Explaining neural networks for detection of tropical cyclones and atmospheric rivers in gridded atmospheric simulation data

Abstract. Detection of atmospheric features in gridded datasets from numerical simulation models is typically done by means of rule-based algorithms. Recently, the feasibility of learning feature detection tasks using supervised learning with convolutional neural networks (CNNs) has been demonstrated. This approach corresponds to semantic segmentation tasks widely investigated in computer vision. However, while in recent studies the performance of CNNs was shown to be comparable to human experts, CNNs are largely treated as a “black box”, and it remains unclear whether they learn the features for physically plausible reasons. Here we build on the recently published “ClimateNet” dataset that contains features of tropical cyclones (TCs) and atmospheric rivers (ARs) as detected by human experts. We adapt the explainable artificial intelligence technique “Layer-wise Relevance Propagation” (LRP) to the semantic segmentation task and investigate which input information CNNs with the Context-Guided Network (CGNet) and U-Net architectures use for feature detection. We find that both CNNs indeed consider plausible patterns in the input fields of atmospheric variables. For instance, relevant patterns include point-shaped extrema in vertically integrated precipitable water (TMQ) and circular wind motion for TCs. For ARs, relevant patterns include elongated bands of high TMQ and eastward winds. Such results help to build trust in the CNN approach. We also demonstrate application of the approach for finding the most relevant input variables (TMQ is found to be most relevant, while surface pressure is rather irrelevant) and evaluating detection robustness when changing the input domain (a CNN trained on global data can also be used for a regional domain, but only partially contained features will likely not be detected). However, LRP in its current form cannot explain shape information used by the CNNs, although our findings suggest that the CNNs make use of both input values and the shape of patterns in the input fields. Also, care needs to be taken regarding the normalization of input values, as LRP cannot explain the contribution of bias neurons, accounting for inputs close to zero. These shortcomings need to be addressed by future work to obtain a more complete explanation of CNNs for geoscientific feature detection.