the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 06 Mar 2026
Integrating total water level data and visual evidence to assess coastal flooding in San Diego County
Abstract. Coastal flooding in Southern California poses a growing threat to communities and infrastructure, exacerbated by climate change and sea level rise. Total water level (TWL), the combination of sea level, tides, and wave runup, is increasingly used to forecast coastal flooding, but validating the thresholds at which flood impacts occur remains a challenge. This study examines the relationship between modeled TWL and photographic or video evidence of flooding in San Diego County from 2010 to 2024. We integrate model output from the Coastal Data Information Program with visual records from community monitoring programs to assess spatial and seasonal variations in flood occurrence. We also evaluate the influence of atmospheric rivers and El Niño conditions. Atmospheric river days were associated with an increase in the likelihood of observed flooding, and El Niño winters showed a positive but weaker correlation. Overall results demonstrate a robust but imprecise correlation between modeled TWL and observed flood impacts, with uncertainty driven largely by convenience sampling in the visual dataset. Despite these limitations, modeled TWL is shown to be a useful proxy for flood risk. Our findings underscore the need for systematic flood impact documentation to refine threshold estimates and improve flood forecasting and coastal management.
- Preprint
(8758 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2026-820', Anonymous Referee #1, 24 Mar 2026
-
RC2: 'Comment on egusphere-2026-820', Matt Conlin, 20 Aug 2026
Pinto et al. combines modeled TWL data in San Diego County with photographic evidence of flooding to attempt to define alongshore-varying TWL thresholds at which flooding is more likely than not to occur. Locally-relevant flood thresholds are notoriously difficult to estimate due to the hyper-local nature of flooding, spatial variability in processes and infrastructure, and difficulty in accounting for wave runup. As such, this study could serve as a valuable source of information for San Diego County coastal planners. This study’s basis in leveraging a large suite of existing quantitative coastal data (pre-existing cross-shore transects, existing hydrodynamic model outputs) and combining it with on-the-ground evidence is commendable.
However, I am concerned that the statistical method used provides flood thresholds that are strongly over-estimated relative to the purported 50% likelihood value. Examining Figures 7, 8, and 11, it seems clear that the majority of documented occurrences of flooding occurred at TWLs below the derived 50% likelihood values, making these values seem incongruent with the observations. Furthermore, an important source of context to these results (which I think should be included in the Introduction and Discussion) are the established NOAA (NWS and NOS) flood thresholds at the La Jolla tide gauge, and while the comparison is certainly not apples-to-apples it is notable that the derived 50% likelihood values are substantially higher than even the NOS major flood threshold (2.0 m above MSL). Based on Figures 7, 8, and 11 it seems that the large majority of cases where no flooding was documented (which, importantly, does not necessarily mean flooding did not occur) is pushing the thresholds higher in the logistic regression than would be expected from examining the confirmed cases of flooding alone (as the Authors state in the caption of Figure 11). Therefore, I do not believe the logistic regression method used is an accurate way of estimating the 50% flood likelihood level. Furthermore, L229-231 seem to imply that the logistic regression was applied to transects with only one documented case of flooding, which should not be possible and gives me concern about how the method was applied. Overall, since it is not possible to disentangle if a “no evidence” case is due to a true lack of flooding or simply a lack of documentation, it may be appropriate in an analysis to weight more heavily the confirmed cases of flooding.
Another strong concern I have relates to the beach slope. The Author’s state that they used CoastSAT to derive beach slopes, however it is unclear if these values were spatially and/or temporally varying or fixed in time/space. And if they were temporally varying, was the beach slope estimated at the closest possible time to the photo? The Stockdon equation is known to be fairly sensitive to the input beach slope, so the method used might have a non-trivial effect on the resulting flood thresholds. Some sensitivity analysis would be beneficial for this, and perhaps could be used to add uncertainty driven by the slope to the confidence intervals of the derived thresholds.
Minor comments:
L105: Should provide more background about the Stockdon equation and CoastSAT.
L105: Were beach slopes temporally varying? As in, did you find the beach slope from the closest possible time to each image? Or was it a single value? Either way, Stockdon is quite sensitive to beach slope, and there should likely be some sensitivity analysis on the results around uncertainty in beach slope, as this could non-trivially affect the final threshold estimates.
L119: I think you probably mean that data from the buoys are assimilated into the wave model and/or provide boundary conditions. Please replace the word “generate” with a more appropriate description.
