Image Similarity in the Atmospheric Sciences: Leveraging Pre-Trained Neural Networks for Classification

Image analysis is a central tool in the atmospheric sciences. With recent massive increases in the size and complexity of atmospheric imagery data, analysis of individual images and simplified image processing techniques are no longer sufficient for many scientific tasks. Deep learning methods have shown promise in this space, however, training these models on large datasets can be computationally expensive. Leveraging the architectures and weights of pre-trained neural networks can alleviate much of this expense, but computational cost savings may have tradeoffs with prediction efficacy.

In this work, we demonstrate the potential benefit of using pre-trained neural networks from a recent release from the deep learning library, Keras. We compare performance using transfer learning techniques to a simple convolutional neural network as a baseline. We evaluate applicability of pre-trained neural networks on several datasets of interest to a wide community and are notorious as difficult learning tasks, namely – cloud morphology characterization, methane plume classification, and air mass back trajectory analysis.