Understanding human–water feedbacks of interventions in agricultural systems with agent based models: a review

The two main methods that have been used to model sociohydrological systems are ABM and system dynamics (Pande and Sivapalan 2017, di Baldassarre et al 2019). In the system dynamics approach, the focus is on the dynamics and evolution of complex overall lumped systems (e.g. a city, population), represented through feedback loops, stocks, and flows, over time and not the micro-level behavior/interactions (Martin and Schlüter 2015, Yu et al 2017, di Baldassarre et al 2019). However, modeling lumped systems misses out on micro-level (e.g. individual farmers) interactions, constraints, heterogeneity, and inequality that give rise to overall system behavior. This also means that inequitable impacts within the population that is at the core of AWM externalities (figure 1) cannot be fully explored.

In contrast, Agent based models (ABMs) can explicitly account for micro-level constraints, individual behavior and their interactions with society and the environment (Berger and Ringler 2002, Berger et al 2006, Troost and Berger 2015, Khan et al 2017). This allows for a natural representation of the real world where social behaviors and dynamics at the macro-level can be attributed to both micro-scale and macro-scale factors (Khan et al 2017, di Baldassarre et al 2019). For this capability, ABMs have been widely used to study the evolution of different systems including land use, urban, forests, ecosystems, epidemiology, social-ecological, and agricultural systems (Le Page et al 2013). This also makes it a promising tool for sociohydrology to understand and explore the evolution of coupled human–water systems, to unravel and understand AWM externalities and resulting in unsustainable and inequitable outcomes. Thus, ABMs can expand and complement conventional AWM model to integrate human–water feedbacks. Table 2 provides some illustrative examples of the strengths of ABMs and how they can expand or complement the AWM studies to capture externalities generated by AWM interventions.

The applications of ABMs in sociohydrology have already begun and are broadening (Michaelis et al 2020, Tamburino et al 2020, Ghoreishi et al 2021). For example, Tamburino et al (2020) developed an ABM to simulate the impact of water use behavior on crop yield and economic gain in smallholder farming systems and how this is influenced by farmer attitudes and behavior. Ghoreishi et al (2021) developed an ABM to study the rebound phenomenon, i.e. increased water demand in response to more efficient irrigation, and its controlling factors in Bow River Basin in Canada.

However, with or without explicit mention of sociohydrology, ABMs have a long history of application in agricultural systems (Berger et al 2001, Berger and Ringler 2002). This includes ABMs for modeling the adoption of AWM interventions (Berger 2001, Schreinemachers et al 2007, 2009), modeling the impact of farmers' agricultural decisions on hydrological systems (Becu et al 2003, van Oel et al 2010) and simulating a range of policy, trade, and market mechanisms (Schlüter and Pahl-Wostl 2007, Farhadi et al 2016, Aghaie et al 2020a, 2020b).

While ABMs have the potential ingredients to capture AWM externalities and applications are increasing in sociohydrology, there is limited understanding of what can be or has been achieved through ABM methodological approaches and what are the remaining methodological gaps that further need to be bridged to unravel the negative externalities of AWM interventions as discussed above (figure 1). With the aim to synthesize the learnings, challenges, and gaps in modeling AWM externalities through ABMs, we here carry out a systematic review of methodological approaches taken in AWM–ABM studies. Since AWM and associated externalities are the focus here, the scope of review is limited to ABM application for modeling AWM interventions. Similarly, other recent reviews have focused more specifically on ABM applications for agricultural policy evaluation (Kremmydas et al 2018), for Food–Energy–Water Nexus (Magliocca 2020) and flood risk models (Taberna et al 2020).

Whether spatially explicit hydrological changes and interactions can be modeled or not depends to a large extent on spatial scales considered and the type of hydrological models integrated/developed in AWM–ABM studies. AWM–ABMs where spatial scale is either individual farm or administrative region (figure 2(a), 27%), is not conducive for modeling hydrological flows and interactions. In these AWM–ABMs, water flows are largely modeled at individual plot/farm levels either using one-dimensional soil water balance (Tamburino et al 2020, Wens et al 2020) or empirical models (van Duinen et al 2016, Zagaria et al 2021). AWM–ABMs with a focus on individual farms are largely concerned with modeling individual farmers socio-economic temporal dynamics resulting from their response, behavior, and adoption of AWM interventions. For example, Wens et al (2020) modeled individual farmers' adaptive behavior, simulated using multiple behavioral theories, to estimate future drought risk in a region in Kenya. In the study, hydrology is modeled at an individual plot scale using The Food and Agriculture Organization of the United Nations (FAO) crop model AquacropOS.

