A Machine Learning Approach for Investment Analysis in Renewable Energy Sources: A Case Study in Photovoltaic Farms

Abstract

Farmland offers excellent conditions for developing solar energy while farmers seem to appreciate its notable revenues. The increasing adoption of photovoltaics (PVs) on farmland raises various concerns with the most important being the loss of productive farmland and the increased farmland prices, which may prevent young farmers from entering the farming occupation. The latter can threaten the future of agriculture in countries that are already facing the problem of rural population ageing. The aim of this paper is to examine the effect of crop type on farmers’ willingness to install photovoltaics on their farmland. To that end, this study applies four machine learning (ML) algorithms (categorical regression, decision trees and random forests, support vector machines) on a dataset obtained from a questionnaire survey on farmers in a Greek agricultural area. The results from the application of the algorithms allowed us to quantify and relate farmers’ willingness to invest in PVs with three major crop types (cotton, wheat, sunflower) which play a very important role in food security. Results also provide support for making policy interventions by defining the rate of productive farmland for photovoltaics and also for designing policies to support farmers to start and maintain farming operations.

1. Introduction

In contrast to other energy production technologies that have fewer space demands, solar energy relies on the availability of large spaces for the installation of panels. Being impressively flat and exposed to uninhibited sunlight, farmland offers excellent conditions for developing solar energy and has become a primary target for solar developers [1,2]. Another advantage of farmland is that it is free of ‘nuisances’ such as infrastructures, trees, wetlands and rocks, which may be present in other types of land. From investors’ perspective, farmland is considered one of the most appropriate installation sites and has been transformed into a strong component in decision-making regarding the uses of agricultural land [2,3].

The installation of solar energy on farmland has been following an increasing trend to the degree that it has been characterized by some proponents as “just another crop” [2]. Farmers’ positive response to the concept of “farming the sun instead of the land” is largely attributed to the benefits flowing from investments in photovoltaics on farmland [2,3]. Perhaps the most notable benefit is that farmers can instantly gain an additional income either by installing solar panels on their land and selling the produced energy or by leasing their land to solar energy producers [4]. In a somewhat indirect manner, the wider community can also benefit from these revenues if they are used for improvements and investments in farming operations.

From a managerial perspective, solar energy on farmland has significant benefits. Compared to fossil fuel plants, whose deployment is particularly time consuming, solar energy entails significantly shorter building and installation lead times. At the same time, the installation of solar projects is much simpler as it does not involve high up-front investments in centralized energy units while stranded asset risks can be avoided [5]. In terms of electricity production, transmission and distribution losses are minimized significantly in relation to other energy production technologies [5].

The continuous development of solar energy on farmland, however, has attracted severe criticism mainly due to the loss of productive farmland that lies underneath solar panels for an extended period of time [6]. This change in land use has been shown to decrease crop yields and in some cases it has even reached up to 40% [7]. Advocates of farmland preservation warn that unless drastic measures are taken, solar energy may further displace farmland and even threaten food security [3]. This warning seems to be reasonable if one considers that most solar energy projects on farmland are only a few acres at first, but they quickly expand to thousands of acres. Phenomena of quick expansion occur because landowners, whose land neighbors solar projects, become quickly familiar with the benefits of solar electricity production and appreciate its benefits over crop production [8]. There is, therefore, a significant risk that solar development will accelerate rural industrial development at the expense of agriculture [9].

Another negative effect concerns the consequences of higher farmland prices for the wider agricultural community. That is, solar producers’ demand for farmland increases the price of the remaining farmland located in areas with significant solar energy deployment. This externality was observed by Lai et al. [4], whose study assessed the impacts of solar installations in monetary terms and indicated that solar energy on farmland drives notable increases in farmland prices ranging from 3.40% up to 37.02% in comparison to average farmland prices. The main conclusion from this study was that higher farmland prices do not benefit neighboring farmers nor advance the national agricultural development [4]. Although increased farmland prices can mean instant revenues for farmers interested in leasing their land to energy developers, expensive farmland can be particularly discouraging for young farmers who seek farmland to rent or purchase in order to enter the faming occupation. In the longer term, the deployment of solar energy on farmland may become not only a permanent land use, but may also alter the identity of traditional rural areas [10]. In this sense, the rural land may cease to be a tranquil green landscape and become industrialized and, at the same time, local residents may abandon traditional rural activities and engage in entrepreneurial activities related to energy production. This kind of disruption to rural landscapes could also have a negative effect on local tourism and decrease tourism value, as the impaired visual landscape quality can be discouraging for rural tourists [10].

