An experimental repo for classifiying BigEarthNet with Sentinel-2 Image Patches.
Used as a learning resource for building a selection of image models.
I wanted to train locally so all models were trained on a Nvidia 4070 Super.
ViT model from scratch based on this guide.
Total Parameters: 6,401,811
| Metric | Value |
|---|---|
| Precision | 0.7254 |
| Recall | 0.7819 |
| F1 Score | 0.7526 |
| Exact Match Ratio | 0.2725 |
| Images per second (Test) | 81,546 |
| Training time | 8.52 hrs |
A fine tuned CNN based model using ResNet18.
Total Parameters: 10,506,195
| Metric | Value |
|---|---|
| Precision | 0.7946 |
| Recall | 0.6919 |
| F1 Score | 0.7397 |
| Exact Match Ratio | 0.2881 |
| Images per second (Test) | 57,467 |
| Training time | 20.46 hrs |
A fine tuned ConvNeXt model using ConvNeXt
Total Parameters: 46,763,827
| Metric | Value |
|---|---|
| Precision | 0.7451 |
| Recall | 0.7906 |
| F1 Score | 0.7672 |
| Exact Match Ratio | 0.2949 |
| Images per second (Test) | 25,040 |
| Training time | 6 hrs |