Preliminary Results of Applying Transformers to Geoscience and Earth Science Data

Transformers with neural networks (NN) have become a dominant technology in AI/ML to achieve better accuracy in classification and prediction. We experimentally compare transformer-based NN and convolutional NN using various applications such as pollen detection and weather ITCZ prediction. Using machine learning in geoscience, earth sciences, and other natural sciences is not entirely new, but applying transformers to data from these environments creates many new opportunities. We apply Facebook's detection transformer (DETR) neural network, developed in 2020, to pollen and weather data to detect forty-four types of pollen and classify earth snapshots into two categories: having vs. not having the phenomena of double Intertropical Convergence Zones (ITCZs) in them. As we conduct our trials, we observe and document how the model performs on both datasets, detect the biases present in each layer of the network, and mitigate them as we tune the model and improve classification results even further.