L121-122: Curious how strong of an effect this midpoint coordinate method has on the results. For the example presented of a parking lot, probably not much. But how un-specific did you allow the “general location descriptions” to be. E.g. what if it said “Imperial Beach” or “San Diego County”? I think some more detail on these points should be added.
L130: Should provide some background and references on what logistic regression and AIC are. Additionally, as related to another comment below, it should be specified that the logistic regression can only be used at transects with multiple photos, which is only a subset of the total number of transects. Finally, while you state that AIC was used, it is never mentioned in the Results or Discussion. Please clarify.
Figures 4 and 5: If I am interpreting these figures and the associated text correctly, if two different transects each had a photo of flooding on the same day– hence driven by the same weather event – this would be considered two different “events” in the plot. If this is correct, these plots are really combining information about both temporal occurrence and spatial extent of individual weather events. As such, it is hard to tell if January had more individual weather events, or just a few large weather events that induced flooding over a large spatial extent. At the very least, I think the word “events” in the captions should be changed to “transect-days” to more accurately describe what the plots are showing and some description of the above should be included.
L201: Looking at the plot I don’t agree that “most” photos aligned with TWL peaks. Many seem to, but not most.
Figures 7 and 8 – Looking at the right-hand panels in these figures, the logistic regression-derived 50% likelihood threshold doesn’t pass a sanity-check for me. In both cases, the vast majority of days with flooding evidence fall below the derived 50% line. The fact that there are many days with high TWL but no evidence, which could be because there was flooding but no one took a picture, seems to be pushing these derived thresholds higher than would seem reasonable based solely on the cases of documented flooding. Relatedly, you label the two categories “No impact” and ‘Impact” but this is somewhat misleading since they are really “No evidence of impact” (though impact could still have occurred) and “Evidence of Impact”.
L217: Here you say the data are in the MSL datum, which is the first it is mentioned. For all figures and in the Methods, please include the datum the data are relative to.
Figure 9: Please describe how the error bars are generated, I assume they are the confidence interval from the logistic regression. This also gets back to previous comment about needing to include more background on logistic regression.
L229: How can a logistic regression be run on transects with only one photo? Logistic regression breaks down for the case of a single positive sample. How were thresholds identified in this case? This must be clarified.
L231: Here you say there are 21 transects with at least 2 photos. Figures 9 and 10 claim to show values for the same set of data (“more than one photo”), yet only 20 transects are shown in the plots. This inconsistency needs to be addressed.
L236: More information about the Bayesian models need to be provided in order for the reader to evaluate the comparison properly.
Figure 11: Similar to Figures 7 and 8, looking at this plot the derived 50% thresholds don’t pass a sanity check for me. The logistic regression purports to give the TWL at which flooding is equally likely to occur or not, but all of these thresholds are on the high-end of TWLs for flooding evidence. This plot tells me that the derived threshold values are those where flooding is very likely to occur. From this figure alone, as you allude to in the caption, taking the intersection of the distributions as the 50% likelihood threshold would seem to make sense, and these thresholds would be substantially lower (~2.0-2.5 m rather than 3.0-4.0+ m above MSL) and indeed would align more closely with the NOAA threshold at La Jolla.
L273-277: I don’t think this paragraph adds very much to the results, especially if plots are not shown. I think it should be removed.
L328-329: This claim needs to be cited.
Something that I think should be discussed is how your derived thresholds compare with published NOAA (NWS and NOS) flood thresholds at the La Jolla tide gauge. Values can be found here: https://tidesandcurrents.noaa.gov/inundationanalysis/analysisParams.html?station=9410230. Converting the published thresholds from MHHW to MSL yields the following: NWS minor = 1.31 m, NOS minor = 1.36 m, NOS moderate = 1.64 m, NOS major = 2.0 m. While it should be expected that your TWL thresholds may be somewhat higher than these SWL-based thresholds, there are differences in what your “50% chance of flooding” and these values represent, and your thresholds should capture much more spatial variability, it is notable that your derived thresholds are generally much higher than even the NOS major threshold. Although, it is interesting that the transects in southern La Jolla (where the tide gauge is) are closer to the 2.0 m value.
