| 20 Aug 2026
Compound Flooding in Surabaya: A Multi-driver Sensitivity and Interaction Analysis
Abstract. Coastal compound flooding, caused by interactions between various flood drivers, presents serious threats to urban coastal regions. In this study, a process-based modelling framework was applied by integrating statistical analysis of historical data with reduced-physics numerical simulations to assess compound flood hazards. A very low correlation between extreme sea level and heavy rainfall events in the study area was demonstrated, which confirmed the initial assumption of flood driver independence. Single-driver and compound-driver scenarios corresponding to a 100-year return period were simulated for present-day and future conditions, accounting for relative sea-level rise and intensified extreme rainfall under SSP2-4.5 and SSP5-8.5 by 2100. Results indicate that non-linear interactions between drivers lead to a considerable amplification of flood hazards. The primary impact mechanism is increased inundation depth within existing floodplains due to backwater effects, which affects 8.6 % to 20.9 % (10.7 to 36.6 km²) of the total inundated area. A secondary effect is the expansion of the flood extent into previously unaffected areas by 0.82 % to 1.1 % (1 to 1.7 km2) of the total inundated area. Climate change is projected to exacerbate these hazards substantially. Under the high-emission SSP5-8.5 scenario, a coastal compound event with a current 100-year return period is projected to occur as frequently as every 2–3 years by 2100. This dramatic increase in frequency, along with a substantial rise in flood intensity, is primarily driven by sea-level rise.
Referee report on egusphere-2026-4592
Compound Flooding in Surabaya: A Multi-driver Sensitivity and Interaction Analysis
Larasari, Reyns, Nawarat, and Dastgheib
Recommendation: major revision. A rejection would be defensible on the probabilistic framing alone (points 2.2 and 2.3), and I would not object if the editor took that route. I recommend major revision because the modelling effort is substantial, Surabaya has no prior process-based compound flood assessment, and the fix is a re-analysis and a reframing rather than a new model.
1. General comments
The manuscript combines a peaks-over-threshold analysis of daily rainfall and GTSM-ERA5 storm surge (1981-2024) with SFINCS simulations of five design events, then repeats the exercise for 2100 with sea level rise and rainfall intensification. It is clearly a lot of work, and I want to believe it was done carefully. As a reader, though, I finish the paper asking what I learned. There is no entry point and no story. The introduction is thin and, in places not correct; earlier flood work in Surabaya is cited but never discussed, and the conclusions restate the scenario inputs more than they interpret the model. It reads as a good MSc thesis, not a journal article.
Three problems I see:
1) What is done is scenario modelling, not compound flooding. Running a 5-year rainfall and a 20-year surge at the same instant on the highest astronomical tide and calling the result a 100-year compound event is not the same thing as quantifying compound flooding. The variability in timing, which is what makes compound events compound, is removed by construction. Compound interaction is found by simulating actual events and switching processes on and off (see, for example, Eilander et al. 2023). The current setup can only say which of three arbitrary pairings is worst under peak-on-peak alignment.
2) The coastal boundary condition dwarfs everything else. Surge is added on top of a fixed HAT in every scenario, including the rainfall-only run. The surge return levels span 0.26 m (1 year) to 0.38 m (100 years). That is why the three compound scenarios return near-identical results (Fig. 5). You cannot superimpose a surge on HAT and keep the dynamics. A water level analysis, with tidal probability included, is required.
3) The model is not validated in any quantitative sense. Appendix B compares four events against social media footage. I understand that local water level data is scarce, and using photos and video is the right instinct, but it has to go further than Appendix B. For example, estimate depths at the photographed sites, a hit rate or CSI, etc.
Beyond these, the manuscript contradicts itself on its central result (Table 5 versus the text and Figure 9), key inputs are asserted without evidence, and the writing and several figures are below standard.
2. Specific comments
2.1 Novelty and framing. The abstract presents "a process-based modelling framework integrating statistical analysis with reduced-physics simulations" as the contribution (L12-13). This is standard practice, formalised in this journal by Eilander et al. (2023). What is actually new, a first process-based assessment for Surabaya and the drainage window mechanism in Section 5.3, is never claimed. Rewrite the introduction and conclusion around those. Smaller points: "strategically positioned" (L61), strategic for what? "The area experiences a tropical climate" (L64), rephrase. The drainage network, which controls the whole result, is described only by reference to two theses (L72-73); describe the actual gates, pumps, and retention basins.
2.2 Joint return period (Section 3.6). The product of marginal return periods is the annual AND probability of two annual maxima, not the probability of the coincident event that is simulated. Under the authors' own independence assumption and a plus or minus 1 day window, coincident exceedances of two 10-year events occur at a rate of order 5e-5 per year, not 1e-2. Section 5.5 admits the phasing is "an artificial maximization" (L599-601) but does not draw the consequence: the 100-year label on R5W20, R10W10, and R20W5 is not supportable, and neither is the conversion back to a frequency in Section 4.3.
