Prithvi-EO-2.0 on Ghanaian Data: Honest Accuracy Numbers
By Samuel Appiah Kubi · 08 Jul 2026
Foundation models are powerful. They are not magic. This post reports what Prithvi-EO-2.0 actually achieves on Ghanaian Sentinel-2 imagery, validated against ground truth we collected independently.
Experimental Setup
We ran pygeovision's PrithviTasks(task='land_cover') on 50 Sentinel-2 L2A scenes acquired during the 2024 dry season, covering all 10 regions of Ghana. Predictions were validated against 800 GPS field reference points collected by CSIR-GIMPA between January and March 2024, classified into 6 land cover categories: Closed forest, Open forest/savanna, Crops, Water, Urban/Built, Bare.
Results
| Class | User Accuracy | Producer Accuracy | F1 |
|---|---|---|---|
| Water | 97.2% | 98.1% | 0.977 |
| Urban | 84.3% | 79.6% | 0.819 |
| Closed forest | 81.7% | 76.4% | 0.789 |
| Crops | 73.2% | 81.5% | 0.771 |
| Open forest/savanna | 68.4% | 71.3% | 0.698 |
| Bare | 71.6% | 68.9% | 0.702 |
Overall accuracy: 78.3%. Kappa: 0.72. These are zero-shot results — no fine-tuning on Ghanaian data.
Where It Fails
The most common misclassification is open forest/savanna being confused with crops. In Ghana's transition zone (Brong-Ahafo, Northern Region), agroforestry systems combine tree crops (cocoa, shea) with food crops in the same 10 m pixel — a class that neither Prithvi's training data nor the ESA WorldCover schema handles well. After 500-sample fine-tuning with local reference data, accuracy on this class improved from 70% to 83%.
We recommend treating Prithvi zero-shot results as a first draft requiring local validation, not as a final product for government reporting.
