Identifying Key Features for Establishing Sustainable Agro-Tourism Centre: A Data Driven Approach

Tourism is a vital economic sector that contributes significantly to the global economy and sustainable development. It acts as a growth engine by stimulating investment in infrastructure, preserving cultural heritage, and fostering environmental conservation [1]. Agro-tourism has emerged as a niche sector that combines agricultural activities with tourism experiences. Agro-tourism allows visitors to engage in farming practices, learn about rural lifestyles, and enjoy farm-to-table experiences [2].

Mainstream tourism practices can be effectively integrated with Agro-tourism by leveraging its unique characteristics and aligning them with broader tourism trends. Numerous researchers have explored the potential of Agro-tourism and its ability to diversify rural economies, promote sustainable agriculture, and attract eco-conscious travelers [3-6].

It supports local farmers by generating additional income while preserving traditional farming practices [1]. It helps for sustainable rural development and also highlights economic, environmental and social benefits to rural population, especially in the regions suffering from high unemployment, rural depopulation, and limited infrastructure [5].

The researchers have explored conceptually the indicators required for the development of Agro-tourism [7-10]. These studies have only focused on potential indicators suggested for Agro-tourism, however, the data driven approach for prioritization and identification of important indicators is lacking. This indicates a strong need for exploring the growth of Agro-tourism. For the sustainable growth of Agro-tourism, it is important to identify its key indicators.

The current study aims to identify indicators for establishing and enhancing Agro-tourism.

1.png

Figure 1. Methodology framework

3.1 Data collection

The area selected for the study was the Pune District of Maharashtra state in India. The study area recorded the highest concentration of Agro-tourism centers. Its proximity to urban populations, diversity in agricultural practices, and successful government and private initiatives make it an ideal region for studying the growth, challenges, and impact of Agro-tourism.

The data collection was done by conducting the field visits to 81 Agro-tourism centers out of 106 which is 76.4% of the study area through a structured questionnaire. The questionnaire mainly focused on collecting the data for the indicators identified in Section 1. Each question from the questionnaire was validated using Cronbach Alpha technique. The questions with Cronbach Alpha values greater than 0.77 were retained for further study.

3.2 Data preprocessing

The collected raw data underwent preprocessing to ensure it was in a usable format. Various techniques were applied, including handling missing values, label encoding, and feature scaling. Categorical data was converted into numerical format using appropriate encoding methods. Other preprocessing steps included missing value treatment, categorical data conversion, and feature scaling to enhance data quality and reliability for the study.

3.3 Feature selection

In this Section, the important indicators which are synonymously called as features in machine learning were identified. The ML methods such as filter approach, wrapper based, and embedded approach were applied to identify important features for growth of Agro-tourism. Feature importance is commonly evaluated using criteria such as information-theoretic measures, performance on a validation set, or accuracy on the training data, to determine a feature's effectiveness in distinguishing between classes [19]. Filter feature selection methods use statistical computation to assess the relevance of features to the target task. Only features that meet specific scoring criteria are retained. Wrapper methods naturally address multicollinearity by eliminating redundant features that fail to contribute unique information to the model. Embedded approaches include classifiers with built-in feature selection methods that perform variable selection during the training process. Unlike filter and wrapper methods, embedded approaches establish an interaction between feature selection and the learning process. Standard embedded methods for feature selection include decision tree-based algorithms such as Classification and Regression Trees (CART), C4.5, and Random Forest (RF) [20]. Although RF is inherently an embedded feature selection method, it is also frequently employed as a classification algorithm to evaluate feature subsets generated by filter and wrapper methods [19, 21, 22].

LASSO (Least Absolute Shrinkage and Selection Operator) is a regression technique that integrates variable selection with regularization, aiming to enhance model's predictive accuracy and interpretability [19]. LASSO applies a penalty to the modulus of the regression coefficients, effectively shrinking some of them to zero, which results in automatic feature selection [19]. In this study, to find the best features for growth of Agro-tourism embedded method LASSO is used.

3.4 Classifiers and predictors

Supervised machine learning plays an essential role in designing prediction models which could significantly enhance operational efficiency [21]. Generally, classification technique is used to predict the result in specific categories or classes which are generally predefined. These models can discover patterns and predict results by training algorithms on labelled datasets in which inputs connect with recognized outcomes [23]. This predictive potential has significance for early detection of the result which helps in taking different prevention measures before failures. In the current study, we used LASSO technique for the feature selection. To find the efficacy of feature selection techniques, machine learning classification models were applied. Considering the small nature of the dataset, simpler models like LR and DT were used to understand model behavior initially. Complex models like SVM and ANN were avoided since they can be unstable and overfit with limited data. The datasets were lacking the non-linearity, complexity in terms of high dimensionality which forced to select the ML models such as DT, LR, RF, and XGBoost over SVM and ANN to generate the result. The analysis was done by splitting the dataset into 70%-30% as well as 80%-20% training and testing dataset. The accuracy of these classifiers was tested using Accuracy, Precision, Recall, and F1-score [24].

Logistic Regression (LR) is one of the commonly adopted models in machine learning. It is typically employed to model binary outcomes [25]. LR predicts a categorical output based on input values, each assigned specific weights or coefficient values. The output variable, YY, follows a Bernoulli probability distribution, taking values between 0 and 1, ranging from 1−π to π. The threshold is typically set at the midpoint, with values exceeding this threshold classified as 1, and those below it classified as 0 [25].

The Decision Tree Classifier is one of the most powerful algorithms in machine learning. This method structures data in a tree-like format, where each branch represents a feature and its corresponding category, which is crucial for evaluating the outcomes [26]. Selecting the most relevant feature and placing it at the correct vertex is essential during tree splitting. This process leads to the generation of valuable features, resulting in highly effective outcomes [21].

The Random Forest algorithm is a robust machine learning technique based on decision trees. It creates multiple decision trees during the training process, each using a randomly selected subset of the dataset. Additionally, at each split within a tree, only a random subset of features is considered. This randomness in Random Forest helps to minimize overfitting and enhances the overall predictive performance of the model [21].

Ensemble learning integrates multiple models to achieve better predictive accuracy and XGBoost is a popular example of such an ensemble method [27, 28]. Ensemble algorithms are typically classified into three main types: bagging, boosting, and stacking [28]. XGBoost falls under the boosting category of ensemble methods [27, 28]. XGBoost is a next-generation boosting algorithm designed to build highly optimized models by identifying the best parameters through the minimization of an objective function.

3.5 Performance analysis methods

Performance evaluation metrics are essential for assessing model effectiveness. In this research, seven performance metrics are used: Accuracy (ACC), Sensitivity (SEN), Specificity (SPE), False Positive Rate (FPR), False Negative Rate (FNR), F1-Score, and Precision. These metrics are calculated using the formulas provided in Eqs. (1) to (7) [29].

$A C C=\frac{T P+T N}{T P+F P+F N+T N} \times 100 \%$                           (1)

$S E N=\frac{T P+T N}{T P+F N} \times 100 \%$                          (2)

$S P E=\frac{T N}{T P+F P} \times 100 \%$                        (3)

$F P R=\frac{F P}{F P+T N} \times 100 \%$                              (4)

$F N R=\frac{F N}{T P+F N} \times 100 \%$                        (5)

$Precision$ $=\frac{T P}{T P+F P} \times 100 \%$                         (6)

$F 1-$ Score $=2 \times \frac{{SEN} \times{Precision}}{{SEN}+{Precision}} \times 100 \%$                     (7)