MowEO connects area selection, 5 km tiling, cloud preprocessing, annual NetCDF acquisition, local grassland mowing detection, quality assessment and geospatial export in one Python package. Version 0.3.0a1 is a prerelease. License selection is pending; see LICENSE_NOTICE.md.
Install Miniconda or Anaconda, extract the repository and open a terminal in
the directory containing pyproject.toml. On Windows, use Anaconda Prompt.
conda env create -f environment.yml
conda activate moweo
python -m pip install --no-deps .
moweo --versionActivate moweo in each new terminal. The first command downloads Python 3.11
and dependencies; the pip command installs MowEO and its command-line interface.
Alternatively, activate a Python 3.11 virtual environment and run
python -m pip install .. Dependencies are also listed in requirements.txt.
openEO is pinned to 0.52.0; other dependencies use version ranges, not a fully
locked environment. See testing below for the environment actually verified.
Create and verify your CDSE account. Check your openEO credit balance in the Algorithm Plaza Billing page before submission. Acquisition uses cloud processing credits; local detection does not. CDSE currently advertises 10,000 free monthly credits, but quotas and terms can change. Start with a small pilot before submitting a whole country.
These service references were checked on 2026-11-12. Registration does not guarantee enough credits for an entire region or year.
Supply a valid polygon AOI with a declared CRS. Use absolute input paths in these examples. Choose a suitable projected CRS in metres; EPSG:32633 is an example for the original study area, not a global default.
moweo --project my_project init-aoi --aoi /absolute/path/aoi.gpkg --epsg 32633 --year 2026 --limit 10
moweo --project my_project prepare
moweo --project my_project acquireinit-aoi --limit 10 selects at most ten tiles for both acquisition and
processing. Omit it to use the whole AOI. The last command is a dry run;
it submits no remote jobs. To start acquisition:
moweo --project my_project acquire --run --parallel 2Open the displayed device-login link, authenticate with the intended CDSE
account and confirm with YES. No grassland mask is needed for acquisition.
After downloading, add private polygons containing only grassland features:
moweo --project my_project masks --grasslands /absolute/path/grasslands.gpkg
moweo --project my_project validate --require-stats
moweo --project my_project process --workers 4Masks are rasterized on the fixed grid without changing tile IDs. They are not distributed with the code. See the input specification.
Select a country from a supplied boundary dataset:
moweo --project country_project init-aoi --aoi /absolute/path/countries.gpkg --country Slovenia --country-field NAME --epsg 32633 --year 2026Use the column and country name/code in your dataset. MowEO does not download
a boundary from a country name alone. Border tiles remain full squares;
subsequent masks exclude pixels outside the AOI. IDs such as
eo_32633_380000_5090000 encode EPSG, west and south coordinates and remain
stable across years in the same grid. See AOI and countries.
To create the grid and masks together from grasslands, use a fresh project:
moweo --project grassland_project init --grasslands /absolute/path/grasslands.gpkg --epsg 32633 --year 2026
moweo --project grassland_project prepareFor existing grids and aligned per-tile masks, use moweo --project my_project setup
and follow the input specification. Once the project and both masks are ready,
connect acquisition and local processing:
moweo --project my_project run --run --parallel 2 --workers 4The first run selects the workflow; --run enables remote submission.
Without --run, no jobs are submitted. Account confirmation still applies.
Use the separate stages when adding masks after acquisition. Installed wheels
use moweo or python -m moweo; numbered scripts and .bat files are legacy
shortcuts for source checkouts.
This release uses the delivery-reference v0.3 detector, including event-local
quality scoring, grow_r=1.3, SEIVE and confidence-aware harmonization. It
changes the detector and evidence score relative to MowEO 0.2.x. See
products, method differences and migration; existing 0.2.x
projects must be deliberately migrated and locally recalculated.
The collection is SENTINEL2_L2A on
https://openeo.dataspace.copernicus.eu; arbitrary collection switching is not
implemented. Cloud preprocessing applies SCL dilation masking and computes
NDVI from B04/B08 at 10 m, NDII from B8A/B11 at 20 m and SWIRMean from B11/B12
at 20 m. NetCDF downloads are cloud-masked index cubes, not raw Sentinel-2
SAFE products or reflectance-band archives.
Local detection identifies NDVI drops and regrowth. Spatial support and NDII/SWIR responses provide diagnostic evidence, not calibrated probabilities. The +5 m detection mask includes boundary pixels; the -7.5 m inner mask reduces boundary mixing in statistics and must be a same-grid subset. Both manifests are required for the full local pipeline, including empty-tile entries.
