Implementation of K-Nearest Neighbors Algorithm for Red Onion Crop Suitability Land Assessment in Selaawi, Indonesia

Shallots are one of the most important vegetables in Indonesian society. In health, shallots have been known since ancient times as a traditional medicine to reduce fever, treat diabetes and ulcers, and reduce sugar and cholesterol levels [1]. Economically, shallots are a strategic horticultural commodity with high economic value that contributes significantly to the economic development of a region [2]. In Selaawi District, Garut Regency, the productivity of shallot farming still needs to improve. According to BPS 2023 data, shallot production in Selaawi in 2020 was only 103 quintals and decreased to 0 quintals in 2021 [3]. This decline is caused by the condition of the rainfed land, which relies only on rainfall as a water source without an irrigation supply. To overcome this problem, proper land management and an understanding of land suitability are needed to make it loose and suitable for shallots. In addition, farmers need to know whether their land is ideal for the crops they will grow. Nowadays, more and more studies are using machine learning models based on land use data as an efficient way to map land suitability [4].

The application of the KNN algorithm can help predict the suitability of shallot farming land by utilizing nearest neighbor data to determine the class or category that matches the land's characteristics. Thus, it aims to utilize machine learning technology to support the development of shallot farming.

Based on Scopus data searches from 2018 to 2023, research related to machine learning agriculture is still widely found with trends, as seen in Figure 1.

Figure 1 shows a graph of the number of agricultural machine learning researches that continue to increase yearly. This indicates that machine learning has great potential to be applied in agriculture. Machine learning research in agriculture can be used for various purposes, such as land suitability classification, crop yield prediction, and pest and disease control. Such research can help farmers to increase agricultural productivity and reduce production costs. Especially in Selaawi, Indonesia, agricultural machine learning research can be used to develop a more accurate land suitability classification system. The land suitability classification system can help farmers select crops suitable for their land conditions. This can increase agricultural productivity in the Selaawi Sub-district. Based on this explanation, agricultural machine-learning research still has excellent potential to be developed and sustained in the future. The research can provide significant benefits for farmers and society as a whole.

1.png

Figure 1. Research trends in machine learning utilization for agriculture 2018–2023

The main contributions of this research are as follows:

•Identification of attributes used to determine the suitability of agricultural land.

•Machine learning model for farmland classification (SMOTE + KNN).

2.png

Figure 2. Dataset details

3.2 Proposed method

Figure 3 is a picture of the proposed methodology used in the machine learning model.

3.png

Figure 3. Proposed method

This research uses a method consisting of several main steps: Data Cleansing, Normalization, SMOTE, K-Nearest Neighbors Hyperparameter Optimization Using GridSearchCV, and Evaluation. The following is an explanation of the flow:

Data cleansing: Data cleansing is correcting errors and incomplete data to be used properly. In this process, missing or invalid data is identified, removed, or corrected. One example of data that is removed is climate data in the dataset because the climate data read by the sensor is not valid. This can prevent inaccurate or incomplete data from spreading and have a negative impact on the use of subsequent data, which can lead to errors in decision-making. With the application of data cleansing, it can improve data quality through data cleaning [10].

Normalization: Normalization is a method used in the preprocessing process of each data sample. This method is performed on datasets that have many 0 values with attributes of different scales. Normalization is used to overcome data outliers [11]. Outlier data refers to observations that have numerical values that are significantly far from other data or look very different from other sample members when the observation occurs [12]. The MinMax Scaler method is used in the normalization process carried out by the author. The MinMax Scaler method is a method in preprocessing used for feature transformation, which scales each feature individually with a specific range. MinMax Scaler is done by subtracting the sample with the smallest sample value on the feature. It will be divided by the largest sample value on the feature that has been reduced by the smallest sample value on the feature [13]. The following is the formula for the Min-Max Scaler method.

$Y=\frac{Y-Y\min }{Y\max-Y\min }$           (1)

where, Y = sample value.

SMOTE: Synthetic Minority Oversampling Technique (SMOTE) is one of the derivatives of Oversampling. The SMOTE technique works by creating a replication of the minority data. The replication is known as synthetic data. This method works by finding the nearest neighbors for each data set in the minority class. After that, synthetic data is created as much as the desired percentage of duplication between the minor data and the k-nearest neighbors chosen randomly [14]. This technique improves the representation of the minority class so that the model can learn better in handling class imbalance. The synthesized data was created based on the K-Nearest Neighbor. All variables used in this research dataset are numeric variables, and the distance between the minor classes is calculated using the Value Difference Metric (VDM) formula as follows [15].

