Hi everyone,

Congratulations to all winners. It was a very interesting problem to work on.

I decided from day one to work with deep learning to see how far it can achieve on such a small, unbalanced and high dimensional dataset without acces to external data or pretraiend models. My idea was to crop a small patch around each field (mostly wider than field size) and patch mask with ones at pixels belong to the field and use both as inputs for a neural network. The first key success to make it work is to apply different augmentations including flipping, rotation, random cropping, mixup and time augmentation.

The model architecture consists of 3-layer Conv-net, Masked Features Averaging layer, 3-layer Bi-directional GRU-net and fully connected classification layer. Masked Features Averaging layer is similar to global average pooling but only averages pixels belong to crop field.

📷

Figure 1: Model Architecture.

The second key sucess is using Bagging Ensemble by training the model 10 times on different 85% subset of training data.

The data preprocessing steps includes:

• Removing B11
• Adding NDVI, NDWI and a third index using blue and infra-red.
• Applying square root on all bands to reduce skewness
• Applying standard scaling

As a future work, I think the reuslts can be improved dramatically using unsupervised pre-training on the unlabeled data but I didn't have time to try it.

You can find the code here.

At the end, I want to thank the organizers for their tremendous efforts.