The financial benefits of solar energy are somehow reduced due to health concerns which are typically expected from the deployment of other energy sources, such as fossil fuel plants and wind turbines [11]. It seems that rural communities have also been expressing fears about the health effects of solar panels. Given that farmers may be torn between the immediate financial benefits of solar energy, the loss of agricultural land and other environmental and socio-economic factors, it is only reasonable to expect that farmers’ decision to install solar panels on their farmland is particularly complex and prone to the influence of various factors.

Policy wise, the understanding of farmers’ decision-making is important because it can point to specific areas that require policy intervention. So far, the existing studies show that farmer decisions are affected both by financial and non-financial variables. Specifically, farmers’ decision to install panels is correlated with financial aspects such as capital availability, increased energy cost, increased total value of production, increased farmland value, increased operation size, farm size, the existence of debts, the level of agricultural revenues, as well as the availability of outside funding [12,13,14,15,16]. Moreover, farmers are more likely to adopt solar energy if they exhibit interest in environmentally friendly practices such as organic farming [13]. In addition to these, positive environmental attitudes, information access and high educational level have emerged as variables that are positively associated with a farmer’s decision to install solar panels on their farmland [12]. Farmers also seem to be affected by their community’s attitude towards solar development; specifically, they are more willing to adopt solar panels if their community is also engaged [17]. It should be noted that the type of cultivation may be an influential factor but there are very few studies on the subject. A study by Beckman and Xiarchos [13] conducted on Californian farmers and ranchers indicated that those engaged in livestock farming are more likely to adopt solar energy compared to farmers engaged in fruit operations.

The European Union (EU) has been described as an energy leader due to the establishment of strong supporting policies. In response to concerns about the use of farmland for energy generation purposes, the EU has already paid considerable policy attention to the risks posed by the rapid expansion of photovoltaics on farmland [18]. Meanwhile, the public in many EU countries still reacts negatively and expresses concerns about the sustainability of this kind of installation. In most EU countries, the criteria for installing panels on farmland have become stricter and it typically takes three to five years for solar projects to issue license and begin operation [6,19].

Greek agriculture is described as having a high level of diversity, due to the fragmentation of agricultural lands and the existence of small farms which spread over diverse regions. This diversity is also due to mountainous and island areas, various microclimate zones, as well as long-standing historical and cultural path dependencies [20]. Because of these characteristics, Greek agriculture favors wide differentiation and variety rather than large-scale industrial production. At the same time, it depends on subsidies provided through EU’s Common Agricultural Policy, which defines the type of crop production to a significant degree [20]. Regardless of modern agriculture requirements, the size of Greek agricultural holdings is relatively small (on average, 6.6 hectares) and most holdings (about 90%) are less than twenty hectares, which is considered inadequate in terms of modern production prerequisites [21]. Another issue concerns the falling contribution of agriculture to the national gross domestic product (GDP). In 1981, when Greece joined the European Union, 30% of the national workforce was engaged in agriculture with agricultural production accounting for 25% of GDP. Today, only 11% of the workforce is engaged in agriculture with the latter contributing only 3.3% to the national added value [22]. In comparison to other EU countries, Greek agriculture seems to perform rather poorly as its contribution to EU-wide agriculture decreased from 3.3% in 1993 to 2.6% in 2015. Based on EU indicators, a similar poor performance is observed for income per worker and entrepreneurial income [20]. Poor performance can be attributed to specific constraints which inhibit the development of the agricultural sector in Greece. Most importantly, the rural population is aging and only 5.2% of crop producers are below the age of 35 years and, at the same time, most farmers lack formal education and training on farming including training on competencies related to running successful farming businesses [23]. Furthermore, the private sector and state institutions related to agriculture are not organized and are not coordinating with each other. Most agricultural cooperatives cannot address existing deficiencies as they are also not well organized [20]. It can be stated that Greek agriculture is subject to multifaceted influences which make it very difficult for policymakers to reconcile land abandonment, biodiversity restoration and sustainable rural development [24].

In view of the challenges involved in the Greek agricultural sector, it is hardly surprising that photovoltaic installations on farmland have drawn significant interest from farmers. This interest is mainly ascribed to the notable reduction in the cost of photovoltaics (purchase, installation, usage and maintenance) in comparison to previous decades, the considerable revenues producers receive for the produced energy, as well as the institutional framework, which is constantly developing in order to correspond to the new reality [22]. As in other EU countries, the legal framework in place in Greece poses certain limitations to the installation of solar panels on productive farmland. To be more precise, Article 24 of Law 4643/2019 foresees that the installation of solar projects with installed capacity equal to or less than 1 MW is allowed on “agricultural land with high productivity” under the condition that solar installations in each regional unit do not exceed 1% of the total area of cultivated lands. However, solar projects which have a capacity over 1 MW on farmland that has been characterized as “agricultural land with high productivity” are not allowed to be installed. The levels of permitted installation capacities for solar projects are fixed for each region and must not exceed 1.0% of the total area of cultivated land in the continental part of the country (the corresponding rate for islands is 0.5%). In the regional unit of Evros, where the study area is located, cultivated land amounts to 1,532,985 decares and thus the largest installation capacity of solar projects on “agricultural land with high productivity” is 15,330 decares. Regarding payments, farmers receive 65 Euros per MWh on the conditions that farming is their main occupation and that the electricity is produced by panels whose installed capacity does not exceed 500 kW. It is also interesting to note that in many regions of Greece, the prices for renting agricultural land have increased significantly. In many cases, landowners can receive a high passive income for an extended period of at least 25 years just by leasing their land.

