*3.3. Influence of Structural, Material and Psychological Factors*

3.3.1. Regression Models Converged for Six Major Variables

The linear regression model selected 12 of the 24 variables to explain the date of first irrigation (Table 5), while the regression tree contained six variables for agronomic practices, irrigation practices, reactivity and assistance. Three of the six variables selected by the tree were decision-making process factors. All variables in the regression tree were also in the linear regression model.

**Table 5.** Statistical results of the two types of regression models that explain the date of first irrigation. Only variables selected for at least one of the two methods are shown (see Table 3 for a description of the variables). Significance codes: 0 < *p* < 0.001: \*\*\*; 0.001 < *p* < 0.01: \*\*; 0.01 < *p* < 0.05: \*.


In the linear regression model, the date of first irrigation was significantly influenced by the number of years of experience with maize production, level of risk aversion, department, type of maize grown, irrigation at sowing, total volume of water used during irrigation, number of intention changes, number of other maize farmers with whom the farmer interacted, and number of weather stations (Table 3). In comparison, the variables in the regression tree, presented by decreasing influence, were the number of intention changes, tillage, maize type, number of other maize farmers with whom the farmer interacted, total volume of water used during irrigation, and number of technologies used to obtain weather information. The first branch of the regression tree, the mean date of first irrigation for farmers with high reactivity, is earlier than the sample mean (9 July vs. 24 June). On the other hand, among farmers with low reactivity, tillage application tends to advance the date of first irrigation. Moreover, farmers with big networks tend to irrigate earlier than others (Figure 6).

To offset the limits of each model (e.g., linearity and distribution hypotheses, multicollinearity, complex interactions, local effects), we compared the results of the models before determining how influential each variable was. The variables selected by both models were the type of maize grown, total volume of water used during irrigation, number of intention changes and number of other maize farmers with whom the farmer interacted. As expected, structural and material factors influenced the date of first irrigation, but decision-making process factors (levels of reactivity and assistance) also had an influence in both models. Notably, reactivity was the variable with the most significant influence in the regression tree and the linear regression model (*p* < 0.001).

The linear regression model explained 77% of the variance (adjusted R<sup>2</sup> = 0.77). The tests of residuals of independence (Durbin–Watson test), normality (Shapiro–Will test) and homogeneity (Breusch–Pagan test) were satisfactory, as was the reliability of the regression tree model, probably due to the choice of a conservative stopping rule (minsplit = 10) to minimize the error.

**Figure 6.** The regression tree model that explained the date of first irrigation. Information on lines includes thresholds or decision variables. Letters in parentheses are variable classes (C: agronomic practices; D: irrigation practices; F: reactivity; G: assistance).

3.3.2. Influence of Structural and Material Factors

All groups of structural and material factors influenced the date of first irrigation in at least one model. The farmer's experiences (Farmer's characteristics) influenced the date of first irrigation in the linear regression model. Experience increases the ability to observe changes in the environment and to rapidly and efficiently make decisions [26,38,39]. The more experienced the farmer was, the later the farmer started irrigating.

The department (Farm characteristics) also influenced the date of first irrigation in the linear regression model. Farmers in the Tarn and Haute-Garonne departments tended to start irrigating later than those in the Gers (mean of +6 and +10 days, respectively). Differences in soil and climate conditions, such as a drier spring season in the Gers (40 mm less rainfall on average), could explain the heterogeneity of the date of first irrigation.

Agronomic practices are of primary interest. In both models, the type of maize had a strong influence on the date of first irrigation. For example, popcorn maize, which has a less dense canopy [40], was associated with a later date of first irrigation in both models. According to the regression tree, seed maize was irrigated later than grain or fodder maize. Later sowing dates for seed maize can explain these later dates of first irrigation. Conversely, fodder maize was associated with an earlier start of irrigation since it is harvested immature and irrigated to optimize early vegetative growth. The influence of grain maize differed between the models due to differences in their mathematical functioning. Maize grain was significantly (*p* = 0.0493) associated with a later date of first irrigation in the linear regression model but with an earlier date of first irrigation in the regression tree. We considered the regression tree to be more relevant since the influence of maize type was based on interactions with previously chosen variables (e.g., tillage, psychological factors). Since the type of maize was significantly correlated with the department (*p* = 0.005), soil and climate conditions in the department could also explain indirect effects.

