Evaluating Pretraining Methods for Deep Learning on Geophysical Imaging Datasets

Machine learning has the potential to automate the analysis of vast amounts of raw geophysical data, allowing scientists to monitor changes in key aspects of our climate such as cloud cover in real-time and at fine spatiotemporal scales. However, the lack of large labeled training datasets poses a significant barrier for effectively applying machine learning to these applications. Transfer learning, which involves first pretraining a neural network on an auxiliary “source” dataset and then finetuning on the “target” dataset, has been shown to improve accuracy for machine learning models trained on small datasets. Across prior work on machine learning for geophysical imaging, different choices are made about what data to pretrain on, and the impact of these choices on model performance is unclear. To address this, we systematically explore various settings of transfer learning for cloud classification, cloud segmentation, and aurora classification. We pretrain on different source datasets, including the large ImageNet dataset as well as smaller geophysical datasets that are more similar to the target datasets. We also experiment with multiple transfer learning steps where we pretrain on more than one source dataset. Despite the smaller source datasets’ similarity to the target datasets, we find that pretraining on the large, general-purpose ImageNet dataset yields significantly better results across all of our experiments. Transfer learning is especially effective for smaller target datasets, and in these cases, using multiple source datasets can give a marginal added benefit.