-
Although this solution only stopped at top 10 in the LB, here is a quick note for feature selection that was used, which can be used as a reference for similar task:
- Raw 6-band satellite data, resized to 24x24, normalized to [0,1]
- Location encoding
- Nitrogen level encoding
- Passing days from the planting day
-
I tested several models, but finally used a simple CNN taking one time-point data of a location as input, while using average of yield scores on all the time points as output for one location. The final model is here
-
Several ideas that were tested but were not showing good results for my experiments:
- cnn-lstm
- contrastive learning to initialize weights for the models
mamba env create --file environment.yml
mamba activate corn_yield