A deep learning model for the automatic classification of tree species from point cloud data. The PointMixer network is chosen for the classification task: PointMixer: MLP-Mixer for Point Cloud Understanding
The point cloud data used for training: >17k individual trees from 33 tree species captured with TLS, MLS and ULS.
The data are divided into training and validation sets. The training set consits of 90% of the data, whereas the rest 10% comprise the validation set. The split is done so that the training data contains a representative number of trees with various diameters from all three sensors.
The PointMixer network is trained with the following inputs and hyperparameters:
- 4098 input points, selected from the original point cloud using the Farthest Point Sampling algorithm
- 42 batch size
- Initial learning rate: 0.1
- Cosine-annealing decay
-- Minumum learning rate: 0.00001
-- Maximum number of iterations: 300 - Training epochs: 300
The best model was chosen based on the performance on the validation set. The best model was obtained during epoch 269 and can be downloaded here.
Metrics computed on the validation set:
| Overall accuracy | Precision | Recall | F1-score |
|---|---|---|---|
| 0.799 | 0.800 | 0.799 | 0.791 |