A Study on Estimating Land Value Distribution for the Talingchan District, Bangkok Using Points-of-Interest Data and Machine Learning Classiﬁcation

: Land is an essential factor in real estate developments, and each location has its unique characteristics. Land value is a vital cost of real estate developments. Higher land costs mean that project developers must create higher valued products to cover the higher land costs and to maintain a proﬁt level from their developments. Land values vary according to surrounding factors, such as environment, social, and economic situations. Machine learning is a popular data estimation technique that enables a system to learn from sample data; however, there are few studies on its use for estimating land value distribution. Therefore, we aim to apply the technique of machine learning to estimate land value and to investigate the factors affecting the land value in the Talingchan district, Bangkok., we used land value level as the dependent variable, with other factors affecting land value levels as the independent variables. Ten points of interest were chosen from Google Places API. Then, three machine learning algorithms, namely CART, random forest, support vector machine, were applied. For this study, we selected 45,032 land parcels as the experimental data and randomly divided them into two groups. The ﬁrst 70% of the land parcels was used to create the training area. The other 30% of the land parcels was used to create the testing area to verify the accuracy of the land value estimation from the applied machine learning techniques. The most accurate machine learning results were produced by random forest, which were then used to measure the factor importance. The academic group factor was school, and the commercial group factors were clothing store, pharmacy, convenience store, hawker stall, grocery store, automatic teller machine, supermarket, restaurant, and company.


Introduction
Since Thailand's 5th National Economic and Social Development Plan, the government has implemented the policy of distributing the main activities from the city center of Bangkok to the central communities of all five perimeter provinces, including Nonthaburi, Nakhon Pathom, Pathum Thani, Samut Prakan, and Samut Sakhon, and thus has continued to spread prosperity to the provincial communities of the country. The critical measure is to distribute some activities to the provinces and to develop a transportation system capable of increasing the connections between Bangkok and the central communities of the provinces, especially the expansions along the BTS/MRT lines; therefore, the economic growth in the metropolitan communities is higher than other communities [1].
Urban expansion is an important phenomenon because it affects political and economic development planning and also affects the livelihoods of people in each community. in recent years, as evidenced by studies on image classification, climate prediction, road traffic injuries, landslide prediction, and poverty prediction etc. [24][25][26][27].
Moreover, many studies have found that the machine learning algorithm technique has higher accuracy than traditional statistical methods [9] and some studies have also used machine learning to estimate house values [28]. We also found that machine learning has been used to determine the factors that affect land value in New York, United States [29], and predicting property prices in Hong Kong [30], but there have been relatively few studies on this type of research. Therefore, from these findings, we aim to apply machine learning techniques to estimate the land value and to identify the factors affecting land value in the Talingchan district, Bangkok.

Study Area
In this study, we chose the Talingchan district as the study area. The Talingchan district is one of fifty areas of Bangkok, located in the outer area on the west side of the Chao Phraya River called "Thon Buri Side", and generally contains rural and agricultural conservation areas mixed with low-density residential areas ( Figure 1). However, currently, the agricultural area is shrinking and decreasing due to construction of residential areas to support city expansion and transportation routes. this technique is that the machine learning can learn from existing data to determine its correct result. These techniques have been widely used in various parts of the world. in recent years, as evidenced by studies on image classification, climate prediction, road traffic injuries, landslide prediction, and poverty prediction etc. [24][25][26][27].
Moreover, many studies have found that the machine learning algorithm technique has higher accuracy than traditional statistical methods [9] and some studies have also used machine learning to estimate house values [28]. We also found that machine learning has been used to determine the factors that affect land value in New York, United States [29], and predicting property prices in Hong Kong [30], but there have been relatively few studies on this type of research. Therefore, from these findings, we aim to apply machine learning techniques to estimate the land value and to identify the factors affecting land value in the Talingchan district, Bangkok.

Study Area
In this study, we chose the Talingchan district as the study area. The Talingchan district is one of fifty areas of Bangkok, located in the outer area on the west side of the Chao Phraya River called "Thon Buri Side", and generally contains rural and agricultural conservation areas mixed with low-density residential areas ( Figure 1). However, currently, the agricultural area is shrinking and decreasing due to construction of residential areas to support city expansion and transportation routes.

Legend
Land parcel

Dataset
In this study, the independent variable was the factor that may affect the land value, and the dependent variable was the assessed land value from the Treasury Department of Thailand; both were accessed from the 2021 dataset.
The independent variable of this study was collected from the Google Places application programming interface (API). The dependent variable of this study is land value data collected from the Treasury Department.
The Treasury Department's Thailand has land value estimation for roadside parcels only, and there is no reference land value estimation for all land parcels. The total number of land parcels in Talingchan district has 138,963 parcels (100%) are divided into 32.4% or 45,032 plots, which are roadside parcels, use for the experiment. Other parcels (67.4% or 93,931), which are not roadside, are not being use for the experiment.

Dataset
In this study, the independent variable was the factor that may affect the land value, and the dependent variable was the assessed land value from the Treasury Department of Thailand; both were accessed from the 2021 dataset.
The independent variable of this study was collected from the Google Places application programming interface (API). The dependent variable of this study is land value data collected from the Treasury Department.
The Treasury Department's Thailand has land value estimation for roadside parcels only, and there is no reference land value estimation for all land parcels. The total number of land parcels in Talingchan district has 138,963 parcels (100%) are divided into 32.4% or 45,032 plots, which are roadside parcels, use for the experiment. Other parcels (67.4% or 93,931), which are not roadside, are not being use for the experiment.