Zoom In Zoom Out Reset image size

AWM–ABMs at an administrative scale in addition to individual farmers socio-economic dynamics can also model spatial dynamics (e.g. crop changes, land-use change, adaptation diffusion) emerging from individual farmers decisions, direct or indirect social environmental interactions (Schreinemachers et al 2007, Barnaud et al 2008, Troost and Berger 2015, Hampf et al 2018). However, hydrology, if modeled, is still mostly modeled at individual farm scales (Schreinemachers et al 2007, Troost and Berger 2015). With hydrological impact not the focus in many AWM–ABMs at an administrative scale, more than 50% of such studies do not employ any hydrological model (figure 2(a)). For example, Troost and Berger (2015)2021 modeled regional land user and crop production dynamics resulting from individual farmer decisions at farm-level to adapt to climate change in a mountainous area in southwest Germany. Water flows were not modeled with the study focusing on analyzing the effect of income, crop changes, and agriculture supply.

Hydrological flows and interactions, via surface and groundwater, can be explicitly modeled in AWM–ABMs where the spatial scale is either watershed/basin (Berger 2001, Becu et al 2003, Schreinemachers et al 2009, van Oel et al 2010, Ng et al 2011) or irrigation systems (Barreateau et al 2004, Schlüter and Pahl-Wostl 2007, Ghazali et al 2018). Overall, 62% and 10% of AWM–ABMs have the watershed and irrigation systems as their spatial scale, respectively (figure 2(a)). In these AWM–ABMs negative hydrological externalities can be captured as agents' actions impact other agents' water flows and availability, simulated as the change in surface water flows (Becu et al 2003, van Oel et al 2010, Pouladi et al 2020), groundwater depth (Noël and Cai 2017, Hu and Beattie 2019, Du et al 2020) and water quality (Pouladi et al 2019). In AWM–ABM modeling irrigation systems, water flows and availability are determined by canal flows, rather than watershed hydrology (Barreateau and Bousquet 2000, Barreateau et al 2004). This is done largely using empirical models (figure 2(a)).

In AWM–ABMs at the watershed scale, both semi-distributed and distributed hydrological models have been used (figure 2(a)). In semi-distributed hydrological models, aggregated hydrological response (e.g. runoff, recharge, drainage) of sub-units (sub-watershed, hydrolgic response unit (HRUs)) is modeled at the overall basin/watershed outlet (Becu et al 2012, Dziubanski et al 2020). Examples of semi-distributed models include using soil conservation service (SCS) curve number method to assess the impacts of land cover changes, aggregated at sub-basin unit, resulting from decisions made by different agent types (Dziubanski et al 2020) or linking hydrologic-agronomic model soil & water assessment tool (SWAT) in Salt Creek watershed in Central Illinois, USA to simulate farmer behavior regarding best management practices and its effect on stream nitrate load (Ng et al 2011). In semi-distributed models, flow at each point/grid is not simulated so they are more useful where the query of interest is assessing the impact on hydrology from the aggregated response of agents. This may limit their utility to assess the impact on individual agents from changes in hydrology, especially when there are significant differences in socio-economic-biophysical capital of farmers in the aggregated units (sub-watershed, HRUs).

In contrast, in distributed hydrological models, hydrology is modeled at each part/grid and can be linked to underlying individual agents. Examples of distributed models are by Becu et al (2003) and Bithell et al (2009) both of which developed spatially distribute models as part of AWM–ABM and linked each point/grid in space with underlying agents. This allows for modeling two-way feedbacks between individual actions/decisions and hydrology. The most often used distributed models in AWM–ABM come from studies assessing groundwater management and sustainability (∼47%, figure 2(b)). In these studies, the use of the distributed modular groundwater flow model (MODFLOW) have been frequent (Farhadi et al 2016, Noel and Cai 2017, Nouri et al 2019). For example, Noel and Cai (2017) developed an integrated ABM–MODFLOW model where farmers' daily irrigation decisions are used as input to MODFLOW which in turn provides updated water-table and baseflow information to agents in Republican River Basin, USA. In contrast, only a few studies (∼10%, figure 2(b)) modeling surface water flows have used spatially distributed models (Becu et al 2003, Bithell et al 2009, Du et al 2020). This could be due to relatively more ease in integrating stock variables (e.g. groundwater head, lake storage) in comparison to output fluxes (i.e. streamflow) in ABMs code (Khan et al 2017).

