A Quantitative Analysis on the Use of Supervised Machine Learning in Earth Science

Abstract:

Review articles on the application of machine learning (ML) techniques to Earth science problems commonly cite the lack of labeled training data as a challenge to supervised learning. Our objective is to understand the landscape of supervised ML within the Earth sciences. To that end, we have analyzed 10 years of articles published in AGU, AMS, IEEE, and SPIE journals to understand trends in the application of supervised ML techniques to Earth science problems. We have further performed intensive analysis of AGU papers from 2018-2019, extracting information such as the Earth science sub-domain, ML algorithm(s) applied, number and type of training data, and methods used to label the training data. We find that ML use in the Earth sciences is growing rapidly, but the plurality of ML algorithms were trained using only hundreds of labeled samples. Our results underscore the lack of training data across the Earth science research community and the need for open sharing of the available data and the development of large training datasets.