Tillage, another agronomic practice, was associated with a later date of first irrigation in the regression tree (a mean of +8 days) but not in the linear regression model. Direct effects of tillage on water availability for a crop are complex and depend on local conditions and practices since tillage can decrease water infiltration into the soil as well as increase evaporation [41,42]. Tillage can also have an indirect effect since it strongly influences other influential variables in the models, such as cover crop and irrigation at sowing. Tillage was negatively correlated with the variable cover crop (*p* = 0.010) since tillage is performed mainly in autumn in this area and, conversely, was positively correlated with irrigation at sowing (*p* = 0.007).

Irrigation at sowing (irrigation practices) was positively correlated with the date of first irrigation and was significant in the linear regression model. Farmers who irrigated at sowing started irrigation later. Irrigation at sowing provides additional water for the maize, which decreases the need for irrigation later.

In both models, the volume of water used for irrigation significantly influenced the date of first irrigation, but the direction of the effect differed. In the linear regression model, increasing volume was associated with an earlier date of first irrigation; the more water the farmer has, the earlier he will irrigate because he does not need to save water since there is no risk of being water-limited later. Conversely, in the regression tree, decreasing volume was associated with an earlier date of first irrigation. Since the volume variable appeared at the end of the tree, only a few of the farmers were concerned by this result, including those who grew fodder maize, who irrigate earlier.

We thus confirmed the influence of farmers' experience, farm location and agronomic practices. The influence of structural and material factors was consistent with the literature [10,11,13,18].

#### 3.3.3. Influence of Psychological Factors

As expected, farmers' risk aversion was negatively correlated with the date of first irrigation: a farmer with greater risk aversion tended to start irrigating earlier. A farmer who is risk-averse will deliberate over a decision as much as possible and will start irrigating earlier to avoid the risk of hydric stress on maize plants before it occurs. Several studies have demonstrated the influence of risk aversion on decision-making [19,26,43,44].

A major result for decision-making process factors was the key influence of the level of reactivity (i.e., number of intention changes). Thus, the more reactive the farmer was, the later the farmer started irrigating. In a previous study of factors that influence fungicide applications on soft wheat [25], a high level of reactivity was associated with adaptive behavior. Similarly, Rodriguez et al. (2011) showed that reactivity (or plasticity) provided greater resilience to change than anticipation (or rigidity) when facing uncertainty since it improved adaptive behaviors and strategies [45].

The level of assistance also had a significant influence. The number of other maize farmers with whom the farmer interacted was negatively correlated with the date of first irrigation in both models. This suggests a mimetic effect: interacting with a larger network of farmers increases the likelihood that one of the farmers in the network will have started irrigating. Several studies indicate that the size of the social network increases the adoption of adaptive behaviors [22,46].

Unlike human factors, technological assistance variables were positively correlated with the date of first irrigation. Farmers who had a weather station or used multiple information technologies were more likely to start irrigating later. The weather-station variable was also significantly correlated with the use of decision-making tools (*p* = 0.03) or weather sensors (*p* = 0.03). We concluded that all types of tools that provide accurate and specific information about the weather could postpone the date of first irrigation. In the same way, Berthold et al. [47] also showed that the use of irrigation tools make it possible to optimize water by making informed decisions. These opposite effects of different types of assistance variables are noteworthy; they suggest that human assistance advances the date of first irrigation, while technological assistance postpones it. In either case, assistance leads to adaptive behaviors.

