OGH2023_hack

How to improve the result?

1. Model hyperparameter tuning
2. Ensembling different models
3. Dataset balancing (you can also combine training and validation datasets)
4. Feature engineering including:
   • Clustering
   • Tasseled cap transformation
   • Spectral indices
   • Using textural features (e.g. SAGA GIS)
   • Reducing spatial resolution
5. Additional features:
   • Radar data from Sentinel 1
   • Normalized digital surface model (nDSM)
   • Panchromatic band (15 m)
   • Satellite scene from another date(s)
6. Post-processing:
   • Modal filter
   • Sieve filter

It is recommended to use a machine learning framework, e.g. tidymodels, mlr3 or scikit-learn. Check out also programming language agnostic frameworks such as mlpack and H2O.