The examples of negative hydrological externalities discussed earlier show that they often result from interactions of SW–GW systems. Examples include the change in potential recharge from surface storage structures and changes in return flows (as brought on by efficiency improvements practices). Resolving these processes requires that hydrological models should be able to capture SW–GW interactions. However, our review shows that there are large gaps in this part. First, only ∼30% of reviewed papers had considered groundwater (figure 2(b)). Even in these studies, many simulate integrated SW–GW systems in a very simplistic way, such as modeling groundwater irrigation but not process-based recharge and storage modeling (Holtz and Pahl Wostl 2012, Tambourino et al 2020, Wens et al 2020).

Second, very limited studies use integrated models with SW–GW interactions in place (∼14% of studies) (van Oel et al 2010, Du et al 2020, Mirzaei and Zibaie 2021). Most studies model surface water (Nikolic et al 2013, Dziubanski et al 2020) or groundwater (Farhadi et al 2016, Noël and Cai 2017, Nouri et al 2019, Aghaei et al 2020a, 2020b) in isolation. One explicit case of distributed integrated SW–GW model use in AWM–ABMs is by Du et al (2020) where GSFLOW (an integrated SW–GW model) was integrated with an ABM to model water use and understand its impact on hydrology in the Heihe River Basin, China, under the influence of collective water management policies. This lack of inclusion of groundwater and integrated SW–GW process means that many of the AWM externalities cannot be captured or predicted.

Individual decision-making and behavior in AWM–ABMs include taking decisions regarding crop production, irrigation, investment in AWM interventions, and other agronomy aspects (fertilizers, labor, etc). These decisions differ among agents based on the assumptions made about three key determinants of human choices: goals and needs, constraints, and decision rules (Müller-Hansen et al 2017, Schlüter et al 2017). Based on these three key determinants, Schlüter et al (2017) categorized theories used for modeling agent decision making. We use these categories to analyze how frequently they appear in AWM–ABMs (figure 3(a)).

Zoom In Zoom Out Reset image size

Our review shows that the most used theories in AWM–ABMs are rational choice and bounded rationality (including heuristics) (figure 3(a)). Rational and bounded rationality are both based on expected utility maximization where an agent's decision-making is goal oriented. Agents choose a strategy, under given constraints, with the best-expected outcome or utility (An 2012, Schlüter et al 2017). The rational choice theory assumes that agents make rational choices. These rational choices achieve outcomes that maximize their advantage or income by optimizing their decision regarding crops, irrigation, and resource use under given constraints. An example of rational theory used in AWM–ABMs is the application of the mathematical programming-based multi agent systems (MP-MAS) model where farmers' investment decisions are simulated to maximize expected long-term average levels of net farm and non-farm incomes (Berger 2001, Schreinemachers et al 2007, 2009).

However, field evidence suggests that farmers are not always rational (Bluemling et al 2010, Howley et al 2015, Dessart et al 2019). Examples include farmers' unwillingness to covert land to forestry even with expected higher economic returns as that does not align with their attitudes (Howley et al 2015) or the economic cost of increased pumping being an insignificant factor in choosing efficient irrigation technology (Blueming et al 2010). This is because human decisions are complex, and decisions are made under the influence of experiences, rules, psychological factors, and social influences (van Duinen et al 2016, Dessart et al 2019, Du et al 2020).

