the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 20 May 2026
Atmospheric Rivers landfalling in Japan: Climatology and physical characteristics causing heavy rainfall
Abstract. This study investigates the climatology and physical factors governing the precipitation efficiency of Atmospheric Rivers (ARs) landfalling in Japan. Using an ERA5 reanalysis-based AR database (1940–2023), we identified typical synoptic patterns of landfalling ARs via Self-Organizing Maps, which effectively categorize the regional moisture transport pathways. Climatological analysis revealed a significant increasing trend in AR frequency specifically in northern Japan. We further examined the relationship between AR characteristics and rainfall using nationwide high-resolution observations. While Integrated Water Vapor Transport (IVT) explains a substantial portion of the overall relationship (R = 0.71), considerable variability in precipitation amounts remains for similar IVT levels. Our analysis demonstrates that rainfall intensity is primarily modulated by a combination of strong moisture convergence (MVIMC), low convective inhibition (CIN), and orographic enhancement over high elevations. Furthermore, precipitable water (PW) emerged as the critical differentiator for the formation of quasi-stationary linear rainbands (QSLRBs), which consistently develop in close proximity to the AR axis. These findings suggest that the synoptic-scale AR provides the necessary environmental conditions for the organization of mesoscale extremes. Enhancing the predictive accuracy of AR landfall location and internal structure is thus a crucial prerequisite for improving the predictability of catastrophic localized rainfall in East Asia.
- Preprint
(1432 KB) - Metadata XML
-
Supplement
(232 KB) - BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2026-2597', Anonymous Referee #1, 03 Jul 2026
-
RC2: 'Comment on egusphere-2026-2597', Anonymous Referee #2, 10 Aug 2026
General comments
This manuscript addresses a useful gap in the literature. Most AR–precipitation studies in East Asia quantify the fraction of extreme events that coincide with ARs, which in effect treats the presence of an AR as a sufficient condition for heavy rainfall. The authors are correct to identify this limitation and to reframe the question as one of precipitation efficiency: given that an AR is present, what determines whether it produces damaging rainfall. The study combines an 84-year AR database with high-resolution radar–gauge observations and establishes an explicit link between synoptic-scale ARs and mesoscale quasi-stationary linear rainbands. The Self-Organizing Map classification is appropriate for the stated purpose, and the use of Huber-regression residuals to stratify events at comparable IVT is a reasonable way to isolate factors beyond moisture transport. I offer the following comments in the hope that they will help the authors further strengthen the manuscript.
Major comments
1. Multiple comparisons are not accounted for, which affects the precipitable water result.
Figure 4 reports ten significance tests, and Section 3.2.2 applies the same set of tests to the QSLRB comparison. The conclusion that precipitable water is the sole significant differentiator between AR events with and without QSLRBs (abstract; L260–263; L298) rests on a p-value of 0.02 obtained from approximately ten tests. This result would not remain significant under Bonferroni correction (α/10 = 0.005) or under a moderate false discovery rate procedure. As this finding constitutes one of the four principal conclusions of the paper, the authors should either apply an appropriate correction and report the outcome, or state explicitly that the result is exploratory.
The manuscript does not identify which statistical test was used. Given the skewed distributions evident in Figure 4, the choice between a t-test, Welch's test, and the Mann–Whitney U test is material, and the omission prevents reproduction of the analysis. The authors should state the test used and report effect sizes, such as Cohen's d or the rank-biserial correlation, alongside the p-values.
2. The precipitation metric does not represent localized extremes.
Section 2.4 states that all fields are spatially averaged to yield a single scalar value per AR event (L100–101). Figure 3 accordingly shows event-total precipitation reaching approximately 25 mm. The stated motivation of the study is catastrophic localized rainfall (abstract, L18–19), and Section 3.2.2 concerns systems capable of producing more than 150 mm in three hours. A domain mean of order 10 mm is a substantially different quantity, and spatial averaging over the full intersection area attenuates precisely the localized signal the study seeks to explain.
The authors should repeat the residual analysis using a percentile-based or maximum-grid-cell metric, such as the 95th-percentile or areal-maximum event-total precipitation within the intersection, and report whether the discriminating variables change. Robustness across metrics would strengthen the conclusions considerably. Sensitivity to the choice of metric would itself be an important result.
3. The trend analysis requires assessment of data homogeneity and field significance.
Two separate concerns apply to Figure 2c and the associated conclusion.
