Boundaries detection using satellite imagery processing

Smart and scalable field boundary detection with EOS Data Analytics for accurate land monitoring and decision-making:

  • Up to 90% accuracy achieved through CNN model trained on multi-country datasets
  • Seamless integration with crop classification for detailed per-field analytics
  • Flexible output formats, including vector masks, statistics, and custom reports via API
Field boundaries detection by EOSDA

Approach and Methodology

Model: Semantic segmentation CNN (encoder-decoder technology) trained on a 10-country dataset

Satellite data: Sentinel-2 L2A images processed with 4 spectral bands, including RGB and NIR

Accuracy: up to 90% by IoU metric (Intersection over Union)

Limitations: Regions with high cloud cover and small fields (<2 ha). High resolution imagery can be used

Flexible integration: field boundary detection can be used as a single solution or in combination with crop classification

Enhanced precision: Field boundary detection results can be combined with a raster classification layer for more accurate, granular per-field classification

High-level CNN encoder-decoder model scheme (EOSDA)

Expected project outputs and formats

Field boundaries detection, Iowa, USA, 2024
  • Vector mask with field boundaries or along with crop classification, if applicable (ESRI shapefile, GeoJSON, KML, GPKG).
  • Aggregated statistics of cropland ares by admin boundaries of regions, districts, etc (xlsx, csv).
  • Analytical report or results interpretation note if requested.

*Delivery in another format if requested, including via API interface.

Required data for machine learning boundary detection project

Output data provided by customer

  • Area of interest
  • table (xlsx/ csv)
  • vector (KML, ESRI shapefile, GPKG, GeoJSON) format
  • Crop сalendar, if available (to choose optimal period for detection)

Data provided by EOSDA

  • Satellite imagery
  • Other additional data layers (cities, roads, water bodies, etc)
  • Validation datasets

Standard project stages

Typical project duration: 1-2 weeks

1

Investigation of vegetation features for AOI

2

Search and download of required satellite data

3

Model launch and support/control

4

Result verification and final outputs preparation

Benefits of field boundary detection solution by EOSDA

  • Accurate identification of land parcels for change monitoring.
  • Improved crop classification, yield prediction models, and field-level performance assessment.
  • Precise data for autonomous farm machinery.
  • Optimized planning for irrigation, drainage, and rural development.