General comments

The value of Agent-Based Modelling as a technique to understand the role of individual decision making in response to hazards, such as flooding, has increased in recent years, since the initial publication of ABM evacuation behaviour was published in 2011. This study is an interesting example of further development of this application of ABM and demonstrates a good synthesis of empirical research to formulate agent behaviour. Through introducing agent characteristics such as age, gender, weight, variable movement speeds and stability patterns, the model introduces an enhanced level of heterogeneity amongst the agent population from previous iterations of the model. 

Major comments

However, there lies two fundamental concerns related to the outputs of the model:

1. It would appear that uncertainty inherent within modelling stochastic agent behaviour is not acknowledged. The impression given is that model simulations were ran only once for each scenario (lines 331 – 333) and outputs were then analysed. There should be evidence of multiple simulations for each test scenario to account for variance in behaviour of agents. For example, confidence intervals and averaging over an appropriate number of simulations would provide a more representative set of outputs that account for stochasticity.
2. The number of agents included in the Hillsborough case study is 4,080. The lowest recorded attendance for Sheffield Wednesday in 2019 was counted at 21,485, meaning that the agent population is only 18% of what is considered ‘real’. Given that the start of the paper states that the FLAMEGPU platform can handle “as large population size as needed (line 79)” and that throughout the paper congestion is frequently stated as a factor that influences flow dynamics and evacuation time, a question emerges as to why such a significantly low agent population has been adopted. The paper states that “using a bigger population size would lead to extreme pedestrian congestion that impacts the movements of individual pedestrians (lines 489 – 490)”. Surely, in the context of a football match, this is an important factor that must be represented as accurately as possible. There is no justification for the chosen figure of 4,080 and it seems arbitrary. I am not wholly convinced that subsequent outputs reflect evacuation times that would be representative of a football match. Either a more realistic agent population in line with actual attendance rates is necessary, or justification that the emergent behaviour is not dramatically impacted by the chosen number of spectators.

Technical/Minor Comments:

• In section 2.3.2, it states “…the more crowding of pedestrians the more energy loss in the floodwater dynamics for low risk floodwaters, which in turn enables the pedestrians located behind to take a faster moving speed (lines 346 – 348)”. Whilst it is well noted that crowding would disperse the floodwater, making pedestrian movement easier, I can’t help but wonder if collective moving speed would in turn be slowed down by the congestion itself, therefore this could be acknowledged.
• In section 2.3.2, it also states “enabling it with the two-way condition (Mode 2) increases slightly the evacuation time as crowding is more likely under the walking condition (lines 435 – 436)”. It would be reasonable to assume that the ‘running condition’ would be more likely to cause crowding as it represents a more ‘excited’ response, causing further bottlenecking in the evacuation process compared to more organised walking agents. Some consideration of this would be insightful.
• The model demonstrates well a fundamental concept inherent within complexity sciences and Agent-Based Modelling; emergence. The adoption of risk perception thresholds seem to have influenced the favouring of destination selection by agents with varying levels of risk perception. As is the case for risk perception and adaptive behaviour in response to flooding (i.e. risk perceptions thresholds increase after an individual’s property is flooded, which increases the likelihood of adopting protective measures in future) as an active area of interest within flood risk management. Some further comments on the overlap between risk perception and evacuation behaviour may be insightful to lay foundations for future research.
• Some more clarity on what solvers are used to dictate the flow of water in the hydrodynamic model should be provided in Section 3.2.1.
• Calibration/validation evidence of the hydrodynamic model should be provided in more detail to highlight the robustness of hydrodynamic outputs. Two separate figures are provided of flood extents; an Environment Agency figure (Figure 8) and your own (Figure 10). These could be combined into one figure, by overlaying and analysing to provide an FSTAT value to indicate accurate representation of simulated outputs versus simulated. Similarly (at a minimum), a figure showing simulated and observed hydrographs should be provided to demonstrate that flood depths are within an acceptable tolerance.
• A supplementary table providing a summary of all agent variables and values upon initialisation (flooding, navigation and pedestrian agents) would be useful. This would promote replicability and reproducibility.

The paper “Flood-pedestrian simulator for modelling human response dynamics during flood-induced evacuation: Hillsborough stadium case study” aims at showing the effect of human-body characteristics and in-model behavioural rules when included in an ABM integrated framework for flood evacuation modelling. The model is largely based on the previous version developed by the authors, and it is tested on a synthetic case study and on a real-world case. The topic of this study is definitely highly relevant and timely. However, I think that the study could be further strengthened by assessing the effect of the evacuation time and characteristics of the inflow hydrograph on the evacuation process. These factors could have a higher impact than the human-body characteristics considered by the authors for flood risk mitigation purposes.

