Interpretable Deep Learning for Identifying and Analyzing Marine Heatwave Patterns

Abstract

Marine heatwaves (MHWs) are intensifying due to climate change, severely impacting marine ecosystems and coastal economies. Identification of MHWs through traditional models is generally not interpretable and it is difficult to understand the factors behind these events. This paper proposes a novel CNN-based attention model to identify and examine the MHW patterns from the sea surface temperature (SST) data. Interpretable in our model the fact that this was emphasized is that we’re able to see the driving variables: atmospheric pressure, ocean currents, and wind speed. We give a visual and quantitative understanding of how the model finds heat waves and what features are most important for predicting them using techniques like Grad-CAM and Shapley Additive Explanations (SHAP). Beyond this, this approach also enhances the detection accuracy and provides better information for decision aiding in early warning systems and coastal management strategies. Using this model, it is shown to be evaluated on a multi-year dataset provides good performance in distinguishing MHWs from nonheatwave periods, and is better predictive than even existing methods. The ability of interpretable deep learning to explain and help understand climate science in terms of the occurrence and evolution of marine heatwaves will provide actionable insights that are essential for the adaptation of the climate to vulnerable coastal regions.