Explaining patterns learned by neural networks for detection of atmospheric surface fronts in gridded atmospheric data
Abstract. Atmospheric surface fronts (SFs) strongly influence midlatitude weather through storms, precipitation, and abrupt temperature changes, motivating reliable detection. Here, we reproduce a U-Net, a type of artificial neural network, to detect SFs in ERA5 reanalysis data with labelled datasets over North America and Europe. The goal of this study is to apply Layer-wise Relevance Propagation (LRP) to interpret the learned patterns. To reduce computational cost, we introduce a, compared to previous studies, reduced U-Net variant and a reduced input set of six variables - temperature, specific humidity, and horizontal wind - without reducing detection quality. Relevance analysis shows that U-Net focuses partly on physically meaningful regions near detected SFs, consistent with rule-based methods, but also noisy patterns and spurious correlations far from the detected fronts. Moreover, U-Net employs two nearly inverted relevance patterns for different parts of a front, which we see as not physically plausible.
Using a simplified synthetic dataset of a 1-D field with SF labels placed at the minimum of the second derivative, we can link the pattern to the front position relative to the frontal zone. Increasing the label resolution enables the U-Net to learn a more plausible pattern without the split in the synthetic case.
When evaluated over a larger region (Pacific to Europe, equator to Arctic), U-Net performs well in midlatitudes but over-detects SFs in the Tropics and Arctic, for example regularly detecting SFs within tropical cyclones. Together with the partially implausible relevance patterns, this indicates that ANN-based SF detection should be used with care outside the training regions. Retraining on the larger domain improves detection skill and the plausibility of relevance patterns, but the split pattern persists. Overall, the results highlight the value of explainable AI for identifying physically implausible shortcuts and the need for a global SF dataset to enable reliable detection worldwide.
Dear authors,
Unfortunately, after checking your manuscript, it has come to our attention that it does not comply with our "Code and Data Policy".
https://www.geoscientific-model-development.net/policies/code_and_data_policy.html
To access the ERA5 data used in your study you provide a link to the Copernicus Data Store, which does not comply with the requirements of the journal, as it is not a long-term data repository. The GMD review and publication process depends on reviewers and community commentators being able to access, during the discussion phase, all the code and data on which a manuscript depends, and on ensuring the provenance of replicability of the published papers for years after their publication. Please, therefore, publish the requested data in one of the appropriate repositories and reply to this comment with the relevant information (link and a permanent identifier for it (e.g. DOI)) as soon as possible. We cannot have manuscripts under discussion that do not comply with our policy. Â
Later, if the Topical Editor decides to continue with the review or publication process of your manuscript and you are requested to upload a new version of it, then The 'Code and Data Availability’ section of your manuscript must also be modified to cite the new repository locations, and corresponding references added to the bibliography.Â
I must note that if you do not fix this problem, we cannot accept your manuscript for publication in GMD.
Juan A. Añel
Geosci. Model Dev. Executive Editor
Â