| # Data Preparation | |
| Fine-tuning inputs are NumPy arrays with at least four columns: | |
| `[x_ft, y_ft, z_ft, las_class]`. Coordinates must use a projected feet-based | |
| coordinate system consistent within a crop. Do not supply geographic degrees. | |
| The public segmentation map is LAS 2 -> ground, 3 -> low vegetation, 4 -> | |
| medium vegetation, 5 -> high vegetation, and 6 -> building. All other classes | |
| map to ignore index `-1`. They remain in sampled encoder input; they must not | |
| be removed merely because they are ignored in loss. | |
| Use non-overlapping train, validation, and test manifests. Record crop size, | |
| normalization, thinning strategy, and class distributions beside each dataset. | |