Explaining neural networks for detection of atmospheric features in gridded data

Detection of atmospheric features in gridded datasets 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 recently published studies that discuss datasets containing features of tropical cyclones (TCs), atmospheric rivers (ARs), and atmospheric surface fronts (SFs) as detected by human experts. We adapt the explainable artificial intelligence technique “Layer-wise Relevance Propagation” 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 for the detection of TCs and ARs, 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. In contrast, for the detection of SFs, we find only partially physically plausible patterns. While U-Net uses regions of changing temperature and humidity as well as strong wind shears to detect SFs, we also find noisy patterns relating to spurious correlations with the background data. To assess whether these implausible patterns reduce U-Net's generalizability, we evaluate it on a different SF dataset. Here, depending on the domain, SFs are often erroneously detected, especially in the Tropics and Arctic, highlighting the importance of analyzing whether patterns learned by a CNN are physically plausible. We also demonstrate application of the approach for finding the most relevant input variables and evaluating detection robustness when changing the input domain.