Vassilis Sitokonstantinou, Mariza Kaskara, Iason Tsardanidis, Thanassis Drivas, Alexandros Marantos, Alkis Koukos, Haris Kontoes AgriHUB | Agriculture, Ecosystems and Environment Group National Observatory of Athens

1. Cultivated Crop Type Maps
25/10/2022
Vassilis Sitokonstantinou, Mariza Kaskara, Iason Tsardanidis,
Thanassis Drivas, Alexandros Marantos, Alkis Koukos, Haris Kontoes
AgriHUB | Agriculture, Ecosystems and Environment Group
National Observatory of Athens
Institute for Astronomy, Astrophysics, Space Applications & Remote Sensing
BEYOND Center, Athens, Greece

2. Multi-temporal Crop Type Maps
• Well-tested Machine Learning and Deep Learning Models Applied
• Dynamic Crop Type Mapping for every cultivation year
• 2 Pilot cases
▪ Lithuania (NPA) ~ 3 millions hectares of agricultural land
▪ Cyprus (CAPO) ~ 0.2 hec average parcel size
• National scale results
• Parcel-based or Pixel-based Approaches
• For every evaluated field S1/S2 band and indices time-series were calculated
using the LPIS buffered geometries
– Sentinel-2 L2A Spectral bands (B01-B12)
• Scene Classification (SCL)
• Vegetation Indices - VIs (NDVI, NDWI, PSRI)
– Sentinel-1 GRD
• Backscattering coefficients (VV-VH)
Multi temporal Crop type mapping
Cultivated crop type maps

3. EO services
• Smart Sampling for OTSC (traffic light system): sophisticated algorithm evolving
dynamically throughout the cultivation period by exploiting the current and the
previously generated Crop Type Maps, to identify the most confident misclassifications
and potentially false declarations.
[1] M. Rousi et al., "Semantically Enriched Crop Type Classification and Linked Earth Observation Data to Support the Common
Agricultural Policy Monitoring," in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 14, pp. 529-
552, 2021, doi: 10.1109/JSTARS.2020.3038152
• Crops Diversification Compliancy Maps (Greening I): a compilation of if-conditions
according to the Greening 1 set of rules which examines the hypothetical impacts
between an actual truth and crop label mapped. Exploits LPIS and the declarations of the
farmers.
Greening I Compliance Map
Precision and Recall of smart sampling algorithmover
cultivation period
Cultivated crop type maps
82%
13%
5% Compliancy
Compliant
Non-Compliant
Not Assessed

4. Case of Lithuania

5. Dynamic Crop Type Mapping
Random Forest
• 14 Sentinel-2 Tiles
• 4 Sentinel-1 Relative Orbits
Case of Lithuania

6. Example 1. NDVI of a case predicted as Spring Cereal and the label
given is Black Fallow
Example 2. NDVI of a case predicted as Winter Cereal and the label
given is Spring Cereal
Wrong Declarations Percentage < 5%
Case of Lithuania

7. Case of Lithuania

8. Case of Lithuania

9. Case of Cyprus

10. Fallow Land Case
Case of Cyprus
Wrong Declarations Percentage > 10%

11. Case of Cyprus
Grasslands
Vineyards
Citrus Fruit Trees
Fallows
Durum Wheat
Barley
Triticale

12. a) Small Parcels Issue → Average size < 0.3 hec (mixels issues)
b) Multiple Cultivations – Polycultures
c) Intense ground vegetation on permanent cultivations
Case of Cyprus
a)
b)
c)

13. Greenhouses (Temporal vs Permanent)
Case of Cyprus

14. From Pixel to Parcel
Case of Cyprus
Dynamic Crop Type Mapping
• 2 Sentinel-2 Tiles
• 2 Sentinel-1 Relative Orbits

15. Case of Cyprus
Custom ML routines in order to
take advantage of the different
hierarchical level information
Support Vector Machine

16. Case of Cyprus
Producer
Accuracy
User
Accuracy

17. Case of Cyprus
Future work
• Evaluate more sophisticated DL approaches in order to exploit the spatial distribution
• Homogeneous vs heterogeneous fields – Fields boundaries delineation
• Train on previous years validated cases
• Continuous communication with CAPO in order to hear about novel problems and find more convenient
solutions

18. Remarks
• Parcel-wise and pixel-wise ML routines that can create dense time-series integrating both Synthetic Aperture Radar
(Sentinel 1) and the cloud free Sentinel 2 acquisitions
• Each pilot case has its own individual particularities → Widely applicable solutions
– Frequent cloud coverage (Lithuania) → Parallel usage of Sentinel-1 and Sentinel-2 imagery
– Small narrow-shaped parcels (Cyprus) →Pixel-based level of inference
• When there is nothing more we can do with Sentinel data we invest on algorithms that work on top of the
classification output, such as our improved smart sampling algorithm to decide on the risk of false declaration even if
the classification accuracy is suboptimal.
• A traffic light system that assist on the planning of campaigns and reduction of their cost
• Disputes arrangement with the farmers using early classification results as warnings
• Greening I compliance map for crops diversification