Abstract

The explosion of scientific development in machine learning and artificial intelligence over the last 30 years, concurrent with the growth in CPU and GPU power to allow for such models to be readily accessible to the average scientist, has resulted in a sea change in statistical studies of the environment. The merging of the two fields - environmental statistics and data science - has resulted in a new focus: Environmental Data Science, which uses a mixture of tools from both statistics and machine learning fields to tackle complex, environmental problems. Two primary concerns statisticians have had of the use of machine and deep learning neural network models is their lack of both uncertainty quantification and inferential potential. This is an open area of research, and the development of "Explainable AI" over the last few years has aimed to mitigate these concerns. Our collaborative working group tackled the problem of explaining which inputs are important, when modelling environmental data. We focused on three general model agnostic methods for explainability, and applied them to a variety of models, all centred around a somewhat classic environmental statistics problem: long-lead forecasting of monthly soil moisture in the North American "corn belt", using sea surface temperature anomalies in the Pacific Ocean. In this talk we will review the findings of this working group, and give some reflections on the lessons learned for pursuing a (successful) collaborative research project across two continents and five time zones.