(a) The trend is computed over 1940–2023. ERA5 prior to 1979 relies on a considerably sparser observing system, and back-extension products are known to contain inhomogeneities associated with major observing-system transitions. A monotonic increase in detected AR frequency across 84 years may therefore reflect improving analysis capability in addition to any physical trend. The authors should repeat the Mann–Kendall analysis for the satellite era (1979–2023) and report whether the northern Japan signal persists. If the trend appears only in the full-period analysis, it should be presented with substantially greater caution.
(b) The test is applied independently at each 0.25° grid point across Japan, amounting to several hundred tests, without an assessment of field significance. The authors should apply a false discovery rate procedure (Wilks, 2016, BAMS) or a bootstrap field-significance test. In addition, annual AR frequency series may exhibit serial correlation. A modified Mann–Kendall test (Hamed and Rao) or prewhitening should be applied, or the absence of autocorrelation demonstrated.
4. Univariate comparisons cannot establish relative importance, and the predictors are likely collinear.
All conclusions regarding relative importance, including the description of MVIMC as a primary driver and of elevation as a decisive mechanism (L227–234), derive from ten separate two-group comparisons. These variables are unlikely to be independent. Lower AR-axis latitude (Fig. 4g) and higher landfalling elevation (Fig. 4j) may represent the same geometric effect, since an AR whose axis lies further south will intersect the mountainous interior of the target region rather than the coastal plain. Similarly, low CIN and high mid-level relative humidity are thermodynamically coupled.
A multivariate analysis is required before any variable can be described as primary. Suitable options include logistic regression on the positive/negative group label, or a random forest with permutation importance. In either case the pairwise correlation matrix among the ten variables should be reported. At minimum, the correlation matrix should be provided.
5. The AR–QSLRB association is established on temporal overlap alone.
L126–128 states that a QSLRB is classified as AR-associated when its occurrence overlaps temporally with the AR lifetime. ARs are present over Japan very frequently during June to August, with 9,037 six-hourly detections in the record, and AR lifetimes are long relative to QSLRB durations. Matching on time alone will therefore produce a substantial number of coincidental associations and will inflate the reported figure of 19.34%.
The authors should add a spatial criterion, such as requiring the QSLRB centroid to lie within the AR footprint or within a specified distance of the AR axis, and report how the statistics change. A null expectation should also be provided: what fraction of QSLRBs would be classified as AR-associated if the AR catalogue were randomly shuffled in time. Without this baseline, the reader cannot assess whether 19.34% is high or low.
6. Sensitivity to the AR detection algorithm is not examined.
The authors cite ARTMIP results demonstrating substantial disagreement among AR detection algorithms (Collow et al., 2022), but employ a single algorithm (tARget v4) throughout. Because the intensity criterion uses a location- and month-dependent 85th-percentile IVT threshold, the detected AR population during the Baiu season is sensitive to this choice. The authors should at minimum discuss how the conclusions may depend on the detection algorithm. A repetition of the key residual analysis using a fixed IVT threshold, for example 250 kg m⁻¹ s⁻¹, would provide a useful robustness check.
Minor comments
7. SOM configuration. L141–143 states that multiple configurations were tested and that a 2 × 2 structure was selected. The quantization error and topographic error for the tested configurations should be reported so that the choice can be evaluated. In addition, the use of a single representative timing per event (L85) discards the temporal evolution of the AR. This should be noted as a limitation, particularly because the discussion (L312–314) identifies sub-event evolution as an important direction for future work.
8. Analysis period in Section 3.2.2. The analysis is restricted to 2006–2023 without explanation. This is presumably because the resolution of the RA dataset changes from 5 km to 1 km in 2006 (L76), which is a sound justification. The reason should be stated in the text.
Citation: https://doi.org/10.5194/egusphere-2026-2597-RC2
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 171 | 88 | 21 | 280 | 26 | 15 | 13 |
- HTML: 171
- PDF: 88
- XML: 21
- Total: 280
- Supplement: 26
- BibTeX: 15
- EndNote: 13
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The paper shows that IVT alone is not enough to explain heavy rainfall from ARs in Japan, using SOM based AR classification, Huber regression residual grouping, and a QSLRB overlap analysis. The Mann Kendall trend result (AR frequency rising in northern Japan) and the finding that PW separates QSLRB from non QSLRB events are the strongest parts of the paper. But the statistics connecting Figure 4 and 5 to the paper's main claim are not strong enough to support how confidently the claim is stated. The paper also never looks at vertical circulation structure even though its own explanation depends on convergence and lifting. And the AR to rainfall relationship is only checked as same window overlap, with no check for whether rainfall happens before, during, or after AR landfall. Major revision is recommended.