Against this background, the decision-making of farmers about solar energy on farmland acquires particular interest as it can reveal the areas that require policy intervention. From a wider perspective, there can be no doubt that farming is a highly demanding activity that is prone to many externalities, making the option of ‘farming the sun’ instead of the land highly tempting for farmers [4]. With the increasing deployment of solar projects on farmland, however, farmland prices increase and consequently the cost for renting or purchasing land is very high for young farmers [10]. In other words, the continuous expansion of solar projects may not only change land use and reduce crop yields, but it may also prevent younger individuals from engaging in agriculture. Combined with the fact that the agricultural labor force in Greece is aging, solar energy expansion can be detrimental to the future of the national agricultural development [4,10]. Although the country has set specific limits to the size of farmland that can be used as installation sites, farmers’ decision to replace crops with solar panels can pose severe threats. For this reason, it is important to examine the factors that affect farmers’ decision to install solar panels on their farmland. Most studies on the subject have focused on financial, sociodemographic and attitudinal variables, but, as of today, there has been little focus on the effect of crop type on farmers’ willingness to invest. In the literature on environmental problems, the application of machine learning algorithms is promising and could be combined with datasets collected from social surveys. For instance, Mishra et al. [25] built a comprehensive arsenic awareness index and indicated awareness drivers by using a questionnaire survey in combination with machine learning models. In addition, Ding et al. [26] combined XGboost algorithm and energy consumption data with data from a questionnaire study in order to predict the energy consumption of public buildings. Moreover, in order to assess the energy consumption in an Iranian city and to formulate strategies for promoting citizens’ willingness for renewable energy generation, Ghadami et al. [27] employed a five-step procedure which included the development of a motivation algorithm based on interviews with experts and the application of artificial neural network.

Based on the theoretical and empirical backgrounds described above, the aim of this study is to investigate the underlying relationship between farmers’ willingness to install solar photovoltaics on farmland and the type of crop they cultivate. The main difference from previous relevant works on RES investment decision-making is the usage of ML algorithms (categorical regression, decision trees and random forests, support vector machines), which are increasingly used for solving, predicting and analyzing various problems; however, from our literature review, it is evident that they have not been used for prediction of farmer decision-making yet. These algorithms are vastly different from classical statistical approaches which have been used throughout the years. Their main differences (and strengths) lie in the way the algorithms deal with uncertainty, their capability to process huge amounts of data and finally (and more importantly), their ability to learn through their training process. As the findings of this study provide a new angle to interpret farmers’ decision-making, they can be particularly useful for environmental managers and policymakers who are engaged in the development of the agricultural and energy sectors.

2. Materials and Methods

2.1. Data Description

Results reported in this paper are based on a dataset that was collected using a questionnaire survey specifically designed to examine farmer decision-making about solar energy investments. In order to ensure that the results would be representative of the population under study, the sampling method that was followed was simple random sampling. This method was preferred due to its simplicity and because it requires very little knowledge about the studied population compared to other sampling methods [28]. The “population” under study was farmers in the Municipal Unit of Didymoteicho, a typical agricultural area in the Prefecture of Evros in Northern Greece (Figure 1).

In the prefecture of Evros and, according to 2019 data, the total arable land corresponded to 1,228,330 hectares. Of these, irrigated cotton cultivations were 223,611 hectares, non-irrigated cotton cultivations were 203,936 hectares, sunflower cultivations were 342,546 hectares, wheat (common) cultivations were 63,213 hectares, durum wheat cultivations were 224,468 hectares, barley cultivations were 26,115 hectares, oat cultivations were 484 hectares and rye cultivations were 2759 hectares [29].

