Explainable Echo-State Networks for Climate Data Applications
Poster presented at Machine Learning for Actionable Climate Science Gordon Research Conference
Echo-state network models are a machine learning method that are promising for modeling climate data due to their ability to capture non-linear relationships and make accurate forecasts. However, many machine learning methods, including echo-state networks, are considered “black boxes” due to their algorithmic complexity, which leads to a lack of interpretability. Explainable machine learning techniques aim to provide insight into how black box models utilize input data for predictions. Recently, there has been growing interest in using and assessing explainability approaches for climate applications. In this study, we contribute to this emerging field by exploring the application of a novel feature importance explainability technique for echo-state networks in two climate applications. The applications have objectives that highlight different uses of feature importance: (1) understanding relationships between variables associated with a volcanic eruption, which serves as a proxy for a stratospheric aerosol injection, and (2) feature selection for subseasonal forecasts of extreme temperature events. We use these applications to discuss the potential benefits and drawbacks of explainable machine learning, with the aim of encouraging further research in this area, especially within the climate community.