2.3 Coastal boundary (Section 3.4). Table C1 shows that every "extreme sea level" is 2.07 m plus a surge return level. A 2.33 m level does not occur every year, so the return periods in Table 2, Table C1, and Figure 10 are surge return periods, not sea level return periods. Also: "total water level" (L140) excludes waves here, so call it still water level. GTSM-ERA5 resolves neither steric variability nor waves; say what that means for the Madura Strait. The "amplification method" (L139-140) cites Xu et al. (2024), which is not in the reference list and is not described. Figure 3 has no axis values. Section 2 describes a mixed semidiurnal regime, but Figure 3 shows one high water in 24 h, and the drainage window argument in Section 5.3 depends on which is imposed.
2.4 River discharge (L198-200). A fixed 100 m3/s for the Surabaya River and 10 m3/s for the Lamongan River is given with no source, and the inflow points are named differently at L174-175 (Wonokromo) and L198-200 (Lamongan). Assumptions are unavoidable, but every grain of evidence must be found to defend them: a bankfull estimate from the channel geometry, a gauge, or the diversion operating rules.
2.5 Extreme value analysis (Section 3.2). The shape parameter is fixed at zero without justification. Report the free shape fit, confidence intervals on the shape and on the 100-year levels, and the sensitivity to the threshold. And an R2 on a Q-Q plot is not a goodness of fit test.
2.6 Dependence analysis (Sections 3.3 and 4.1). Kendall's tau on paired magnitudes, conditioned on rainfall, does not test whether joint exceedances occur more often than independence predicts. Table 4 shows tau rising with the threshold to 0.20 (p = 0.016) at the 99.5 % quantile, the one closest to the design events, which the text then dismisses. Report a tail dependence diagnostic (chi, or observed versus expected joint exceedance counts), test seasonality, and explain how n is counted. The abstract's "confirmed the initial assumption of flood driver independence" and the Conclusion's "statistically significant, very low correlation ... there is no common meteorological event" (L614-617) are both too strong and contradict each other.
2.7 Climate scenarios (Sections 3.4 and 3.8). Vousdoukas et al. (2018) is an RCP4.5/RCP8.5 study; the text labels the values SSP2-4.5/SSP5-8.5 and the Figure 10 legend still says RCP. State which component and which percentile. The values are called "relative" throughout, yet subsidence is excluded (L585-587, L643-645). With the manuscript's own rates of 0.2-83.3 mm/yr, subsidence plausibly exceeds the SLR term. Drop "relative" or add a subsidence case. The rainfall factors +17.44 % and +23.61 % come from Indonesia-wide seasonal indices (R95p, Rx5day) in Kurniadi et al. (2024) and are applied to a 6-hour storm; justify the transfer and drop the false precision.
2.8 Table 5 contradicts the text. The text (L362-374) attributes 11.9, 30.2, and 36.6 km2 to the coastal-dominated scenario; Table 5 gives them to the rainfall-dominated. Figure 9 and Table 6 support the text, so the table rows appear inverted, but the authors must confirm this from the source data because Sections 5.2-5.3 and the Conclusion rest on it.
2.9 Interaction metrics (Section 3.7). "Amplification" is tidal terminology; use something like compound-only area and depth increase area. The metric compares a (5-year, 20-year) compound run against 100-year single-driver runs. That answers which combination is worst, not whether the interaction is nonlinear, as the text claims. Nonlinearity needs the compound run compared with the sum of its own single-driver components. The "total inundated area" in Figure 5 is identical across compound scenarios.
2.10 Model setup (Section 3.5). The 0.4 m DEM has no source, date, datum, or status, and how the DEM relates to the GTSM levels is not described. State whether open water and ponds are masked (Figure A1 renders the strait as flooded). Manning and curve number values are not tabulated. The rationale for removing the major gates (L180-184) is inverted: gates close at high tide to keep the sea out and open at low tide to drain. Removing the structures that exist to prevent backwater, in a paper whose finding is that backwater controls the hazard, must be framed correctly. Rainfall is uniform over roughly 450 km2 in a regime the authors call convective.
2.11 Writing and figures. The writing is broken in places (for example, "based on strictly by the digital elevation model", L497; "would expectantly demonstrate", L457) and needs a full language pass. Replace "dramatic", "substantially", and "confirmed" with numbers. Figure 2 is not legible in Task 1. Figure 6 cannot be read, since the compound area it shows is under 1 % of the extent. Figures B1 and B2 are web viewer screenshots with an undefined datum.