Bounded rationality theory, a modification of rational choice theory, aims to account for these factors by putting constraints or bounds on the agent's information receiving, understanding, and cognitive capacity (An 2012, Schlüter et al 2017). There are many different approaches to formalize bounded rationality with respect to limited information, quality of information, and cognitive capacities of decision-makers (van Duinen et al 2016, Schlüter et al 2017). The most often used approach is heuristics, where agents are assigned rules, derived from empirical data or observations, that drive their decision-making (Schlüter and Pahl-Wostl 2007, van Oel et al 2010, An 2012). In heuristics, decisions emanate from farmers' experience, accumulated knowledge, and p (Schlüter and Pahl-Wostl 2007). Examples of heuristics include 'if/then/else' rules where agents make cropping decisions based on the predefined threshold such as capital, soil pH, and groundwater levels (Castilla-Rho et al 2015) or sensitivity to crop water stress (Noël and Cai 2017). Though heuristics can mimic an agent's behavior and decisions, it fails to explain the underlying reasons for the same as this is without a strong theoretical basis (An 2012). While this can suffice for modeling behavior to known stimuli/changes/options but has limited utility in case of unexpected and unforeseeable scenarios.

Thus, to drive actual motivations and incentives behind the decisions, there is an increasing realization and call for grounding agent decisions in established social-science theories (e.g. protection motivation theory (PMT), theory of planned behavior (TPB), learning) rather than rational or simple heuristics (Schlüter et al 2017, Taberna et al 2020, Wens et al 2020). PMT, a version of bounded rationality, offers an example (Dziubanski et al 2020, Wens et al 2020, Zagaria et al 2021). In PMT, farmer adaptation is simulated as the integration of farmers' perceived risk and appraisal of their capacity to adapt (Wens et al 2020, Zagaria et al 2021). Wens et al (2020) applied PMT to explore the adaptation decisions of farmers in Kenya. Their results show that bounded rationality can model complex human adaptation decisions more realistically over theory based on rational agents.

In contrast, there is a relatively lower application of other theories in AWM–ABMs, namely the habitual or reinforcement learning theory, TPB, and prospect theory (PT) (figure 3(a)). In Habitual or Reinforcement learning, positive and negative experiences (history) are stored in the state (knowledge) and reflected in the habit formation of agents (Nikolic et al 2013, Schlüter et al 2017, Yuan et al 2021). Castilla-Rho et al (2015) partially include this in heuristics behavior by including 'history' of risk accumulating where agents learn to avoid risky investments. TPB focuses on farmer intention, shaped by agent attitudes, subjective norms, and perceived control, as the main determinants of implementing a certain behavior (Kaufmann et al 2009, Pouladi et al 2019, Yang et al 2020). Pouladi et al (2019) used TPB to assess farmer decisions on the conservation of water resources in the Zarrineh River Basin, Iran. PT takes into account the differences in risk preferences of agents with the idea that people are much more sensitive to losses (risk-averse) and evaluates possible future outcomes differently based on the subjective probabilities rather than objective probabilities (Ng et al 2011, Balbi et al 2013, Ding et al 2015, Gonzalez-Ramirez et al 2018). Ng et al (2011) applied PT to model farmers' crop and best management practice decisions where farmers maximize total utility as a function of their perceptions of future conditions and risk attitude.

Social interactions among individuals play a critical role in influencing individual responses and decisions (Barreute et al 2004, Schreinemachers et al 2007, 2009, Ng et al 2011). The specific sets of individual behaviors influenced by neighbor's decisions and behavior are also referred to as sideward looking theories (Müller-Hansen et al 2017). Agents can interact, observe, or share information with other similar agents (i.e. horizontal interactions) or with higher authorities, governments, markets (i.e. vertical interactions), or both. We focus on the former as the latter act more like constraints or incentives for individual behavior (Aghaie et al 2020a). Of the reviewed papers, only one-third incorporate agent social interactions or sideways looking theories (figure 3(b)).

These AWM–ABMs have used social interactions to model diffusion and adoption of adaptation practices (Berger 2001, Schreinemachers et al 2007, 2009, Ng et al 2011, Schreinemachers and Berger 2011), mimicking of behaviors (such as cooperative or non-cooperative behavior) and decisions regarding cropping practices (Barreteau et al 2004, Nikolic et al 2013, Castilla-Rho et al 2015, Farhadi et al 2016, Cai and Xiong 2017, Ghazali et al 2018, Bazzana et al 2020).

