the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 29 Jun 2026
Explainable AI shows that a neural network learns extratropical cyclones as predictors of heavy precipitation
Abstract. Neural networks are increasingly used in weather and climate science, not only for prediction tasks but also for process understanding and scientific discovery, where model outputs must be linked to physically meaningful processes. Explainable artificial intelligence (XAI) helps establish this link by providing tools to interpret the information a neural network uses to make its predictions. However, most approaches rely on spatially aggregated or composite analyses that do not reveal the physical basis of individual predictions. Here, we present an object-oriented XAI framework that enables such prediction-level evaluation. We use this framework to analyse a simplified prediction task in which a neural network is trained to predict the occurrence of daily heavy precipitation in Western Norway across multiple prediction lead-times, several days in advance. In this study area, heavy precipitation is mainly associated with mid-latitude cyclones, providing a clear criterion: regions of high relevance identified by XAI should correspond to detected cyclones in the input fields. We find that most predictions are indeed associated with cyclones and that relevance patterns match key physical features such as the low-pressure centre and the zone of maximum winds. Furthermore, we show that predictions are based primarily on strong cyclones that travel along the North Atlantic storm track. This study provides a controlled benchmark that demonstrates that neural network predictions of heavy rainfall can align with established physical understanding. More generally, it illustrates how an object-oriented XAI framework can be used to assess physical realism at the level of individual predictions, representing an important step toward building the trust necessary to use these models for research and decision-making applications in weather and climate.
Competing interests: At least one of the (co-)authors is a member of the editorial board of Weather and Climate Dynamics.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.- Preprint
(1830 KB) - Metadata XML
- BibTeX
- EndNote
Status: open (until 14 Aug 2026)
- RC1: 'Comment on egusphere-2026-3362', Anonymous Referee #1, 02 Aug 2026 reply
-
RC2: 'Comment on egusphere-2026-3362', Anonymous Referee #2, 13 Aug 2026
reply
Review of “Explainable AI shows that a neural network learns extratropical cyclones as predictors of heavy precipitation” by Guillaume-Castel et al.
In this study, the authors train a convolutional neural network (CNN) to predict the occurrence of extreme precipitation over western Norway based on large-scale meteorological variables (specifically Z850 winds and sea level pressure). Predictions are made for the current day and up to 7 days ahead. The emphasis of the study is on analyzing the information used by the network to make extreme precipitation predictions. The authors apply layerwise relevance propagation (LRP), an explainable AI method, to identify pixels in the input domain that are used by the CNN to make predictions in days leading up to extreme precipitation. These pixels are then compared to the location of extratropical cyclones from an existing cyclone database. The results show that many predictions made by the CNN leading up to extreme precipitation are based on the presence of extratropical cyclones.
Overall, I think the study presents an interesting method. However, some of the results puzzling and I think the scientific contributions are relatively weak unless the authors go deeper into understanding the CNN predictions and the association between cyclones and extreme precipitation predictability. I also think the framing of the study could be improved to better highlight the contributions. The paper was generally easy to follow but could also benefit from some re-organization.
Main Comments:
General framing: I think the framing that other studies have not used XAI to look at individual predictions is incorrect. LRP specifically produces heat maps for a single prediction at a time, and many studies (including many of those referenced in the paper already) have presented results for individual predictions. Also, in the current study, most of the results are composites or summarized across many predictions (i.e. Fig. 6-9). I think the contributions of this paper are developing methods to link XAI relevance maps with specific meteorological features (cyclones). I think the introduction could be revised to convey this more clearly and provide a more accurate and nuanced review of the literature.
The paper organization is somewhat confusing. Section 2 is titled “Methods”, but includes results related to the CNN performance. The, most of the novel methods related to comparing LRP maps to cyclone location are presented in Sections 3 and 4. I suggest the authors separate the methods and results into clearly organized and titled sections.
Performance Metrics: It would be helpful to provide more background on how the PR-AUC is calculated (i.e. based on a graph that plots performance vs. recall at different probability thresholds). Second, it would be helpful to see the actual Precision-Recall curve, and to report statistics like precision and recall in the text (e.g. at the 0.5 probability threshold typically used to distinguish positive and negative predictions).
I am curious about the persistence baseline for the prediction of extreme precipitation on subsequent days. Why was this chosen as a benchmark? I am also curious what is driving the persistence skill. Is it truly that extreme precipitation is likely to occur in multi-day events? Or is the slightly elevated persistence skill compared to climatology somehow reflecting seasonality (i.e. where extreme precipitation is more likely in a certain season, and thus there is some elevated chance of extreme precipitation happening within the 7-day period following an event)? These are minor points because the persistence model is not the main focus of the paper, but looking into the temporal patterns of extreme precipitation might be useful for understanding other aspects of the CNN prediction.
When performing the cyclone-based LRP comparison, the authors only focus on positive cases (I take this to mean cases where extreme precipitation actually occurred regardless of whether the CNN prediction was correct or not). I don’t think this choice makes sense, nor is it clear why the authors make this choice. When interpreting the results, there is no context about whether the CNN predictions are correct. For example, are CNN predictions that were based on cyclones more likely to be correct or vice-versa? Do false positive predictions occur where there are cyclones that don’t result in extreme precipitation? It would be helpful to see the results of the object-based XAI analysis split up for true positive, false positive, true negatives, and false negatives.
