AGRS – Agricultural Remote Sensing Library

Description

AGRS is a domain-focused Python library that turns Sentinel-2 imagery into agronomy-ready features for tasks such as yield modeling, stress analysis, and site-specific fertilizer (NPK) recommendation.

Instead of re-implementing STAC queries, cloud filters, band math, and geometry clipping in every project, AGRS provides opinionated, agriculture-centric pipelines:

• Search & fetch Sentinel-2 L2A scenes (via Microsoft Planetary Computer STAC).
• Compute a comprehensive set of classic vegetation, water, moisture, and burn indices.
• Select snapshots using agronomic strategies (fractions of season, specific dates, all scenes, best cloud-free scenes).
• Aggregate to field-level features per index and time-fraction of the season.
• Optionally return per-band statistics (mean, median, percentiles, …) for QA and custom modeling.
• Return tidy pandas.DataFrame objects ready to join with yield, NPK, and management tables.

The goal is to maximize value for agricultural ML workflows: field trial analysis, site-specific NPK optimization, and crop monitoring.

Key features

• Sentinel-2 access

  • Microsoft Planetary Computer STAC API.
  • Automatic bounding-box search from field geometries (GeoDataFrame).

• Band handling

  • Automatic band detection (default: all B* assets) or user-defined band list.
  • Robust geometry clipping for each field.

• Built-in index formulas (Sentinel-2 friendly)
  (only computed if required bands are present)

  • Vegetation / chlorophyll:
    • NDVI, EVI, EVI2, SAVI, MSAVI, OSAVI
    • GNDVI, VARI, GCI
    • NDRE, RECI, ARVI
    • MCARI, MCARI2
  • Water / moisture / burn:
    • NDWI (McFeeters), MNDWI
    • NDMI (a.k.a moisture index)
    • NBR, NBR2
  • Safe division with NaNs for invalid cases (no hard crashes on zero / missing).

• Snapshot selection strategies

  • fractional – nearest scenes to temporal fractions (e.g. [0.3, 0.7] of the season).
  • fixed_date – snapshot closest to a single target date.
  • dates – snapshots closest to a list of key dates.
  • top_n_cloudfree – N scenes with lowest eo:cloud_cover.
  • all – use all scenes between start_date and end_date.

• Return modes

  • return_mode="features" (default):
    • Field-level aggregated index features (per field, per index, per time-fraction).
  • return_mode="bands":
    • Per-field, per-date statistics for each band:
      • mean, median, min, max, std, p10, p90.
  • return_mode="both":
    • Tuple (features_df, bands_df) with both outputs.

• Field-level aggregation

  • For each field and index:
    • Aggregation over the selected snapshots with statistics such as mean, median, min, max, std.
  • Uses time fractions to encode position within the season (0–1).

Installation

pip install agrs

AGRS requires Python 3.9+ and depends on geopandas, shapely, rasterio, pystac-client, and planetary-computer.

Quick start (features only)

import geopandas as gpd
from shapely.geometry import Point
from agrs.client import s2agc

# 1) Single test field
field_geom = Point(-8.0, 32.0)

fields = gpd.GeoDataFrame(
    {"field_id": [1], "geometry": [field_geom]},
    crs="EPSG:4326",
)

# 2) Create client
client = s2agc(
    source="planetary_computer",
    max_cloud=0.2,
)

# 3) Get field-level index features with fractional snapshots (e.g. 30%, 60%, 90%, 100% of season)
features_df = client.get_features(
    fields=fields,
    field_id_col="field_id",
    start_date="2018-10-01",
    end_date="2019-06-30",
    crop="wheat",
    snapshot_strategy="fractional",
    fractions=[0.3, 0.6, 0.9, 1.0],
    return_mode="features",  # default
)

print(features_df.head())

Examples of other behaviours

1. Custom fractional snapshots (e.g. 30% and 70% of season)

features_df = client.get_features(
    fields=fields,
    field_id_col="field_id",
    start_date="2018-10-01",
    end_date="2019-06-30",
    crop="wheat",
    snapshot_strategy="fractional",
    fractions=[0.3, 0.7],
    return_mode="features",
)

2. Bands only (all snapshots between two dates)

bands_df = client.get_features(
    fields=fields,
    field_id_col="field_id",
    start_date="2018-10-01",
    end_date="2019-06-30",
    crop="wheat",
    snapshot_strategy="all",   # use all available scenes
    return_mode="bands",       # per-band stats per date
)

print(bands_df.head())

3. Features + bands together

features_df, bands_df = client.get_features(
    fields=fields,
    field_id_col="field_id",
    start_date="2018-10-01",
    end_date="2019-06-30",
    crop="wheat",
    snapshot_strategy="dates",
    key_dates=["2019-02-15", "2019-04-01"],
    return_mode="both",
)

More usage patterns (fixed-date, top-N cloud-free, multi-field examples) are available in the examples/ folder.

Contributing

Contributions, bug reports, and feature requests are very welcome.

Please see CONTRIBUTING.md for guidelines on how to:

• Set up a development environment
• Run tests and examples
• Propose new indices, snapshot strategies, or data sources
• Open issues and pull requests

License

AGRS is released under the MIT License. See LICENSE for details.