The model of diffusion is based on a principle that agents mimic and learn from other farmers' decisions. Most AWM–ABMs have employed social influence as a model of diffusion (Young 2009), where adoption of practices is modeled as threshold functions. In these models, agents adopt practices or interventions once a certain threshold of the population has adopted them (Schreinemachers et al 2007, 2009, Schreinemachers and Berger 2011, Farhadi et al 2016, Cai and Xiong 2017). Order of adoption between agents is based on agent behavioral values such as innovativeness or risk behavior, which can be either based on empirical data or randomly allocated to agents.

Another model of diffusion used is the contagion model, where agents adopt interventions when they meet others who have adopted them (Young 2009, Holtz and Pahl Wostl 2012). In this model, the diffusion of an innovation is modeled as a self-reinforcing process that tends toward a final saturation level of adopters (Holtz and Pahl-Wostl 2012). For example, Nikolic et al (2013) modeled social interactions where farmers are able to imitate the cropping patterns of neighbors resulting in higher yields during the previous season.

The third type of diffusion mode is the social learning model of diffusion where agents also rationally evaluate, rather than adopting it based on whether others have, the evidence of proposed benefits of interventions generated by prior adopters (Young 2009). The use of social learning in AWM–ABMs is however limited (Ng et al 2011, Daloglu et al 2014, Perello-Moragues et al 2019). For example, Ng et al (2011) used social learning where agent adoption is influenced by variances of the net return on the adoption of interventions, which decreases as more people adopt it.

Extensive use of the social influence diffusion model, with its roots in the study of hybrid seed corn in the USA in the 1940s (Rogers 2004), has been a leading theory of agriculture extension work employed in many international rural development programs and research (AgriFutures 2016). Application of the diffusion model in the field often includes identifying lead or progressive farmers (more innovative or more risk-taking) who are trained or provided support for interventions with the assumption that others will learn and mimic their practices (Tsafack et al 2015, Franzel et al 2019).

However, the application of the theory can be a source of inequity as the expectation that introduced practices will trickle down from lead farmers (mostly more progressive and economically well-off) may not happen (Monu 1995, AgriFutures 2016). This is so because diffusion models often assume homogenous social systems with respect to the introduced technology, which is often not the case (Monu 1995). Empirical field research has shown that the decision making on adoption is influenced by a range of factors including preferences and socio-economic and ecological constraints (Shilomboleni et al 2019), social groups, clans, acceptability (de Roo et al 2019), attitude, cultural norms, and abilities (Kaufmann et al 2009, Daniel et al 2019). Thus, there is a need to internalize and incorporate the wealth of empirical field research and move away from the use of a simplistic threshold-based approach as often done in AWM–ABMs (Kaufmann et al 2009).

Our reviews show there are two broader methods of representing spatially distributed farmers: modeling individual farmers (Berger 2001, Schreinemachers et al 2007, 2009, Arnold et al 2015) and modeling aggregate farmers (Hu et al 2015, Farhadi et al 2016, Hu and Beattie 2019). The latter has also been termed as areal agents by Wens et al (2019). There can also be non-spatial agents such as institutions and markets (Wens et al 2019). These are not reviewed here explicitly as the focus is on farmers or farms, but they are implicit in agents' behavior where they set rules and constraints.

In AWM–ABMs modeling individual agents, agents are assigned to discrete spatial units (e.g. plots, grids) in the model spatial domain where each agent interacts and provides feedback to the underlying environment and hydrological flows (Berger 2001, Schreinemachers et al 2007, 2009, van Oel et al 2010, Arnold et al 2015, Noël and Cai 2017). These AWM–ABMs differ depending on whether the entire population is modeled (Schreinemachers et al 2009, Arnold et al 2015) or only a subset of the population is modeled (Ng et al 2011, Holtz and Pahl-Wostl 2012). For example, Schreinemachers et al (2007) modeled soil fertility and poverty dynamics of all 520 farmers in two village communities in Uganda by dividing the spatial domain into grid cells of area 0.5 ha, corresponding to the size of the smallest agricultural field cultivated in the study area. In contrast, Holtz and Pahl Wostl (2012) divided the farmers based on land size and simulated only 100 farmers per land size class in Upper Guadiana, Spain. Results were extrapolated from this representative population to assess the influence of farmers' characteristics on land-use change and associated groundwater over-use.