Turning back to the estimation of the sample size, since simple random sampling without replacement was used, the correction of finite population can be ignored because the size of the sample n is small in relation to the size of the population N [30].

$$n=\frac{t^2\cdot \overline{p}\cdot \left(1-\overline{p}\right)}{e^2}=\frac{1.96^2\cdot 0.6\cdot \left(1-0.6\right)}{0.0{63}^2}=232.59\cong 233$$

(1)

In the above equation, t stands for the value of the Student’s t-distribution for probability (1 − α) = 95%, and n − 1 stands for the degrees of freedom. As the size of pre-sampling is relatively big (bigger than 50), the value is taken from the probability tables of the normal distribution for the desired probability. In practice, for probability 95% the value corresponds to 1.96 [28]. Moreover, p represents the proportion estimate and e is the maximum acceptable difference that can occur between the sampling mean and the unknown population mean. We can thus accept that it is 0.063, i.e., 6.3%.

In order to estimate the sample size, pre-sampling on a size of 50 individuals was performed. More specifically, it became possible to estimate the real analogy of the population for every variable. The questionnaire should not be able to estimate only one variable but many, as it is necessary to calculate the required sample size for each variable. Variables “Gender” and “Are you willing to invest in photovoltaics on your farmland?” gave the highest sample sizes. These are also the most significant variables in our research. According to research rules, if sample sizes of these variables are similar and fall within the economic ability of the research, then researchers should select the highest sample size. By doing so, the most ‘varying’ variable is considered with high accuracy, while the remaining variables are estimated with higher precision than was initially planned [28].

Following the principles of simple random sampling, respondents were administered questionnaires. In total, 233 landowners participated in the study, which was conducted from June 2020 to September 2020. To ensure the validity of responses, all questionnaires were completed through personal interviews with the respondents. The questionnaire consisted of 22 closed-ended questions and most items employed five-point Likert scales to enable respondents to express their views with great precision. More analytically, the questionnaire was structured around six thematic sections. The first section involved somewhat introductory questions that examined respondents’ views on renewable energy in general. The second section collected information on respondents’ farming operations such as type of crop cultivation and size of owned farmland. The third section examined respondents’ willingness to invest in photovoltaics on their farmland while the fourth section examined respondents’ views on various reasons for investing, including economic, environmental and social drivers. Then, the fifth section investigated farmers’ preference and use of information sources. Finally, the sixth section collected information on their sociodemographic characteristics. The analysis reported in this paper utilizes the items examining the crop type that farmers cultivate (wheat, cotton, sunflower) and the size of the cultivated area (quantitative variable). Specifically, respondents were asked to state “Which of the following crop types do you cultivate and which is the cultivated area for each crop type?”. Then respondents were asked “Are you willing to invest in solar panels on your farmland?”(qualitative variable); for this question, there were two possible answers: “Yes, I am interested in investing in the future” and “No, I do not want to invest in the future”.

2.2. Machine Learning

Initially, the data were imported to the Statistical Package for the Social Sciences (SPSS) and descriptive statistics was applied on all variables. However, machine learning (ML) was used to discover the underlying connections between crop type, cultivation area and farmers’ willingness to invest in solar energy as a means to obtain alternative income. The algorithms used for ML build models based on sample data, which are known as training data. This process allows the algorithm to understand the inherited data connections and after the end of the training period to make predictions, decisions or classifications to newly imported data (data which were not used during training and thus they are unknown to the algorithm). ML is the answer to computers solving problems without being explicitly programmed for doing so [31].

2.2.1. Linear Regression

Linear regression (LR) is a methodology suitable for data analysis, capable of predicting the value of unknown data. The prediction is based on the usage of known data values which were previously collected [32]. LR is considered a type of supervised machine learning algorithm capable of computing the linear relationship between dependent variable and one or more independent variables. In our case, we used a modified version of LR called categorical regression (CR) which is capable of using data which contain qualitative variables.

2.2.2. Decision Trees

Decision trees (DTs) are a supervised ML methodology capable of performing both classification and regression [33]. Decision trees are considered excellent algorithms in cases of pairwise dissimilarities such as categorical sequences [34].

2.2.3. Random forests

Random forest (Rf) is a ML methodology applied used for classification, regression, prediction and other tasks which used decision trees during training [35]. Rf is an ensemble method and as such is made up of a set of classifiers (decision trees). Their predictions are aggregated to identify the most popular result [36].

2.2.4. Support Vector Machines

Support vector machines (SVMs) are supervised learning algorithms capable of analyzing data for classification and regression analysis. By using a set of training examples, an SVM algorithm builds a model which assigns new data to one category or the other, making it essentially a non-probabilistic binary linear classifier [37].

2.3. Data Preparation

Building ML models is the initial and most important task. During the data preparation, we must perform data cleaning, data curation and removal of redundant features from the initial dataset, define the input and output variables, perform data splitting (if necessary—depends on the algorithm), train and cross validate the model (if necessary—depends on the algorithm) and finally, evaluate model performance by comparing the predicted values with the actual values [38].

