
low_confidence low_margin strategies_disagree insufficient_valid_pixels aoa_below_thresholdconformal_discrepancy declaration_discrepancy ambiguous_prediction_set calendar_inconsistentno_harvest_evidence tillage_detected insufficient_mowing_events heterogeneous_parcelA selection of the documented codes (parcels, markers). Semicolon-separated and not mutually exclusive.
Models and pretrained adapters
Every model and feature source runs behind the same spatial CV, ablation and prediction, so a new idea is compared on identical folds.
An optional feature group emb from Presto, computed locally on the CPU from GeoCrop’s own composites, or from AlphaEarth and TESSERA annual embeddings. Ablation adds handcrafted, embeddings-only and combined rows.
ftw-boundaries runs pretrained field-boundary models on two-window composites: CC-BY checkpoints by default, SHA-256-pinned download, watershed parcels, scored like OBIA and the U-Net.
Pick a diverse, well-spaced batch among the most uncertain parcels, grouped into trips. Ingest labels with per-round provenance, and simulate label budgets against random sampling.
Random Forest, ExtraTrees, HistGradientBoosting, LightGBM, CatBoost, TempCNN and L-TAE, plus parcel-level classification on signatures, on one benchmark harness.
Heavy dependencies and checkpoints are optional and fetched at runtime; nothing is bundled. Licences and attribution: embeddings, field boundaries, active learning.
Built for real data
Preflight
Ready, warn or blocked
Scenes per window and orbit, clear-sky coverage, class support under the spatial blocks, CRS overlap, and disk, RAM and runtime estimates before a long run starts.
Acceptance · Semberija 2023
21 of 21 checks
One versioned run of the whole chain on real Sentinel-2 and Sentinel-1 data over about 4 km²: radiometry, S1/S2 co-registration, reproducibility, COG, STAC and GeoParquet validity.
Reliability gate
Killed, then recovered
A real run is killed inside features, train and predict and must recover to the uninterrupted result. Partial outputs are quarantined; a disk guard stops before writing.
Report: acceptance-0.10-dev-2026-10-26. The acceptance checks cover the technical integrity of the chain on that area, not crop-map accuracy.
Open by architecture
A layered core — local logic, adapters for networks and external programs, application, interfaces — enforced by architecture tests. Extend it through plugin registries for classifiers, indices, segmenters, exporters and validators.
geocrop serve binds to loopback and requires a token on every API request. Jobs are queued, logged, cancellable and their outputs downloadable. Sentinel-2 cloudless and OpenStreetMap basemaps sit under the imagery.
A standalone plugin adds run-command, demo and load-job algorithms to the Processing Toolbox. It talks to the engine API, streams logs, reconnects to jobs and loads results as layers. QGIS keeps its own Python.
Apache-2.0 (licence, notice). Targets QGIS 3.28–3.x; tested in 3.40.6.
EngineClient submits the same commands, polls jobs, reads logs and downloads artifacts, without putting the token in the notebook. It uses /api/v1 and falls back on older servers.
# GEOCROP_WEB_TOKEN comes from the environment.
from geocrop.client import EngineClient
with EngineClient() as engine:
job = engine.submit(["indices", "--format", "json"])
engine.wait(job["id"], timeout=60)
print(engine.logs(job["id"])["text"])Every endpoint is served under /api/v1, the documented stable contract with an OpenAPI schema. Requests need a Bearer token; request bodies and job arguments are size-limited.
GET /api/v1/health
GET /api/v1/commands
POST /api/v1/jobs {"argv": ["indices", "--format", "json"]}
GET /api/v1/jobs/{id}/logs
POST /api/v1/jobs/{id}/cancel
GET /api/v1/jobs/{id}/artifactsGet started
Install, run the synthetic demo, then open the workspace. Python 3.10–3.12.
git clone https://github.com/ZeljkoDzafic/geocrop-engine.git
cd geocrop-engine
python3 -m venv .venv && source .venv/bin/activate
python -m pip install -e '.[all,web]'
geocrop demo --root outputs/my-first-demo --quick
# Workspace: first set GEOCROP_WEB_TOKEN (32+ random characters)
geocrop serve --workspace . --port 8765git clone https://github.com/ZeljkoDzafic/geocrop-engine.git
cd geocrop-engine
py -3.12 -m venv .venv
.\.venv\Scripts\python.exe -m pip install -e ".[all,web]"
.\.venv\Scripts\geocrop.exe demo --root outputs\my-first-demo --quick
# Workspace: first set $env:GEOCROP_WEB_TOKEN (32+ random characters)
.\.venv\Scripts\geocrop.exe serve --workspace . --port 8765