Using Machine Learning and XAI techniques for Convective Mode in Future Climate Change Scenarios

This research examines the morphology of simulated deep, moist convection in a suite of high-resolution regional climate simulations. To accomplish this goal, we first trained a convolutional neural network (CNN) on hand-labeled images of various storm morphologies extracted from GridRad’s composite radar reflectivity archive. We selected the CNN model architecture using an independent validation dataset, and the best-performing model was then evaluated using an independent testing dataset. The selected CNN was tasked with classifying the storm morphology of qualifying events identified in simulated composite reflectivity grids from three sources—namely, a retrospective (1990 ­– ­2005) and two future (1985­ – 2100) climate change simulations. The latter two simulations are based on representative concentration pathways (RCP) 4.5 and 8.5, and thus provide two future storm morphology projections based on an intermediate and extreme greenhouse gas concentration scenario, respectively. The spatiotemporal shifts in storm morphology are discussed, along with potential societal implications for many aspects of society. Finally, we use explainable artificial intelligence (XAI) techniques like backward optimization and input * gradient plots to interpret the CNN and build trust in the classification decisions made by the model.