2.4. Model Building

All ML learning algorithms can be categorized in three general types:

• Supervised. In this type, both X and Y values are known; therefore, the model is trained to predict the output from the input.
• Unsupervised. In this case, only X values are known; therefore, the model is trained to understand to model these values.
• Reinforced Training. In this case, the model decides on the next course of action, and it does this by learning through trial and error in an effort to maximize the reward.

Evidently, as both the dependent (Y) variables (crop type and cultivation area) and the independent (X) variable (farmers willingness to invest on RES) are known, our models will use the supervised learning methodology.

Different approaches for model building and sample usage have been followed depending on the type of ML algorithm we used. In the case of CR, we used the entire collected sample of 233 answers. In the case of DT and Rf, the sample was divided into two sets (training and test set) In more detail, of the total of 233 answers, we used 186 for the training dataset and 47 for the test dataset (Figure 2).

2.5. Hyperparameter Optimization

It is possible to improve the results produced from a ML algorithm by using a technique called hyperparameter optimization (HO). HO usage on a ML learning algorithm has a direct impact on the learning process and the prediction performance. There is not one hyperparameter setting that can be used on all ML algorithms nor is HO applicable to all algorithms. Therefore, we must perform hyperparameter optimization based on the algorithm we are using. In our case, we used the RandomizedSearchCV optimization method for the Rf algorithm.

2.6. Coding

For developing purposes, we used Google Colab (GC). GC is a product from Google Research which allows users to write and execute PYTHON code from their browser. GC is especially well suited for ML data analysis, by incorporating a plethora of required algorithms. Additionally, it provides access to advanced cloud resources including the ability for the user to use graphics processor units (GPUs) and tensor processing units (TPUs) [39]. Additionally, GC supports the capability of installing other ML libraries and allowing users to develop code using a variety of capabilities. For our project, we used the scikit-learn library which is a free software ML library for Python v3.11.4. It includes support for a variety of tasks (classification, regression and clustering) using support-vector machines, random forests, gradient boosting, k-means and DBSCAN [40].

3. Results

3.1. Classified Regression

The results from the application of CR are presented on

Table 1. CR results.

Independent Variables	Beta	Std Error	Df	Importance	F	p
Wheat	0.013	0.082	1	0.028	0.024	0.878
Cotton	0.203	0.069	1	1.007	8.562	0.004
Sunflower	−0.023	0.089	1	−0.035	0.07	0.791

. The analysis gave co-efficient value of multiple determination R2 = 0.04 and F = 3.210 and it is statistically important (<0.05).

The values of standardized coefficients indicate that the dependent variable “Farmers’ willingness to invest in photovoltaics on their farmland” is mostly affected by the type of crop cultivation and specifically, cotton cultivation. The measures of relevant importance of the independent variables suggest that cultivating ‘Cotton’ made the highest contribution to the dependent variable. Taking into account the above and in combination with the level of significance of the independent variables, we come to the conclusion that categorical regression is unable to predict “Farmers’ willingness to invest in photovoltaics on their farmland”.

3.2. Decision Trees

The DT algorithm was used with a variety of configurations (i.e., number of leaves, the maximum depth of the tree and the sampling of the features to consider when looking for the best split at each node). Setting up the algorithm is in many cases defined by using trial and error for the definition of its parameters. In our case, after several training sessions, optimal results were produced when we set the parameter random_state equal to 0.

DT after training and testing produced the following results:

Table 2. Decision tree results.

	Precision	Recall	f1-Score	Support
No	0.39	0.61	0.47	23
Yes	0.18	0.08	0.11	24
Accuracy			0.34	47
Macro avg	0.29	0.35	0.29	47
Weighted avg	0.56	0.53	0.29	47

presents the classification accuracy using the accuracy score function. In a multilabel classification, this function returns the subset of accuracy. If the entire dataset of predicted labels (in our case Yes and No) for a sample strictly match with the set true set of labels, then the subset accuracy is 1, otherwise it is 0.

In general, if ${\widehat{y}}_i$ is the predicted value of the i-th sample and $y_i$ is the corresponding correct value, then the fraction of correct predictions over $n_{samples}$ is defined using the following Equation (2):

$$accuracy\left(y,\widehat{y}\right)=\frac{1}{n_{samples}}\sum _{i=0}^{n_{samples}-1}1\left({\widehat{y}}_i=y_i\right)$$

(2)

In Equation (2), 1(x) is the characteristic function (the function that maps elements of the subset to Yes and all other elements to No).

