Appeared at AI for Earth and Space Science workshop, ICLR 2022. In this paper we examined if our previously proposed method to overcome data quantity and resolution issues is able to work independently of the machine learning algorithm used to make predictions. We compared three neural network architectures with a reference Random Forest model in a case study of nitrogen response rate prediction. The architectures were:
- a Multilayer Perceptron
- an MLP Autoencoder where we replace the decoder with a regression head after training
- an MLP dual-head Autoencoder which optimizes the reconstruction and prediction losses simultaneously
On the repository there is also the LSTM version of the dual-head Autoencoder which was not included in the paper:
