Thank you for inviting me to comment on the manuscript: “Adaptive Behavior of Over a Million Individual Farmers Under Consecutive Droughts: A Large-Scale Agent-Based Modeling Analysis in the Bhima Basin, India”. I am an expert on agent-based land use change models. This may limit my expetise on the hydrological aspects of the presented work.

The authors use two coupled models (one ABM and one hydrological model) GEB (please spell out what GEB stands for) to model the land use of presumably 1.4 million farmers in the face of consecutive droughts over several decades. Unfortunately, the text is very rich combing description of drought modelling, theory, ABM model, and many more things. Therefore, at least for me it is impossible to understand the model(s) in detail and to put the results into context. Therefore, I am not able to  appreciate the model and its results sufficiently, although the topic is timely and I guess the approach is relevant and promising.

Maybe the use of a protocol such as ODD+D (Mueller et al. 2013, Describing human decisions in agent-based models - ODD+D, an extension of the ODD protocol) could help to present the model in a more digestible way. At least the authors should summarize somewhere (maybe as a table) an overview of the properties of the agents and or agent types. The authors present many responses from their model: “Our analysis examines how these adaptations affect profits, yields, and groundwater levels, considering, e.g., farm size, risk aversion and drought perception.” Maybe it would be helpful to reduce the number of responses and/or scenarios? Especially in Figures 5 and 7 I would encourage the authors to present less panels and to focus on a more narrow narrative.

I appreciate that the authors follow a theory to justify their decision model. However, the authors need to guide the reader more carefully, since the SEUT is not a standard theory for all Economists and certainly not for all land use change modellers. Fishburn (1981) is a review of several theories and the list of papers suggested as examples of the application of SEUT needs to be critically revised (Groeneveld (review), Haer and Wens do not mention SEUT and do not cite Fishburn). When SEUT is introduced it also should be mentioned that the authors use imitation and “bounded rationality” (line 215) as well in their decision modelling. Later on also prospect theory is considered.

In general, I am missing an argument why it is useful to model > 1 million agents. Other authors decided to gain knowledge by aggregating actors to agent types in their land use models (e.g. https://landchange.imk-ifu.kit.edu/CRAFTY or the work by Millington et al. https://www.jasss.org/11/4/4.html). It would be great to see an argument developed why this computational demanding approach is seen more appropriate to answer questions of land use change. This is especially critical since the authors argue that it is not computationally feasible to compute for all agents the SEUT for all 300 options (“unique crop rotations”). Would it make sense to compute less agents and therefore consider all 300 options?

I have difficulties to understand the results. To my understanding imitating the strategy of more successful agents in the neighbourhood of an agent is at the heart of the presented dynamics. In the abstract this is not mentioned: “n adaptive scenarios, farmers can either do nothing, switch crops, or dig wells, based on each action’s expected utility.” If my reading is correct the imitation aspect should be mentioned early on and should be discussed in a diffusion of strategy/technology context. What are the updating rules – synchronous or asynchronously? How many neighbours are considered? What is the initial trait distribution of actors. Is there a spatial structure in the inital trait distribution? Is the number of farmers constant over the years? Form Figure 4 it seems that the model does not show much variation between runs. Would have been less agents sufficient? What source/help of A.I. have the authors used for what?

The role of the “spin-up” period (21 years) needs to be explained in more detail. The model is initialized with data from some point in time (when). Given the substantial temporal dynamics of the responses (see Figure 4) the choice of the length of the “spin up” period should have a strong effect on the results? It is written that the calibration has been done in the period after the spin up from 2001 to 2010? It is difficult for me to understand the evolution of Figure 4. The starting point at 2001 is the result of the spin up period? The period from 2001 to 2010 is calibrated and after that it is the model projection? How are small, medium, and large field farmers defined in terms of hectare?

Thus, overall I have the impression that potentially great insights are hidden in the current text. More specific and potentially less research questions could help to narrow down the story to allow easier access to the main highlights of the study. And at least for me it would be necessary to have a clearer motivation why it is beneficial to consider so many agents at the very same time.

Specific comments

Abstract: “realistically simulate” – That is maybe personal but I would avoid phrases like “realistically simulate” since it is a model and the best one can do is to model something useful in respect to the research question.

The models are written in Python?

Lines 73/74 What means “one-to-one scale”?

Lines 78: What “simple assumptions of human behaviour”?

Figure 1: it is not clear that some boxes are empty – please explain.

Line 101: reservoir operators – is this agent type considered in the presented study? “reservoir operators” are never mentioned again.

Figure 2: The land cover map presents less classes (e.g. agricultural land) than used in the results? How did the authors discriminate the different crop types?

Lines 132ff – Can you define in term of your indicator (SPEI) what a severe, moderate and so on drought is? I guess there are thresholds? Please specify.  

Line 156: Why 5 km radius – is this decision based on a sensitivity analysis?

Line 165: C_adapt is not considered in equation 4. Does it needs to read “C_input in eq. 4 on current market prices”?

Equations 1-4 please explain subscripts x, d, and m.

Lines 197: crop costs – crop costs C_input are dependent on the type of crop? If so how?

Equation 9: Where is the “crop coefficient” Kc used? The duration of different harvesting stages are not crop-specific?

“Agent initialization”: Spell out IHDS, the authors should give the ranges of the agent properties. Showing boxplots or other useful representation of the distributions of attributes would be useful. I have no intuition for example how net income is initially distributed among the 1.4 million agents and how it is developing over time.

Line 326: (7g)?

Figure 4c: Is it reasonable that one crop is going from 0.05 fraction to the dominating crop?

Figure 5: Very busy graph (6 panels). The authors may want to focus on some panels.

Figure 6 and elsewhere: Specify the unit Rs

Figure 8: Too many panels. Legend unreadable. Please focus on the important aspects.

Technical corrections

The figure labels are too small and therefore hard to read.

Figure 4c and others: colours are too similar – hard to differ crops

Lines 65-66 and elsewhere: Inconsistent in-text citation of Udmale et al. 2014/2015

Line 114 and elsewhere: “95 %” -> “95%”

Line 263: The authors refer to figure 3 in Jun et al. 2014? Jun et al. 2014 is a comment in Nature without Figures to my understanding. Please check.

“Sensitivity Analysis”: What are 300 distinct samples. Sampling from what distribution?

Line 306: Where does stochasticity enters the model(s) and how?

Thank you for the invitation to review this manuscript. In this work, the authors extend the GEB, a coupled agent-based hydrological model, with the Subjective Expected Utility Theory and apply the model for analysis of the Bhima River basin in India under consecutive droughts. The manuscript is impressive for the complexity of model integration and the breadth of analysis conducted. I especially commend the authors for the extensive sensitivity analysis that is conducted using the model, which is often a critical gap of coupled human-water systems analyses. However, the extensiveness of the manuscript is a double-edged sword, with the manuscript very challenging to wade through given the sheer amount of material (as reviewer #1 also noted). In this sense, I reiterate reviewer #1’s comments in regards focusing the analysis. I have additional comments in regards to the manuscript:

1.    My first and foremost comment is that the authors should demonstrate the validity and reasonability of the model in relation to real-world observation / understanding. While I understand that a full-scale, spatiotemporal validation of the model isn’t likely possible given the sparsity of real-world observations and the complexity of the model, one can still ask the question: does the model better capture real-world patterns of the complex system in comparison to alternative approaches (e.g., the no adaptation alternative). For example, model results indicate that there is a very significant uptake in groundwater wells for large farms (growing from 30 percent of farms to 65 percent of farms) over the course of the model run. Is there any real-world quantitative or qualitative data that supports these model results? The onus in this case would be demonstrating that the adaptive model outperforms the non-adaptive model in replicating these large-scale patterns observed in reality. Similarly, do we in reality see the significant increases in groundwater depletion associated with the adaptive behavior (~10 meters in relation to the non-adaptive version); I would imagine that even apart from point groundwater level measurements, such a stark difference in depletion could be corroborated by GRACE, or even other qualitative sources. Cropping patterns are another example, the adaptive model shows large-scale crop switching that could likely be corroborated, in a broad scale sense, via agricultural census information or remote sensing data. While the modeling integration and advances are impressive, there are so many choices that are made in regards to theory and implementation (as is the case with nearly all coupled human- natural models), that it becomes nearly impossible to assess the value of these model improvements in the absence of such evaluation.

2.    As I understand, the region is also heavily managed in regards to the surface water supply system (reservoirs, diversions, manmade canals, etc.), which influences water availability for irrigation and associated demand for groundwater and farm decisions to install a groundwater well. Can the authors speak to the capabilities or limitations of CWatM in effectively representing surface water deliveries for irrigation in this region and how this may be influencing results?

3.    In this discussion, the authors note that groundwater well drilling is potentially maladaptive, as farmers then rely on wells that can go dry during subsequent droughts. These are important findings that seem to be largely glossed over in the results section. For example, there isn’t a figure reporting on the drying of these wells during subsequent droughts.

4.    It would seem to me that the imitation technique (described in lines 155-156) would very quickly lead to homogenization of crops across farmers using the same irrigation technology. Is this not the case? Could the authors further comment?

5.    I understand that the political economy of sugarcane is particularly influential on water security outcomes in the region (e.g., https://iopscience.iop.org/article/10.1088/1748-9326/ab9925/meta). Could the authors speak at all to how such considerations factor into the analysis? More broadly, crop prices are a significant driving factor of farm behavior, but the subjective expected utilities are only formulated in relation to subjective drought perception. Can the authors comment on whether/how farmer perceptions of economic conditions might influence results (even if outside the scope of this analysis)?

6.    The above article is conducted as part of the Stanford FUSE project, which was an outgrowth of the Stanford Jordan Water Project (JWP) which also introduced a coupled agent-hydrologic model for similar types of analysis (e.g., https://www.nature.com/articles/s41893-023-01177-7; https://www.pnas.org/doi/abs/10.1073/pnas.2020431118). While much of this work was focused in Jordan rather than India, these are important studies to note as part of the literature context. Can the authors speak more to how the current effort relates to and is distinguished from this line of coupled agent-hydrological model?

7.    Figure quality throughout could be improved. Resolution is often poor with text difficult to make out and colors often hard to distinguish (e.g., couldn’t distinguish crops in the cropping figs). Fig 1 is also difficult to interpret and missing text in boxes.

8.    Lastly, I agree with reviewer #1’s comment regarding the >1 million agents. Even if such # of agents is warranted, headlining the # so prominently throughout the paper (in title, abstract, etc) in my opinion misplaces focus and potentially signals the wrong message (e.g., model complexity for the sake of model complexity). This ability to model of large # of agents was already heavily featured/highlighted in the original GEB paper, so in this case I’d rather see the spotlight placed on the insights drawn from the modeling improvements and analysis, rather than the # of agents that can be modeled.