The F1 is a harmonic mean of the precision and recall. An F1 score reaches its best value at 1 and worst at 0. The relative contribution of precision and recall to the F1 score are equal. The precision is the ratio tp/(tp + fp), where tp is the number of true positives and fp the number of false positives. The precision is intuitively the ability of the classifier not to label as positive a sample that is negative. The recall is the ratio tp/(tp + fn) where tp is the number of true positives and fn the number of false negatives. The recall is intuitively the ability of the classifier to find all the positive samples. Finally, support is the number of occurrences in each label.

The accuracy predictions include the following measures:

Macro calculates metrics for each label and finds their unweighted mean. This does not take label imbalance into account.

Weighted calculates metrics for each label and finds their average weighted by support (the number of true instances for each label). This alters ‘macro’ to account for label imbalance; it can result in an F-score that is not.

When the DT is applied on the test set, the following confusion matrix is created.

In Figure 3, the correct predictions are displayed on the main diagonal, 14 correct predictions for No and 2 correct predictions for yes.

Cultivation type plays a very important role in the farmers final decision. The feature importance property of the DT classifier calculates the importance of each feature. The importance is computed as the (normalized) total reduction of the criterion brought by that feature. In our case, cotton is the most important feature (cultivation) with a weight coefficient (WC) of 0.5839, followed by sunflower (WC = 0.3185). Wheat is the least important cultivation with a WC = 0.097 (Figure 4).

3.3. Random Forest

The RF algorithm was used with a variety of configurations (i.e., number of trees, the maximum depth of the tree and the sampling of the features to consider when looking for the best split at each node). Setting up the algorithm is in many cases defined by using trial and error for the definition of its parameters. In our case, after several training sessions, optimal results were produced when we set the following numbers to the parameters: n_estimators to 10, max_depth to 7 and random_state equal to 42.

On

Table 3. RF results.

	Precision	Recall	f1-Score	Support
No	0.51	0.83	0.63	23
Yes	0.60	0.25	0.35	24
Accuracy			0.53	47
Macro avg	0.56	0.54	0.49	47
Weighted avg	0.56	0.53	0.49	47

, we present the RF results after training and testing.

When the RF is applied on the test set, the following confusion matrix is created.

In Figure 5, the correct predictions are displayed on the main diagonal, 19 correct predictions for No and 6 correct predictions for Yes.

Cultivation type plays a very important role in the farmers’ final decision. The feature importance property of the RF classifier calculates the importance of each feature. The importance is computed as the (normalized) total reduction of the criterion brought by that feature. In our case, cotton is the most important feature (cultivation) with a weight coefficient (WC) of 0.4476, followed by wheat (WC = 0.3066). Sunflower is the least important cultivation with a WC = 0.2456 (Figure 6).

The application of hyperparameter optimization (HO) on the trained dataset defined the best hyperparameters having a max depth of 2 and the number of estimators at 151. The application of these parameters to the RF model produced the following confusion matrix (Figure 7).

Evidently, the results were improved, although not significantly. On the produced results, we applied the accuracy measure (

Table 4. Hyper optimized RF results.

	Precision	Recall	f1-Score	Support
No	0.49	0.91	0.64	23
Yes	0.50	0.08	0.14	24
Accuracy			0.49	47
Macro avg	0.49	0.50	0.39	47
Weighted avg	0.49	0.49	0.38	47

).

Finally, feature importance’s after the application of hyper optimizations are presented in Figure 8.

3.4. Support Vector Machine

The SVM algorithm was used by applying the radial basis function (rbf) kernel type. This particular kernel was selected because it is capable of approximating given functions. After the training of the SVM, the algorithm was tested using the testing data (

Table 5. Support vector machine results.

	Precision	Recall	f1-Score	Support
No	0.51	0.91	0.66	23
Yes	0.67	0.17	0.27	24
Accuracy			0.53	47
Macro avg	0.59	0.54	0.46	47
Weighted avg	0.59	0.53	0.46	47

).

When SVM algorithm is applied on the test set, the following confusion matrix is created.

In Figure 9, the correct predictions are displayed on the main diagonal, 21 correct predictions for No and 4 correct predictions for yes.

Cultivation type plays a very important role in the farmers’ final decision. The feature importance property of the SVM classifier calculates the importance of each feature. The importance is computed as the (normalized) total reduction of the criterion brought by that feature. In our case, cotton is the most important feature (cultivation) with a weight coefficient (WC) of 0.5839, followed by sunflower (WC = 0.3185). Wheat is the least important cultivation with a WC = 0.097 (Figure 10).

4. Discussion

The results from the application of the four ML algorithms indicate that support vector machines are more capable on providing the insights behind farmers’ decisions, followed by random forests and decision trees. The more classical methodology (categorical regression) failed to identify a statistical significance on two out of the three cultivation types, indicating that it is less capable and thus less likely to be used for solving this type of problems. Additionally, the results showed that even with a relatively small dataset, we could provide results with acceptable accuracy.