No variable related to deliberation appeared in either model; thus, unlike reactivity and assistance, deliberation did not influence the date of first irrigation. This result differs from that of Daydé (2017) for whom deliberation increased the adoption of more sustainable practices.

#### 3.3.4. Synthesis of Results

Figure 7 summarizes results regarding factors that influence the decision of the date of first irrigation.

**Figure 7.** Variables identified as influential factors for the decision of the date of first irrigation.

#### *3.4. Advantages and Disadvantages of the Method*

The use of different inquiry methods allowed us to identify robust indicators to describe the decision-making process. We removed the subjectivity of personal statements by using methods such as role-playing and different scenarios with farmers.

Preselecting variables based on correlation and agronomic expertise was important to minimize the types of bias that collinear variables can create in linear regression models: high variance in predictors, large or unstable regression coefficients, and coefficient signs that run counter to intuition [48] Because predictors change when explanatory variables are strongly correlated, we preselected only independent variables. However, we could not eliminate all complex interactions and correlations that can disturb linear regression models. To obtain a relatively equal distribution of variables among the groups, variables were excluded only if they were simultaneously in the same group and had high correlations

between each other (i.e., *p* < 0.05 for qualitative variables, and Pearson correlation >0.4 for quantitative variables). We used a regression tree to offset these limits of the linear regression model, but it was subject to more local effects since it divided observations into groups and sub-groups until the stopping rule was activated. In particular, variables at the end of the tree must be carefully interpreted because, in this study, they were based on 3–4 individuals. Deep learning from our database was challenging due to its small sample size.

We obtained more robust results by using two types of regression models that have complementary advantages and disadvantages. Although linear regression and regression trees are based on different statistical approaches, each yielded similar results, particularly the strong influence of decision-making process factors (assistance and reactivity) on the date of first irrigation. However, the models sometimes yielded different results due to their functioning or initial descriptions of the data. For example, regression trees can highlight local effects of variables, such as the volume of irrigation water, which obscure the overall influence of these variables for the entire sample. The linear regression model always considered all observations of the sample. However, when two variables were strongly correlated, it selected only the one that best explained the date of first irrigation, and this approach can ignore the influence of the second variable.

The main disadvantage of this study is its relatively small sample size (34 farmers). Since the sample is not entirely representative of the region, the results cannot be considered generic. However, they provide knowledge about the adaptive capacity of large maize farming systems. Moreover, our goal was not to describe or predict behaviors of farmers in the region, but to test the hypothesis that decision-making process factors can influence irrigation practices. We met this goal since we revealed the strong influence of reactivity and assistance on the heterogeneity of the date of first irrigation. For example, the linear model selected the number of intention changes because it had the largest influence on the date of first irrigation, but it ignored the number of technologies because it was redundant.

We studied the influence of multiple factors on the date of first irrigation, which is only one aspect of farmers' irrigation practices. Thus, it could be interesting to study other aspects such as irrigation equipment or duration, which would make it possible to test the influence of decision-making process factors on the entire irrigation strategy. However, the current study did not include multiple factors due to time, means and budget limitations.

#### *3.5. Improving Adaptive Capacity*

Although adaptation strategies are studied in the agricultural extension literature, farmers do not always adopt them. According to Öhlmér et al. [49], adaptive capacity can explain the difficulty that farmers experience when implementing new practices recommended by experts. Adaptive capacity is defined as the capacity of actors to implement new adaptation strategies, which leads to resilience [50]. Farmers' behaviors can explain much about their adaptive capacity [51]. In particular, the decision-making process needs to be studied to improve adaptations [52]. Thus, a better understanding of the influence of farmers' decision-making mechanisms on the adoption of practices could improve their adaptive capacity through the design of specific supports and policies.

Understanding farmers' adaptive processes is crucial for improving adaptation strategies. Behavior models that model the decision process using decision-making process factors, such as that of Daydé [23], help explain the heterogeneity of practices and, thus, the reasons for adopting practices. Our study reveals that farmers adopt practices in part due to their decision-making process. For agricultural water management, levels of assistance and reactivity strongly influence the date of first irrigation.