Citation: https://doi.org/10.5194/egusphere-2026-820-RC2
Data sets
Integrating Total Water Level with Photo Flooding Evidence Dataset Authors/Creators Denali Pinto and Tom Corringham https://doi.org/10.5281/zenodo.18729177
Model code and software
Integrating Total Water Level with Photo Flooding Evidence Code Denali Pinto and Tom Corringham https://doi.org/10.5281/zenodo.18729150
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 789 | 332 | 76 | 1,197 | 92 | 94 |
- HTML: 789
- PDF: 332
- XML: 76
- Total: 1,197
- BibTeX: 92
- EndNote: 94
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
GENERAL COMMENTS
The manuscript “Integrating total water level data and visual evidence to assess coastal flooding in San Diego County” aims to link modeled total water levels (TWLs) with photographic and video evidence of coastal flooding to estimate impact thresholds across locations in San Diego County. The authors use logistic regression to quantify the probability of flood occurrence as a function of TWL and explore temporal patterns, spatial variability, and the potential influence of atmospheric rivers (ARs) and ENSO on coastal flooding impacts. The study addresses an important and timely problem: translating modeled TWLs into meaningful, impact-based thresholds by integrating TWL estimates with visual observations. This approach is valuable and potentially impactful. However, in its current form, the manuscript has several major limitations that hinder its scientific robustness and reproducibility.
The most critical issues are: (1) the absence of clear, consistent, and physically grounded definitions for key concepts throughout the text, such as ‘coastal flooding’, ‘flood impact’, and ‘impact threshold’; (2) a Methods section that is incomplete and does not provide sufficient detail to allow reproducibility, with several statistical analyses introduced for the first time in the Results; (3) a paper structure in which the boundaries between Methods, Results, and Discussion are frequently violated; (4) a tendency to overstate conclusions, especially regarding the physical interpretation of the lagged relationship between ARs and the occurrence of flood-impact photos; and (5) figures that do not effectively communicate the specific results cited in the accompanying text.
Below, I provide detailed comments outlining where improvements are needed.
SPECIFIC COMMENTS
1. Lack of clear and consistent definitions of key concepts - the manuscript does not provide rigorous, consistent definitions of “coastal flooding”, “coastal flood impact”, and “impact threshold”. These concepts are used interchangeably throughout the manuscript or introduced without clear explanation, making it difficult to interpret the results. For instance, it remains unclear whether “coastal flooding” is identified on photos/videos based on water levels reaching a physical morphological feature (e.g., dune) or infrastructure (e.g., seawall) or based on observed impacts such as the need for a cleanup response (as suggested in line 91). This distinction is critical, as it directly affects how photographic evidence is classified and how thresholds are derived. The lack of clarity limits the interpretation of the full manuscript and reduces confidence in the reported results. Clear, physically meaningful definitions should be established early in the manuscript and applied consistently throughout.
2. Methods section is insufficient for reproducibility - a key NHESS criterion is that methods are described with enough detail to allow replication, which is not fully met in its current form. Several critical elements are unclear or missing: the beach slope definition used in the Stockdon runup model is not clearly specified (is it foreshore beach slope, mean beach slope, or satellite-derived tidal datum-referenced slope?); the TWL calculation procedure is not fully described; the social media search strategy is incomplete – what other terms were searched besides San Diego coastal flooding and king tides? Are these use together or individually?; what are flood observations and how different they are from flooding events listed on section 1.1?; the handling of duplicate or ambiguous photo submissions is unclear (line 95 mentions consolidation without specifying how this process is done); and several statistical methods (linear regression, GLMM, and Bayesian logistic regression) are introduced for the first time in the Results section without prior description in the Methods. All methodological steps should be clearly described in the Methods section.
3. Paper structure - the separation between Methods, Results, and Discussion is frequently unclear. Methodological descriptions appear in the Results (e.g., lines 267–272, 273, 306), while interpretation and discussion of mechanisms or prior literature are also included in the Results (e.g., lines 161–162, 194–198, 261–265, 294–295, 330–335). In contrast, the Discussion reads more like a Conclusions section, lacking a summary and a deeper interpretation of findings and research contribution. Reorganizing these sections would improve clarity and ensure a more consistent structure.
4. Overclaiming and overstating results - several conclusions are stated with greater certainty than supported by the analysis. For instance, the repeated claim that modeled TWL and photographic evidence “validate” each other is not fully supported by the analysis – this is explicitly mentioned from lines 337 - 339; the analysis demonstrates correlation, not validation. TWL-max-daytime is stated to outperform daily maximum TWL, but no quantitative comparison is presented among the different regressors used in the logistic regression models, and Figures 7 and 8 show nearly identical time series between these two variables.