So far, most research on the drivers of investments in photovoltaics on farmland have employed questionnaire and interview surveys, with some of them providing conflicting results [41]. The latter brings forward the need to develop new methodologies in order to predict investments with higher accuracy. Being able to extract knowledge and patterns embedded in data, machine learning, a branch of artificial intelligence is suitable for identifying the underlying connections between dependent and independent variables, managing categorical and numerical values and calculating their weight coefficients in the final decision. Moreover, since all machine learning algorithms are optimized for large datasets, they can be combined with social science projects [42]. From this perspective, it is surprising that very few studies have so far employed machine learning to predict investments in renewable energy [43,44].

This study, however, is different from others in the field not only in terms of methodology, but also in terms of the examined drivers. With the exception of Beckman and Xiarchos [13], who examined the effect of type of agricultural operation on the decision of ranchers to adopt solar energy in the US, most previous studies have focused on drivers that concern financial aspects of the investment [12,14,15,16,17] and overlooked the relationship between crop type and farmers’ willingness to invest. This relationship includes, apart from the financial aspect of income produced by crops, other parameters like, the labor intensity of the cultivation, environmental and socioeconomic factors. The influence of some of the examined financial aspects on farmers’ willingness-to-invest, however, is somewhat expected and intuitive; for example, it only makes sense that farmers would be more inclined to invest if they anticipated high returns and if there was outside funding [15,16]. A new insight from this study is that the type of crop plays a major role in farmers’ decision-making. In the study area, farmers are more willing to invest in photovoltaics on their farmland when they are engaged in the cultivation of sunflower, followed by cotton and wheat. The increased willingness of sunflower growers to engage in solar electricity could perhaps be ascribed to the low prices in the sales contracts for sunflower and sunflower cultivations’ high need for irrigation water. The latter suggests that low prices for crops render farmers more willing to engage in electricity production and move away from agricultural activity. Regarding cotton growers’ increased willingness to invest in photovoltaics, a possible explanation is that cotton growers may have become somewhat fatigued by the production of cotton, which not only has high demands in water, energy and pesticides, but is also labor intensive. It is thus possible that the combination of high production costs and excessive labor renders cotton farmers more likely to seek alternative ways to create income perhaps in the form of solar electricity. This raises serious concerns because in the prefecture where the study area is located, cotton cultivations account for a high share of the total cultivated land and, over time, the area has developed long-term partnerships with cotton traders. If a notable share of producers abandons cotton cultivations for the sake of solar electricity, however, this would possibly have a negative impact on the sales contracts in the wider area. Reductions in production would also be detrimental to the agricultural character of the study area which has long been linked to cotton. Policy wise, this finding brings forward the need to design measures to support growers so that they do not abandon cotton production. For instance, it would be meaningful to provide subsidies in order to ensure that cotton growers can deal with the rising energy costs and the lack of laborers in the area.

Apart from cotton growers, farmers engaged in wheat production also present pronounced willingness to invest in solar energy on their farmland. If this willingness-to-invest is transformed into actual investments, a severe implication would be reduced wheat yields in one of the largest plains in the country. Unless measures are taken, the replacement of crops with solar panels could reach levels that are observed in other EU countries such as Germany, where the decrease in crop yields reached 40% [7]. Also, given that experts warn about food shortages due to the conflict in Ukraine, the replacement of wheat cultivation can further threaten food security in the longer term [3].

Taken as a whole, the findings here provide support for proposing changes in the existing legislation. According to Law 4643/2019, in the continental part of the country, the total productive area that solar installations occupy must not exceed 0.8% of the total cultivated land in the continental part of the country. Given that growers exhibited a keen investment interest in solar electricity production, there is a live risk that solar installations will expand at the expense of sunflower, wheat and cotton production. It is thus recommended to exclude from the allowed 0.8% those cultivations that are presenting shortages in the national and European markets or cultivations in which the country does not have autarky. Another recommendation is to establish measures to ensure that in agricultural areas with significant solar energy deployment, land prices remain stable so that young farmers are able to rent or purchase land. Along with these efforts, one more measure could be to utilize agricultural and photovoltaic complementary power stations. That is, the combination of crop production and solar applications does not alter the nature of land use thereby maintaining crucial land resources but also enabling farmers to engage in energy production. Newer studies have shown that complementary power stations can achieve efficient crop cultivations and may thus provide a new path towards agricultural development [44,45].

