2024-10-04
WP 200
Iterative Algorithm Desing
and Data Collection Sprints
Each including relevant data collection.
Morphometric classification homogenisation protocol
momepy
, libpysal
, sgeop
AI model design and development
80/20%
Model | Architecture | #Labels | Images |
---|---|---|---|
Satlas | SwinT | 302M | Sentinel-2 |
Clay | MAE/ViT | 70M | Sentinel-2/Landsat/NAIP/LINZ |
Prithvi | MAE/ViT | 250 PB | Sentinel-2/Landsat |
Model | Satlas | Clay | Prithvi |
---|---|---|---|
Run time (per epoch) (with GPU) | 9 mins | 8 mins | 20 mins |
# parameters | 90M | 86M | 120M |
Implementation | 5/10 | 6/10 | 7/10 |
Hyperparameter tracking | Own setup | Wandb.ai | Tensorboard |
Satlas | Clay | Prithvi | |
---|---|---|---|
Accuracy (weighted) | 0.57 | 0.72 | 0.62 |
IoU (weighted) (0-1) | 0.33 | 0.58 | 0.41 |
F1 (weighted) | 0.41 | 0.69 | 0.58 |
Without hyperparameter tuning!
Satlas | Clay | Prithvi | |
---|---|---|---|
Accuracy (weighted) | 0.25 | 0.72 | 0.59 |
IoU (weighted) (0-1) | 0.2 | 0.58 | 0.42 |
F1 (weighted) | 0.21 | 0.69 | 0.59 |
Urbis 24, Frascati
World Urban Forum, Cairo, November 6th @ 15:30
Monitoring urban fabric for data-driven planning and decision-making