Figure 6: it would be helpful if the authors could elaborate on the results in this figure with some further analysis. There is a relatively large gap between the orange and black lines. If I understand correctly, this indicates that, at 1- to 6-day lead times, the CNN is making predictions based on different cyclones than the cyclone this is associated with extreme precipitation on Day 0. Are the CNNs making correct predictions at these lead times? Or is it just coincidence that the CNN identifies a cyclone at those lead times, and the prediction isn’t necessarily correct? Splitting these results into true/false positives/negatives could help show what is going on.
Line-by-line comments:
Line 23: “XAI has notably been used to identify the physical drivers of weather events (e.g. Alessi et al., 2025;…”. I think this reference applies to mean precipitation and not weather events and may not be appropriate here.
Line 101: “Each input field is converted into anomalies by removing the temporal mean from each grid cell, then dividing by the global standard deviation.” Does global standard deviation refer to the entire globe or the entire input domain?
Section 2.2.3 Cyclone tracking dataset: In this section, it would be helpful to clarify that cyclones are tracked from one day to the next, so cyclones can be “backtracked” from D0.
Line 70: “Similar improvements in training are expected with lower thresholds of the 70th, 80th, and 90th Percentiles”. I don't understand what this sentence is trying to say. Is this referring to different thresholds for the label smoothing regularization, or for the extreme threshold?
Figure 2: I find the labels for the convolutional blocks to small/squished to read.
Line 245: “We start by defining a relevance patch as a contiguous area of high positive relevance within which every pixel must account for a relevance greater than 0.1%” . Are relevance values summed across all input variables?
Line 252: “If a large patch spans multiple cyclones, we first split the patch into individual patches by applying watershed segmentation to the distance-transformed masked relevance field, then assign cyclones to the resulting sub-patches.” I have two comments on this. First, how often does a large patch span multiple cyclones? It would help to have some context about this. Second, the statement on watershed segmentation is confusing. What is the watershed segmentation procedure? What is the distance-transformed masked relevance field?
Line 257: “Patches with a total relevance of less than 10% are discarded.” How often does this occur?
Figure 4: Why was this particular example chosen? It might also be helpful to see details of other individual predictions, for example in cases where relevance was not associated with a cyclone, or associated with a cyclone outside the Atlantic/Europe regions.
Figure 5: Caption refers to 5b but should be 4b. Also, caption for panel c indicated the cyclone is shown in red, but appears to be green.
Line 273: “To test this, we track the cyclones associated with D0 predictions backwards in time, and find that the decrease in the number of associated cyclones is broadly consistent with this physical constraint: the neural network generally does not assign relevance to cyclones that have not yet formed (Fig. 6, black line).” It took me a minute to understand what this is saying. It would be helpful to describe the methods and results from Figure 6 more in the text.
Line 310: “For example, moderately strong cyclones with a minimum SLP of 960 to 970 hPa are associated with predictions about 18% of the time they exist in the input fields, while strong cyclones deeper than 940 hPa are associated with predictions more than 50% of the time.” It might make more sense to perform this analysis (and that in Figure 8) for cyclones close to western Norway, since one wouldn’t expect cyclones in certain regions (e.g. Pacific or Tropics) to be associated with extreme P in western Norway.
Figure 8 (and related text): The wording “associated with predictions” is confusing. Does this mean associated with a correct, positive prediction? Or any prediction? I also found this terminology to be confusing throughout the text.
Figure 9: I find it interesting in panel c that there are more cyclones in the West Atlantic than are used by the neural network to make predictions for D-3 and D-4. This suggest that the CNN is not as accurate in tracking the cyclones from day to day leading up to extreme precipitation? Perhaps the authors can comment on this.
Sections 5.2, 5.3, and 5.4: These sections are pretty broad and make sweeping claims about the potential novelty or impact of the work. I would encourage the authors to reframe the conclusion section to focus on summarizing the key results, and discussing more specific contributions of the study rather than sweeping generalizations.
Citation: https://doi.org/10.5194/egusphere-2026-3362-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 42 | 22 | 6 | 70 | 5 | 4 |
- HTML: 42
- PDF: 22
- XML: 6
- Total: 70
- BibTeX: 5
- EndNote: 4
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The manuscript is a method paper focussing on the introduction of a novel explainable AI (XAI) method that facilitates the application of XAI for scientific discovery. To this end the authors combined a post-hoc attribution-based XAI method (LRP) with a large scale spatial field analysis to identify impact of atmospheric objects. Building on a clearly defined case study of cyclone impact on Norwegian strong precipitation events, the authors construct a CNN architecture that predicts a heavy precipitation event label for the following seven days (0 or 1 for each day) based on several atmospheric variable fields of cyclone-specific inputs. Across several experiments the author demonstrate that the network relies on the established cyclone patterns and driver. While I find the paper to be well written, structured and the work highly interesting, I have stronger concerns regarding the framing. I suggest reframing the work. I outlined why and how in the document below. Depending on the authors preferences, this might only be minor revisions.