Additionally, policymakers should also consider the communal reactions to the installation of PVs. Residents, particularly those who are not engaged in solar energy, have expressed acute concerns about the safety of solar projects. Using qualitative semi-structured interviews with residents, Crawford et al. [8] observed that members of rural communities were particularly concerned about the toxicity of panels and were afraid that chemical compounds could leach into and contaminate groundwater. The glass of photovoltaics generates a reflection effect with light reflectivity of about 4%, which causes light pollution affecting ecological balance [45]. Light pollution can mainly affect human vision while the effect on mental health is negligible [46]. If photovoltaic modules are placed on the sides of the road at a certain height and angle, the reflected light will directly meet drivers’ eyes and affect their vision while driving. Another area of concern is that the expansion of solar installations may increase local temperatures due to changes caused in the way that incoming energy is reflected back to the atmosphere or absorbed. Concerns about temperature increases were corroborated by the findings of Barron-Gafford et al. [47], which indicated that temperatures on photovoltaic plants were 3 to 4 °C warmer than wildlands during the evening. It is thus recommended to additionally explore the opinions of the entire community apart from the farmers tendency to invest in PVs.

The findings reported in this paper should be considered within certain study limitations. Most importantly, the conflict in Ukraine and the subsequent increases in the prices of fuels may have affected the responses of farmers and perhaps have rendered more willing to invest. In addition, our analysis has not included other factors, such as sociodemographic and attitudinal variables, and, therefore, it is not possible to compare the influence of these factors. Our results could also provide some directions for future studies. Since results reported in this paper point to a pronounced willingness to replace crop production with solar energy production, it would be highly meaningful to examine the effect of more variables as well as the reasons for which farmers are willing to invest. The former could be examined by the same algorithms applied in this study, while the possible reasons, the problems encountered by farmers and the ways to address them may be examined through in-depth interviews because they can reach a deep level of analysis and thus may provide answers to more complex matters. More specifically, such insights may help policymakers design policies to encourage growers to maintain farming operations. In addition, it would be useful to examine farmers’ attitudes towards the existing agricultural subsidy schemes and measure their satisfaction with them. Moreover, research on the relationship between farmers’ willingness-to-invest and their values regarding the preservation of the agricultural character of rural areas can reveal the extent to which such values can act as drivers for maintaining agricultural operations and resist the wider tendency to replace crops with solar panels. If these values act indeed in the favor of farmland preservation, then it is possible to recommend strategies aiming at raising the awareness of farmers about land preservation ethics. Also, farmers’ attitudes towards the combination of agriculture and photovoltaic applications could be investigated as such techniques do not cause changes in the nature of land use and can thus can preserve land resources [48,49].

5. Conclusions

The purpose of this paper has been to examine the relationship of crop type and farmers’ willingness to invest in photovoltaics on their farmland. In other words, this work has focused on farmer decision-making in order to inform policy decisions seeking to ensure that food production is not compromised by a disproportionate level of photovoltaic installations on farmland. While most studies on the subject have used statistical analyses typically used in social research, such as regression analyses, this study combined the application of machine learning algorithms and compared their efficacy. As categorical regression here failed to identify the statistical significance of two out of the three cultivation types, our results provide support that machine learning algorithms are more capable of making predictions. There is thus adequate evidence showing that this methodology can be highly advantageous when seeking to establish high involvement decisions such as investments in renewables. As scholars in the field stress the importance of maintaining agricultural activities, which promote local rural economies and sustainable land use practices, this study added a new parameter that should be considered by policymakers seeking to preserve agricultural development. That is, this study has indicated that, beside other influential factors, crop type plays a major role in farmers’ investment decision. Being involved in the cultivation of specific crops increases farmers’ willingness to install solar energy on farmland which, in turn, suggests that there is a live risk for the yields of specific crops as well as national agricultural development per se. In particular, farmers engaged in cotton, wheat and sunflower cultivation are more likely to invest in photovoltaics on their farmland compared to farmers involved in other types of crop cultivation. Our study indicates that overall, farmers cultivating cotton are more willing to invest in RES. This willingness may be attributed to the current low crop prices, the difficulties that farmers encounter (such as energy costs and lack of laborers) with these types of crop cultivations and other environmental and socio-economic factors. If farmers’ willingness to invest in photovoltaics is transformed into actual investments, there can be severe consequences. Most importantly, the ongoing adoption of solar energy on farmland may possibly drive reductions in the yields of crops, which are important for food security, but also have negative effects on sales contracts and the agricultural character of the study area. At the same time, the demand for farmland for solar development could drive increases in farmland prices and prevent young farmers from entering farming. Hence, it is recommended to make changes in the legislation that defines the rate of productive farmland that can be used for the installation of solar energy. Moreover, policies aiming at encouraging farmers to maintain farming operations should involve financial support for farmers engaged in demanding crop cultivations or for those farmers who receive low prices for their yields. Finally, there should be policies to ensure that farmland prices do not rise due to land demands for solar installations so that young farmers are able to rent or purchase farmland.