- It has been recently shown that the timing in which the evacuation is issued is crucial for reducing flood risk (Alonso et al., 2020). However, in section 2.3.1 the authors state that “When the floodwater starts to propagate over the walkable area, simulation time (t) of 0 min, the pedestrian agents start the evacuation …” In a no-flooding situation, agents are randomly moving based on their behaviours and on their daily routines. Why an agent should start moving exactly when the water starts entering the building and not before or after? Are there supporting evidence to justify such an assumption? Then, when flooding occurs, there can be two extreme situations. On the one hand, if the agent is doing something else it may not notice the flooding until the pedestrian agent goes on a flooded area. On the other hand, the agent could be informed earlier about a coming flood and start to evacuate earlier. These two scenarios could have dramatic consequences. I suggest including more modes (table 6) accounting for different evaluation timing. This study could show the role of human-body characteristics and in-model behavioural rules in reducing flood risk when evacuation is issued late.

- My second concern relates to the shape of the hydrograph considered in the synthetic experiment. I understand that using a hydrograph with a high flood peak would lead to significantly bigger HR and the consequent loss of life. However, is it not the scope of this model to represent worst-case scenarios to improve flood risk management and reduce loss of people? Also, not necessarily using a shorter hydrograph can lead to loss of life. I invite the authors to run different scenarios keeping the same volume of the input hydrograph but changing the timing of the peak. The timing of the peak is a crucial factor in any flood risk management application and I do not understand why its influence was not included in this study.

- Section 2 is a very large and dense section. There are many headings and sub-headings and I found myself lost with a need to scroll up and down. Would not be better to move 2.3.1 and 2.3.2 in a new section 3 called synthetic case study used to test the model and then introduce section 4 (now section 3) on the real-life experiment?

Reference:

Alonso Vicario, S., Mazzoleni, M., Bhamidipati, S., Gharesifard, M., Ridolfi, E., Pandolfo, C., & Alfonso, L. (2020). Unravelling the influence of human behaviour on reducing casualties during flood evacuation. Hydrological Sciences Journal, 65(14), 2359-2375.

It is an important topic and well within the remits of the journal. They have used a single agent-based modelling (ABM) framework running on Graphical Processing Units (GPU). However, in the review’s view, further clarification and revision is required before paper is suitable for publication.

• This paper is a follow up from previous papers from the authors and the authors need to clearly state what are the novelties and to demonstrate clearly the effects of these novelties.
• There have been recent experimental and numerical studies which has shown improved HR assessment, e.g. Martinez-Gomariz et al. (2016). Why the authors used Equation quoted in line 133 for flood Hazard Ratio while they used more recent studies to evaluate flood HR?
• Flood Hazard Ratio introduced in this study (line 133) also include a debris factor which has a significant impact on the HR. This is missing from the equation. what is the significance of removing this from the equation?
• In some places the paper is confusing. In particular, section 2 where there are many references to the other sections.
• Justification to use Mode 2 with the simulator to plan evacuation case study involving severe flood is based on the lack of major changes as a result of implementation of Mode 4 (line 394). This seems to be insignificant and based on 1 simple case study. The reviewer expects a more detailed analysis for such decisions.
• The authors provided a good literature review of the topic. However, their methodology and novelty only compare with their previous study, e.g. use of BMI (line 170), age and gender (line 10) are considered as novelty when compared to Shirvani et al. (2020). However, these have been used in flood hazard assessment previously.
• The paper seems too long in reviewer’s view. It can be more concise by shortening some of the text which are less relevant to the novelty of this paper. Some of the figures and tables are also reproduced from other publications, e.g. Figure 2 and Table 1 and 2. Removing those and just citing them seems more appropriate. This will reemphasise the point about the overlaps between this and other publications.
• It seems the characteristics are assigned randomly to the agents (line 292). What is the significance of such random assignment? For instance, how repeatable the simulations will be if they are repeated and what is the significance of this repeatability.
• What is the significance of including conditions from both toppling-only and toppling-and-sliding (paragraph starting line 402)? Furthermore, a case study considered a shopping centre where the floor material expected to be different to the conditions where stability experiments were conducted. Furthermore, the slope and the direction of walking has impact on stability. The authors are expected to consider this somehow.
• Similarity between the predicted flood inundation levels and those of the EA flood maps were used as the validation of the model. It is stated that “the generated inflow hydrograph is able to replicate realistic floodwater depths and extents within the walkable area” (line 531). It is important to ensure that both simulations have the same conditions to be able to draw such a conclusion. For instance, is the EA flood maps produced under similar hydrographs? Furthermore, velocity is as important if not more important than the water levels in case of evacuation. It is important to show the velocities across the domain and provide reader with an assessment of uncertainties associated with such predictions.

Minor comments:

• What was the reason behind the orientation of the model used in this study? Was it not more efficient to have y-axis parallel to A61?
• Line 152: what is n_M?
• Equation 2: where is the source of the equation and calculation of M? It might be better to be cited in the same way as Equation 1.
• It will be useful to also include whether the discharge was added only as mass or there is also momentum (velocity) associated with the flow.

Reference:

Martínez-Gomariz, M. Gómez, B. Russo, Experimental study of the stability of pedestrians exposed to urban pluvial flooding. Nat. Hazards, 82 (2016), pp. 1259-1278, 10.1007/s11069-016-2242-z