Major Comments:
1. In lines 219-238, Figure 4. Ten separate tests are run (IVT, PW, CAPE, CIN, MVIMC, RH500, latitude, duration, overlap, elevation) on groups of 46 and 50 events, with no correction for running that many tests at once. At a 0.05 significance level across 10 tests, you would expect about one false positive purely by chance. Five out of ten come back significant here. Without a correction such as Bonferroni or FDR, this result cannot be trusted as strongly as the paper treats it, and it is central to the main claim of the abstract.
2. In lines 198-217. The way events are split into positive and negative residual groups may already guarantee part of the result. IVT and PW are physically related to each other. When you regress rainfall on IVT and look at what is left over, you also remove any effect that comes from things correlated with IVT, including PW. So, the finding that "PW shows no difference between groups" (line 222) might just be a side effect of how the groups were built, not real evidence that PW does not matter. The authors should report the correlation between IVT and PW directly, and discuss whether their method could even detect a PW effect if one existed.
3. In lines 219-238. MVIMC, CIN, RH500, elevation, and latitude are treated as five separate independent factors, but some of these are likely connected. Higher elevation naturally increases convergence through orographic lifting, so elevation and MVIMC may be reflecting the same physical process rather than two separate ones. A partial correlation or multiple regression is needed before listing both as separate conclusions (lines 326 to 328).
4. In lines 141-143. The choice of a 2 by 2 SOM grid is only explained in words, not backed by numbers. Since Cluster D at 44.7 % of events (line 143) is the basis for the whole climatology section, the grid size choice needs a quantitative check, such as quantization error or topographic error across different candidate grid sizes.
5. In Section 3.1 and 3.2, Figures 1, 5, 6. All circulation analysis in the paper is based only on IVT composites. The discussion (lines 288 to 297) argues that convergence and low CIN drive rainfall, but this is inferred from a single CIN number rather than shown directly. Adding vertical velocity and geopotential height anomaly composites at 500 hPa and 850 hPa for the positive and negative residual groups, alongside the existing IVT panels, would let the authors actually show this mechanism instead of inferring it. This kind of analysis is used in Pandidurai et al. (2026), https://doi.org/10.1002/joc.70462.
6. In lines 96-128 and 245-248. The relationship between ARs and rainfall is only checked through overlap in time and space (lines 96-99, 127-128), never through timing offset. Since both ARs and QSLRBs last several hours on average (Figure 2b shows 10 to 13.5 hours), simple overlap cannot tell apart rainfall that happens at the same time as the AR from rainfall that happens as a delayed response, for example moisture building up first and heavy convection organizing later. The authors should compute a lag relationship, either a cross correlation or a distribution of the time gap between peak rainfall or QSLRB onset and AR axis passage, going out to at least 48 hours given the six hourly resolution of the data. This should be done separately for each SOM cluster and for QSLRB versus non QSLRB events before combining everything, because pooling all 479 events together first could hide a real lag if different AR types behave differently. This connects to the sequential moisture approach in Pandidurai et al. (2026), cited above in M5.
Minor Comments
1. Lines 218 to 221: The sentence describing what Figure 4 shows is repeated almost word for word right after itself. Remove the duplicate.
2. Lines 65 to 68: "we utilized a suite of ERA5 based variables and indices except for precipitation" is awkward. Consider "excluding precipitation, which came from RA data instead."
3. Lines 219, 260 (Figure 4, Figure 5c): The statistical test used to get the p values is never named. Since the paper already points out that rainfall data has heavy tails and uses Huber regression for that reason (lines 108 to 111), using a plain t test elsewhere would be inconsistent. State and justify whichever test was used.
4. Line 232: "Landfalling Elevation" is never clearly defined. Is it the elevation at the point where the AR axis meets the coast, or the average terrain elevation across the AR footprint? Needs one clarifying sentence.
5. Lines 51, 172, 320 and reference list: The name "Hirockawa" is used throughout, but the more common spelling in the AR and QSLRB literature is "Hirokawa." Worth checking if this is correct or a repeated typo.
6. Lines 245 to 250: State clearly that the QSLRB analysis only covers 2006 to 2023 because of the RA data limits, while the residual group analysis in section 3.2.1 covers a different period. This is currently left unstated.
7. The authors note that the causes behind shifting AR tracks remain unclear (lines 315–317). One direction for future work would be tracing moisture sources for landfalling AR events and using higher time resolution moisture budget datasets to better capture the fine-scale convergence signal discussed as central to rainfall efficiency in Section 3.2.1. Relevant studies include: https://doi.org/10.1029/2026EF008305, https://doi.org/10.1029/2026GL121973, and https://doi.org/10.1038/s41597-025-06044-y.