Settings are in local project_config.json. Defaults, fixed assumptions and
cloud settings are described in parameters and
project_config.example.json. Acquisition defaults
to March 15–October 31 and detection to April 1–October 31. Review these windows
for your region and the backend's temporal endpoint convention.
my_project/workspace/<YEAR>/
acquisition/
openeo_jobs_<YEAR>.csv
process_eopatch/<PATCH_ID>/
<PATCH_ID>_NDVI_<YEAR>.nc
<PATCH_ID>_NDII_SWIRMean_explicit20m_<YEAR>.nc
processing/
process_eopatch/<PATCH_ID>/ # baseline, final, diagnostics and event tables
national_qa_<YEAR>.csv
national_analysis_<YEAR>/ # summary tables and figures
national_maps/
AREA_MOWING_COUNT_<YEAR>.tif
AREA_MOWING_COUNT_SEIVE_<YEAR>.tif
AREA_MOWING_COUNT_SPATIAL_HARMONIZED_<YEAR>.tif
AREA_MOWING_DOY_1_<YEAR>.tif ... AREA_MOWING_DOY_6_<YEAR>.tif
AREA_MOWING_CONFIDENCE_<YEAR>.tif
AREA_VALID_OBSERVATIONS_<YEAR>.tif
AREA_SCL_MASKED_PERCENT_<YEAR>.tif
products_manifest.json
AREA_TILE_QA_<YEAR>.gpkg
provenance/ # settings, versions and process graphs
The count mosaic uses 0 for grass pixels with no detected mowing, 1–6 for
detected counts and 255 for NoData. Per-tile float rasters use -9999 as
NoData; band descriptions identify counts, day-of-year dates and diagnostics.
The twelve single-band raster products match the delivery inventory. DOY 1–6
are original date slots; SEIVE is a separate count derived from filtered dates.
Confidence uses rounded UInt8 0–100 with NoData 255 and is not a probability.
GeoPackage output contains tile-level QA summaries, not vectorized mowing
pixels or field boundaries. output_prefix controls the AREA prefix.
moweo --project my_project status
moweo --project my_project process --workers 1
moweo --project my_project process --dry-runUse fewer workers when RAM or storage becomes a bottleneck. Keep the computer awake during download and processing. Repeat acquisition to resume completed products and logged jobs. Interrupting the client does not cancel remote jobs. Do not run concurrent acquisition clients against the same workspace.
For another year on the same grid, change year in the configuration and
repeat acquisition and processing. Outputs are separated by year; review
whether the mask vintage is appropriate. Changing area, CRS, seasonal windows
or cloud processing requires a fresh workspace: existing products/jobs are
reused by tile and product, not a complete parameter fingerprint. Never change
settings during an active run. Relative input paths refer to the project folder.
python -m unittest discover -s tests -vThe tests create synthetic inputs and need no CDSE account or credits. They cover AOI/country grids, stable IDs, persistent tile limits, later mask addition, dry-run acquisition, detection, reports, GeoTIFF/CSV/GeoPackage outputs and rejection of incomplete mosaics. Temporary examples are removed afterward. This verifies software behavior, not scientific detection accuracy.
All three tests passed on Windows/Python 3.13.5 from source and an installed 0.3.0a1 wheel. A ten-tile real-data pilot for 2026 was also locally recalculated; its SEIVE, harmonization and rounded confidence outputs matched the reference export functions pixel-for-pixel. This is an implementation check, not an independent scientific accuracy assessment. See the release notes for details. The documented Python 3.11 Conda environment has not been executed here; CI targets Python 3.11 on Windows and Linux but has not yet run on GitHub. macOS has not been tested. See release notes for versions.
The CLI records settings, versions, backend and collection in provenance/.
Actual openEO process graphs are saved before remote submission. Preserve these
records and input vintages; backend/archive changes can still prevent exact
reproduction. Tokens remain in the private workspace/<YEAR>/.openeo_auth/.
Thresholds were calibrated for the source study, not validated worldwide. Clouds, drought, grazing, senescence and observation gaps can affect detections. Zero detected events does not establish no mowing. Dates are time-series estimates, not exact observed operation times. Extensive countries require a suitable projection, sufficient credits and independent regional validation.
No project masks, reference observations, credentials, cubes or results are included. Public upload, license assignment, author/citation metadata and DOI archival remain separate release tasks. See release status.