A group project for the "Data Science in Earth Observation" module (WS 23/24, TUM)
Project tasks:
- Use Sentinel-2 multi-spectral imagery to derive the dominant tree species for forests in Germany
- Dominant tree species shall be classified by a machine-learning approach trained through the provided reference data
How to start:
- Create an environment by typing
conda create --name <name> python=3.10 - Activate the environment by typing
conda activate <name> - Run the command
pip install -r requirements.txt
| Model | Dataset | Accuracy |
|---|---|---|
| SVM (Handeul) | ᛫ top10 ᛫ split 0.8/0.2 ᛫ augm? |
47.22 % |
| KNN (Yi) | ᛫ top10 ᛫ split ? ᛫ augm? |
40 % |
| MLP (Yi) | ᛫ top10 ᛫ split ? ᛫ augm? |
22 % |
| ConvNet (Colin) | ᛫ top10 ᛫ split 0.7/0.3 (42) ᛫ mirror, 90°, 180° |
46 % |
| ResNet (Colin) | ᛫ top10 ᛫ split 0.7/0.3 (42) ᛫ mirror, 90°, 180° |
49 % |
| ResNetDropout (Colin) | ᛫ top10 ᛫ split 0.7/0.3 (42) ᛫ mirror, 90°, 180° |
49 % |
| Random Forest Classifier with hyperparameter optimization (Chris) |
- split 0.7/0.3 - 57928 samples - data augmentation included - no autumn bands - no nan-values |
44 % |