AI Applications to the Earth Sciences: 35 years through the lens of the AMS AI Committee

The development and application of artificial intelligence methods to problems in the earth sciences has evolved at a fast pace. Its progress has been made possible by developments in the computer sciences, the availability of larger and more comprehensive environmental data sets, without forgetting the ever-increasing availability of affordable computing power. The past 35 years of this evolution is examined through the lens of the activities of the AMS AI committee. The committee, or its ancestor, started organizing workshops in 1985. It second workshop, in September 1987, took place in Boulder Colorado and was organized by Rosemary Dyer and William Moninger. The workshop gathered 80 participants with topics familiar to present AI practitioners in the earth sciences including meteorology, hydrology, environmental protection, etc. In the eighties the AI methods of choice were focused on expert systems and their inference engines. The proceedings of the 1987 workshop includes only one mention of Neural Networks.

The first AMS AI conference took place in 1998 and included 47 presentations with many earth science topics similar to those presented today: precipitation predictions, satellite retrieval and pattern recognition, climate classification and prediction, image processing, decision aids and natural language systems. Looking back at the abstracts and presentations of past AMS AI conferences allows to track the popularity of different AI/ML methods and how they were implemented in the earth sciences. While Neural Networks were the most popular technique in 1998, mentioned in 49% of all abstracts, this percentage drops to 27% in 2008 and 10% in 2019. The implementation of other AI techniques for earth science problems varied through the conferences and included fuzzy logic, tree-based methods, genetic algorithms and support vector machines. The field continues to evolve and by the 2019 AMS AI conference, deep learning was by far the method of choice with 36% of all abstracts, far ahead of any other AI method. We are expecting this trend to continue in parallel with an increase in the popularity of AI/ML as a method to better predict and gather a deeper understanding of a wide variety of complex and nonlinear processes in the earth sciences.