Reactivity could improve the adaptive capacity of farmers since a reactive decisionmaking process is associated with changes in irrigation practices (i.e., later date of first irrigation). The more reactive farmers are, the more they are able to postpone the date of first irrigation if necessary. Therefore, if farmers are facing a heatwave forecast, they would

be able to change their date of first irrigation in order to find a balance between saving water and avoiding water stress.

Better support of maize farmers in southwestern France could encourage them to become more reactive. One way to increase reactive behavior is to encourage greater consideration of new information, increase the ability to observe changes in the environment and make better use of past experiences. One starting point is for farmers to share experiences and re-frame self-criticism of past decisions in discussion groups.

Encouraging access to specific information tools such as weather stations and new technologies is a way to obtain more adaptive behaviors, which may help to optimize water use for irrigation. Communicating with and educating farmers about the use of decision-support tools and technologies could increase adaptation practices. In addition, financial support from agricultural policies for farmers to invest in these tools would be relevant.

Future research should focus on a better understanding of decision-making strategies and the identification of relevant methods to measure them. Research should also focus on understanding how to improve adaptive behaviors. A key element of the decision-making strategy is the information received by the farmer and the farmer's ability to process information. Helping farmers find, access and understand information, compare sources, and rapidly select the relevant information according to the context are initial elements required to improve adaptive behaviors.

Our study contributes to research on adaptations by highlighting the important role of farmers' decision-making strategies. We revealed the need to improve reactive and assistance behaviors to increase adoption of adaptation practices, and to provide ways to improve these adaptive behaviors. These elements should be considered by advisors and included in public policies.

#### **4. Conclusions**

To explain the heterogeneity of the date of first irrigation among farmers, we surveyed 35 maize farmers. Our results confirm the role of structural, material and risk-aversion factors. They also highlight the strong influence of decision-making process factors on the date of first irrigation. Reactivity influenced the date of first irrigation more than any other variable. A high level of reactivity is associated with adaptive behaviors. Assistance from decision support tools and technologies also helps farmers adopt more adaptive behaviors. Conversely, other types of assistance such as social networks decrease adaptive capacity. However, assistance always influenced the date of first irrigation, whether it advanced it or postponed it. Advisors and public policies in the agriculture sector could consider these elements as ways to improve adaptation. In the context of water scarcity, our findings could help agricultural advisors to assist maize farmers with their water management practices. Future studies of farmers' irrigation practices could focus on exploring the influence of decision-making process factors on other key explanatory variables such as equipment, irrigation sources or water volumes. Their results would help us to understand the extent to which decision-making process factors influence the irrigation strategies of maize farmers.

**Author Contributions:** Conceptualization and methodology, M.A., S.C., J.-E.B., M.W.; Writing— Original draft preparation, M.A., S.C., J.-E.B., M.W.; Investigation, M.A.; Formal Analysis, M.A., R.F.; Writing—Review and editing, M.A., S.C., J.-E.B., M.W., R.F.; Funding Acquisition, M.W. All authors have read and agreed to the published version of the manuscript.

**Funding:** The research was supported by INRAE as part of the VACCARM project of the ACCAFMetaprogram.

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Informed consent was obtained from all subjects involved in the study.

**Data Availability Statement:** Survey instruments and code used in this study are available from the authors upon request.

**Acknowledgments:** This study was funded by INRAE as part of the VACCARM project of the ACCAF Metaprogram. The authors thank the trainees who helped collect the data and the English proofreaders.

**Conflicts of Interest:** The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analysis or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.

#### **Appendix A**

"In this part, we will assess your preferences toward risk using a lottery game. Nine lottery games are proposed. For each game, two profits are possible (a low one and a high one) with identical chances to occur. We will ask you to choose your favorite lottery game from among the nine proposed".
