Project aim
Test whether an Earth observation foundation model produces better 10 m crop-type maps for Ghanaian smallholder landscapes than a conventional CNN baseline, without leaking spatial information between training and test data.
- >0.70F1 (spatial CV)
- 10 mMap resolution
- ClayFoundation model
- ResNet-50Baseline
Illustrative graphic
Overview
The project fine-tunes the Clay foundation model on multi-temporal Sentinel-2 stacks over Ejura and benchmarks it against a ResNet-50 baseline. Evaluation uses spatial cross-validation across agro-ecological zones, and the fine-tuned model reaches an F1 score above 0.70.
The challenge
Smallholder fields are small, irregular and often intercropped, and labelled data is scarce. Random train/test splits overstate accuracy because neighbouring pixels are nearly identical, so the evaluation has to hold out whole regions to mean anything.
Data & inputs
Approach
-
01
Multi-temporal stacks
Sentinel-2 imagery is assembled into multi-temporal stacks covering the growing season.
-
02
Foundation-model fine-tuning
The Clay EO foundation model is fine-tuned for 10 m crop-type classification.
-
03
Baseline
A ResNet-50 is trained on the same data as a conventional point of comparison.
-
04
Spatial cross-validation
Folds are split by agro-ecological zone so test areas are never adjacent to training areas.
Results
Under spatial cross-validation across agro-ecological zones, the fine-tuned model achieves an F1 score above 0.70.
Validation
Spatial cross-validation across agro-ecological zones, benchmarked against a ResNet-50 baseline.