5. Ambiguity and redundancy in TWL predictor definitions and temporal aggregation - the distinction among daily maximum TWL, daytime maximum TWL, and lagged TWL variables (e.g., TWL-max, TWL-max-daytime) is unclear and appears, in some cases, redundant. It is not evident how different daytime maximum TWL is from daily maximum TWL, nor why it should be a more appropriate predictor of flooding. If the rationale is that photos are predominantly taken during daylight hours, this should be explicitly stated and justified. Even in that case, a more direct and physically meaningful approach would be to match each photo to the closest hourly TWL value, or to a short time window around the photo (e.g., ±3 hours), rather than relying on aggregated daytime maxima.
Additionally, the definitions of the predictor variables are ambiguous—for example, TWL-daytime and TWL-max-daytime may represent the same quantity depending on how they are constructed. This lack of clarity makes it challenging to interpret the logistic regression results and assess whether the different predictors provide independent information. This issue is further compounded by the lack of clarity on how flooding is identified in the photos (i.e., based on observed impacts vs. water level extent on land), which directly affects how these TWL metrics should be interpreted.
Furthermore, the authors note (lines 203–204) that photo evidence and TWL do not always align. This mismatch may stem from the definition of flooding used. If flooding is identified based on impacts (e.g., cleanup), rather than actual water levels, then timing mismatches between peak water levels and photo capture are expected. Conversely, if flooding is defined based on observed water levels in the photos, then discrepancies with TWL magnitude likely reflect uncertainty in the modeled TWL. This raises additional questions regarding how images are georeferenced and whether a consistent vertical datum is used to relate observed water levels to TWL estimates. Without a clear framework linking observed water levels, model estimates, and thresholds, it is difficult to interpret these mismatches.
A clearer definition of all variables, along with a quantitative comparison of their performance (e.g., daily max vs daytime max vs time-of-photo TWL), is needed.
6. Physical interpretation of the AR–photo relationship is not supported - the manuscript’s TWL formulation (not explicitly stated, but TWLs are generally = tides + nontidal residual + wave runup) could, in principle, include AR-driven sea level anomalies within the nontidal residual component, though the authors do not decompose or formally demonstrate this. More importantly, the main AR result, a lagged GLMM indicating increased odds of photo occurrence following AR events, relates ARs to the likelihood of a photo being taken, not to the physical occurrence of flooding. These are not equivalent. The observed lagged relationship is likely influenced by behavioral and observational biases from researchers and the community (e.g., increased documentation during and after major storms), rather than a direct physical mechanism. This distinction should be clearly acknowledged, and the results should be interpreted with caution rather than as evidence of a causal AR–flooding relationship.
7. Figures do not effectively support or communicate the results - several figures present the data in ways that introduce bias or do not align with the claims in the text, specifically:
a. Figure 3 shows absolute photo counts by month without normalizing for total photo availability, which can bias the interpretation of seasonal flood frequency. It is unclear how many photos were taken each month (e.g., December vs April) and what proportion of those actually show flooding. Without this context, the counts alone are difficult to interpret and may reflect sampling effort rather than true variability in flooding occurrence.
b. Figure 6 presents absolute counts per transect rather than proportions, similarly to Figure 3, making spatial comparisons sensitive to uneven sample size.
c. Figures 7 and 8 show nearly identical time series, yet the text claims one predictor outperforms the other without quantitative support.
d. Figure 9 presents thresholds only for the preferred predictor (TWL-max-daytime), while claims of improved uncertainty relative to other variables are not supported by comparable results.
e. Figure 12 does not clearly convey that the odds ratios represent the probability of a flood photo being taken, rather than the physical probability of flooding.
8. Wave direction is not considered - given the physical importance of wave direction in determining whether wave runup produces impacts at a coastline with a given orientation, the absence of wave direction as a potential explanatory variable for spatial patterns of flooding represents an important gap. Beach orientation relative to the dominant wave approach can be critical for whether runup reaches and exceeds impact thresholds.
CONCLUSION
Overall, as currently presented, the study does not yet quantify physically meaningful flood thresholds, but rather statistical relationships between modeled TWL estimates and the occurrence of photographic evidence. While the approach is promising, the current implementation is limited by unclear definitions, insufficient methodological transparency, and a lack of physical grounding linking TWLs to observed impacts. Addressing these issues is necessary before the results can be interpreted as robust or transferable.