MSc Thesis · My Proposal
Satellite Crop Classification
ConvLSTM + Conv3D on Sentinel-2 multispectral time-series to classify crops at pixel level — and automatically flag farmer declarations where the satellite evidence contradicts the filed crop type.
Self-proposed: identified the gap from ERP work, proposed the satellite solution, built the ML pipeline, and integrated outputs into the ERP for internal declaration-consistency review.
Why this matters
Greek farmers submit annual crop declarations to OPEKEPE to claim subsidies under the EU Common Agricultural Policy. Subsidy amounts depend on what crop is declared on each parcel. False or incorrect declarations — whether intentional or not — cost the Greek state millions annually.
The European Space Agency's Copernicus programme makes Sentinel-2 imagery freely available at 10-metre resolution every 5 days. Every growing season leaves a distinct spectral signature in the time-series. A well-trained classifier can read that signature and tell you: wheat, cotton, maize, or fallow.
My thesis built that classifier and connected it to a declaration-consistency check.
Results
94.2%
Overall Accuracy
on held-out Greek test patches
0.96
F1 — Wheat
dominant crop class
0.91
F1 — Cotton
highest-subsidy class
0.93
F1 — Maize
summer crop
10
Sentinel-2 bands
spectral bands per timestep
12
Temporal depth
satellite passes per growing season
Evaluation on Greek region test patches. Training used the PASTIS benchmark (France) with transfer fine-tuning on Greek data. Models: ConvLSTM (best temporal performance) and Conv3D (best overall).
Pipeline
Data
Sentinel-2 multispectral time-series. PASTIS benchmark + Greek region patches. 10 spectral bands, 12 timesteps per season.
Preprocessing
Cloud masking, temporal interpolation for missing passes, patch extraction (64×64 px) centred on declared parcels.
Model
ConvLSTM encoder captures spatio-temporal dynamics across timesteps. 3D convolutional decoder produces pixel-level class maps.
Validation
Predictions overlaid on farmer declaration maps. Inconsistent parcels — where the satellite class differs from the declared crop — flagged for review.
Prediction overlays, confusion matrix, and class activation maps
Assets to be added: drop exported PNG overlays + confusion matrix into content/hero/crop-classification/source-assets/