Modeling a subset of the population, taken as representative of the total population, limits model runs when the spatial domain is large, saving computational costs. Conclusions on broader dynamics may be drawn from this representative population (Ng et al 2011, Holtz and Pahl-Wostl 2012, Troost and Berger 2015). However, this may restrict the complete representation of all possible spatial and social interactions among the agents. The challenge is also to build the best representative typologies that can explain the farmer's decision/behavior.

In AWM–ABMs modeling aggregated agents, individual agents are aggregated and are represented as one super-agent, over a larger region such as a sub-basin, watershed, or a city (Nikolic et al 2013, Xiao et al 2018, Hu and Beattie 2019, Nouri et al 2019). It is the aggregated responses and feedbacks of agents that are simulated and integrated with biophysical systems (Nikolic et al 2013, Hu et al 2017, Hu and Beattie 2019). For example, Hu and Beattie (2019) modeled 46 counties with each county aggregated as one farmer, Farhadi et al (2016), and Nikolic et al (2013) modeled 13 and 28 sub-watershed/basins, each acting as one independent agent. Aggregation of agents can facilitate practical model development, especially where large basins are modeled. However, aggregated agents limit the model's capability to include local variability and heterogeneity, missing out on equity dynamics within a population (Berger and Ringler 2002). This is critical, especially in an unequal society where the adoption and response to AWM and the impact of AWM externalities could be quite different within the population.

Agent characterization in AWM–ABMs is a way to represent the heterogeneity of a population. Representing population heterogeneity is important to model inequities in cost and benefits sharing and capacities of the agents to adapt AWM practices. Agents are characterized by their socio-economic characteristics, biophysical endowments, and behavioral characteristics. Behavioral characteristics define agent behavior and decision-making and are discussed in the next section.

Our review shows that most of the studies consider socio-economic characteristics of households and farms (family, family composition, household composition, age, sex, area) (table S1). This determines the availability of labor, consumption, and expenses of agents. Other often used socio-economic characteristics, based on the objective of AWM–ABMs, are ownership of assets, machinery, and capital, access to extension services, credit, markets, and off-farm income sources. These all determine the economic, social, and knowledge endowment of agents. The use of a wide range of characteristics already shows the importance and centrality of considering the heterogeneity of agents in AWM–ABM studies. The data for these socio-economic characteristics are either collected from existing microeconomic datasets (obtained from sample surveys, censuses, and administrative systems) (Noël and Cai 2017) or through primary surveys (such as household surveys and focus group discussions) (van Oel et al 2010, Pouladi et al 2019, Wens et al 2019).

Biophysical endowments of agents are mostly derived from underlying maps of biophysical datasets (e.g. soil, elevation, rain). Biophysical endowments characteristics considered (table S1) differ markedly between studies but most consider data on soil type, elevation, precipitation, and irrigation sources. In addition, relative locations of the agents' farms (such as upstream or downstream of other agents, command area, flood plains, etc) have been used to differentiate agents. These data are mostly acquired through secondary data and geographical databases such as cadastral maps, digital elevation models, land use maps, soil maps, etc.

One critical aspect that is of importance while providing biophysical endowments to agents is how agent's location in space is determined. Agent location in space is of paramount importance as this determines their biophysical endowments (e.g. soil quality, water availability), interactions with hydrology, neighbors, and social groups. Our review shows that despite the importance of location, only a few studies use real location data to distribute agents spatially (Schreinemachers et al 2007, van Oel et al 2010, Arnold et al 2015, Noël and Cai 2017). For example, Noël and Cai (2017) use certified irrigated acres from the existing database on pumping wells to delineate the agents. The results of our review are similar to the conclusion of Kremmydas et al (2018), who found that only 2 of the 32 reviewed papers used observed location data.

Despite AWM interventions being intricately linked with hydrology (section 2), our review shows that a quarter of AWM–ABMs simulate dynamics at individual farms or administrative regions (figure 2(a)) where spatial scale is not conducive to model hydrological interactions. In these studies, the subject of inquiry is not hydrological changes but dynamics such as emergent land use, adoption of interventions, and changes in the cropping system. Given that AWMs are intricately linked with hydrology, the simulated dynamics can cause hydrological externalities leading to unsustainable and inequitable outcomes. Thus, there is a need to supplement/complement these studies with hydrological models to account for and predict any negative hydrological externalities.