$\Delta(A, B)=\sum_{i=1}^n \delta\left(V_{1 i}, V_{2 i}\right)$           (2)

where,

∆(A,B): Distance between observation A and observation B.

N: Number of independent variables.

δ(V_1i,V_2i ): Distance between observations A and B for each calculated variable.

The mathematical formula in determining the distance between observations A and B for each variable is as follows [15].

$\delta\left(V_1, V_2\right)=\sum_{i=0}^n \frac{C_{1 i}}{C_1}-\frac{C_{2 i}}{C_2}$            (3)

where,

N: Number of categories in the i-th variable.

C_1i: The number of the 1st category included in the i-th variable.

C_2i: Number of 2nd categories included in the i-th variable.

C₁: The number of times the 1st category occurs.

C₂: The number of times the 2nd category occurs.

K-Nearest neighbors hyperparameter optimization using GridSearchCV: Hyperparameter optimization is a step in machine learning models that involves finding the best values for parameters. This process involves minimizing the function Ψ(λ) with respect to λ ∈ Λ, also called the response surface. The function Ψ(λ) is a function that represents the response surface or search space Λ, which must be optimized to find the best fit λ. The variable λ is indexed by Λ, which is the set of configurations that must be optimized in the learning process [16]. Grid search is an approach to parameter tuning that methodically builds and evaluates a model for each combination of algorithm parameters specified in a grid [17]. GridSearchCV is part of the scikit-learn module that validates multiple models automatically and systematically, providing hyperparameters for each model.

In this case, the GridSearchCV KneighborsClassifier is used where the method is a step to find the optimal value of the parameter K (number of nearest neighbors) in the KNN algorithm. The process involves varying the value of K and evaluating the performance of the model using cross-validation techniques. The result is the K parameter that provides the best performance for the KNN model. In this study, various values of K for the n_neighbors parameter in the K-NN algorithm were evaluated. The K values tested were 3, 5, 7, 9, 11, 13, and 15. In addition, several other parameters, such as weights and metrics, were also tested. Weights have two options, namely 'uniform' and 'distance,' while metric has several options, such as 'Euclidean,' 'Manhattan,' 'Chebyshev,' and 'Minkowski.' Based on the evaluation results, here are the accuracy values for various K values (Table 2).

Table 2. Predicted parameters in the model

From Table 3, it can be seen that the best K value that provides optimal performance in the K-NN model for land suitability classification in Selaawi is K = 15 with Euclidean metric and accuracy of 0.9221. This value provides the best balance between bias and variance and avoids overfitting the training data.

Table 3. Parameter accuracy results

Evaluation: Evaluation is needed to measure the performance of a model. There are several matrices to estimate performance, one of which is the confusion matrix. Table 4 shows the confusion matrix for agricultural land suitability where each entry has the following meaning: TP is true positive, FN is false negative, FP is false positive, and TN is true negative.

Table 4. Suitability of agricultural land

Besides the confusion matrix, there are also recall, accuracy, precision, and F1 scores, which are used as tools to evaluate machine learning methods. Recall, accuracy, precision, and F1 score are evaluation metrics that are calculated based on the values in the confusion matrix. The goal is to provide a deeper insight into the model's performance. The following is an explanation and formula for recall, accuracy, precision, and F1 score:

Accuracy: Accuracy is an evaluation matrix to measure how well the model makes correct predictions (True Positive and True negative) from the total predictions that have been made. To calculate the accuracy value of the model can use the following mathematical equation [18].

Accuracy $=\frac{T P+T N}{T P+T N+F P+F N}$            (4)

Precision: Precision is a metric that evaluates how well the model predicts positive classes from the total optimistic predictions (True Positive and False Positive) made. The following is a mathematical equation to calculate precision [19].

Precision $=\frac{T P}{T P+F P}$          (5)

Recall: Recall is an evaluation measure that describes how well a model correctly identifies positive classes. The following mathematical equation can be used to calculate the value of recall [14].

Recall $=\frac{T P}{T P+F N}$           (6)

F1 Score: F1 score is an evaluation metric that shows the balance between precision and recall in the classification model. The value of F1 score will provide information on how well the model has been made in combining precision and recall capabilities so that the F1 score value can also provide an understanding of how effective the model is in classifying. The following mathematical formula generates the F1 score value [20].

$F 1$ Score $=2 \times \frac{\text { Recall } \times \text { Precission }}{\text { Recal }+ \text { Precission }}$            (7)