Additionally, even in AWM–ABMs with the capability to model water flows (i.e. the spatial scale of the watershed, and basins), methodological gaps limit their capacity to completely resolve the hydrological externalities of AWM interventions. This includes a lack of incorporation of spatially distributed models, limited inclusion of groundwater systems, and almost non-existent integrated SW–GW models (figures 2(a) and (b)). Spatially distributed hydrological models are required to capture the spatial heterogeneity of both biophysical systems and agents in the region and capture spatially explicit hydrological externalities of AWMs. The lack of spatially distributed models means that the impact of hydrological changes on individual farmers and vice versa cannot be modeled. This limits the capability of AWM–ABMs to resolve inequitable impacts. Additionally, the non-inclusion of the groundwater system and lack of integrated SW–GW limits AWM–ABM capability to capture the holistic hydrological impact of AWM interventions that often leads to reallocation/changes within SW–GW systems.

Our review shows a clear need to enhance the representation of hydrological systems in AWM–ABMs if they are to be used to assess the negative hydrological externalities of AWM interventions. This requires coupling ABMs with spatially distributed and integrated models. This can be done by developing hydrological models as part of ABMs or coupling ABM code with existing open-source models (e.g. GSFLOW, spatial processes in hydrology (SPHY)). An example of the latter is by Du et al (2020) where GSFLOW, an integrated SW–GW model, was tightly coupled with ABM at the source code level.

A realistic representation of individuals' behavior and interactions forms the basis of modeling society's unexpected and emergent dynamics. This requires a suitable, accurate, and dynamic representation of agent behavior and decision-making processes. Though a range of farmer decision-making behavior has been simulated, there remain gaps in terms of incorporating appropriate behavioral theories in AWM–ABMs. The empirical field research has shown that human behavior is shaped by a range of factors such as socio-economic, cultural norms, risk attitudes, perceptions, and other psychological characteristics (Kaufmann et al 2009, Daniel et al 2019, Pouladi et al 2019). However, there is a large gap in incorporating the same in AWM–ABMs. Our review shows that the use of rational choice theory and simple heuristics is still dominant (figure 3(a)). The rational theory assumes agents make rational choices and discount the impact of a range of factors such as socio-economic, cultural norms, risk attitudes, and other psychological characteristics or both. Similarly, farmers' heuristics devised based on experience, accumulated knowledge, and preferences lack the theoretical background to explain the underlying reasons for the same.

There is limited but increasing use of theories grounded in social science and field research to account for these constraints (e.g. PMT, PT, and TPB). There is a greater need to formalize these theories in AWM–ABMs. A general lack of sufficient and good-quality primary data on agent behavior makes derivation, validation, and verification of agent behavioral rules difficult (Hu et al 2017). Multiple studies have shown that this can be done with primary data collection through surveys or focus group discussions (Kaufmann et al 2009, Pouladi et al 2019, Wens et al 2020). Additionally, there is a need to incorporate further behavioral models such as risk-, attitude-, norm-, ability-, self-regulation- (RANAS) model originally developed for the water, sanitation and hygeine (WASH) sector (Mosler 2012). RANAS combines multiple important behavioral theories (including the TPB) to explain and change behavior and can be adapted to a range of situations and already provides a standard template of questions to quantify behavioral factors and analyze the behavior (Callejas et al 2021).

Another gap in AWM–ABM studies is the lack of incorporation of social interactions among agents (figure 3(b)). In limited studies where social interaction is in place, social interaction is largely modeled simplistically following simple thresholds or contagion-based diffusion model approaches. These approaches assume agents adopt interventions or behaviors once certain other people have adopted, or they come in touch with someone who has (Schreinemachers et al 2007, 2009, Ng et al 2011). These are found to ignore a range of factors influencing adoption, including preferences and socio-economic and ecological constraints as has been showcased in multiple empirical field research studies (Kaufmann et al 2009, Daniel et al 2019). Thus, like individual theories, there is a need to expand the AWM–ABMs social interactions theories in use, employing more holistic adoption and diffusion models.