Points-of-Interest Data
In this study, we mainly used the Google Places application programming interface (API) to collect the factors, since we found that the points-of-interest data from Google Places API were always up to date, unlike the databases from other organizations that were updated periodically. Through examination of documents, we identified the related factors for land value, which included school, fashion store, pharmacy, grocery store, convenience store, automated teller machine (ATM), company, market, and restaurant, after we derived the points-of-interest from Google Places API, and then estimated the spatial density using kernel density estimation to determine the density of the related factors (Figures 2 and 3). In this study, we mainly used the Google Places application programming interface (API) to collect the factors, since we found that the points-of-interest data from Google Places API were always up to date, unlike the databases from other organizations that were updated periodically. Through examination of documents, we identified the related factors for land value, which included school, fashion store, pharmacy, grocery store, convenience store, automated teller machine (ATM), company, market, and restaurant, after we derived the points-of-interest from Google Places API, and then estimated the spatial density using kernel density estimation to determine the density of the related factors (Figures 2 and 3).

Schools
Fashion stores

Pharmacy Hawker stalls
Grocery stores Convenience Stores

Land Parcel and Land Value Data
The land parcel data were obtained from the Department of Lands for the dependent variable in polygon format. The land value data were obtained from the Treasury Department in datasheet format. The features and tables of the land value dataset were joined with the land parcel ID in GIS software. The land value dataset includes only the essential road in the district (Figure 4). The lowest-priced land value parcel (blue parcel) was close to the highest-priced land value parcel (red parcel) because the lowest-priced land value parcel was a plot in a small alley from the major road and this dataset was derived from the Treasury Department.

Land Parcel and Land Value Data
The land parcel data were obtained from the Department of Lands for the dependent variable in polygon format. The land value data were obtained from the Treasury Department in datasheet format. The features and tables of the land value dataset were joined with the land parcel ID in GIS software. The land value dataset includes only the essential road in the district (Figure 4). The lowest-priced land value parcel (blue parcel) was close to the highest-priced land value parcel (red parcel) because the lowest-priced land value parcel was a plot in a small alley from the major road and this dataset was derived from the Treasury Department. Descriptive statistics of the independent variables of the Talingchan district, Bangko are given in Table 1 and show the number of points-of-interest in the study area. The tab includes the number of points in the POI datasets, from school to grocery store, in poi units. Descriptive statistics of the independent variables are shown in Table 1. Descriptive statistics of the dependent variables or land value of the Talingchan di trict, Bangkok are provided in Table 2 and show the frequency of the land value level The table includes the number of parcels in land parcel units.  Descriptive statistics of the independent variables of the Talingchan district, Bangkok are given in Table 1 and show the number of points-of-interest in the study area. The table includes the number of points in the POI datasets, from school to grocery store, in point units. Descriptive statistics of the independent variables are shown in Table 1. Descriptive statistics of the dependent variables or land value of the Talingchan district, Bangkok are provided in Table 2 and show the frequency of the land value levels. The table includes the number of parcels in land parcel units.

Modeling Framework
We selected 45,032 land parcels as the experimental data for this study, or 32.4% of the 138,963 land parcels throughout the Talingchan district. These were the land parcels that had land value data from the Treasury Department. Then, the other independent factors that affected land value were verified from the primary documents and added to the experimental data. The next step was to divide the experimental data into two groups randomly. The first 70% of the land parcels was used to create the training area. The other 30% of the land parcels was used to create the testing area to verify the accuracy of the land value estimation from applied machine learning techniques. Then, we compared the accuracy of all of the models to find out which one had the highest accuracy and to identify the factors that were related to the model and the land values ( Figure 5).

Imbalanced Data
The dataset of this study faced the imbalanced data problem. The synthetic minority oversampling technique (SMOTE) is a unique data sampling technique for additional sampling instead of using existing data. SMOTE was applied to synthesize the new data from the existing data using the nearest neighbor principle to expand the model's decision boundary, affecting the existing data's average and standard deviation value. In this study, there were imbalanced data in each land value level, which could affect the learning process of the machine learning algorithm, and therefore the study's accuracy. The training area distribution of each land value level and the factors from additional sampling are shown in the detail in Table 3.

Imbalanced Data
The dataset of this study faced the imbalanced data problem. The synthetic minority oversampling technique (SMOTE) is a unique data sampling technique for additional sampling instead of using existing data. SMOTE was applied to synthesize the new data from the existing data using the nearest neighbor principle to expand the model's decision boundary, affecting the existing data's average and standard deviation value. In this study, there were imbalanced data in each land value level, which could affect the learning process of the machine learning algorithm, and therefore the study's accuracy. The training area distribution of each land value level and the factors from additional sampling are shown in the detail in Table 3.

CART
The classification and regression tree (CART) is a predictive tree model for investigating data structure. It creates visualized decision rules for predicting a categorical and continuous variable. The regression tree does not generate classes of dependent variables, unlike the classification tree (which splits the input space of many variables into subspaces, each subspace associated with a specific class of output variable). However, in the matrix of the independent variables, dependent variables represent the response values for each observation. Because regression trees do not contain preassigned classes, the stage's output is a response value for each of the new dependent variable observations. The squared residuals minimization procedure is used to make the splitting rule in regression trees, which indicates that the predicted sum variances for two resulting nodes should be minimized.
The classification and regression tree, proposed by Breiman et al. [31], is one of the most widely used approaches for dealing with classification and regression issues. CART models execute the Gini and the least-squared deviation measurements for categorical and numerical issues, respectively [32]. Let the pth sample be illustrated as I p,1 , I p,2 , . . . . . . I p,n . . . O p , where I p,n is the value of the pth sample with n features, and O p is the corresponding output value of the sample. The minimization of the least-squared deviation measure of impurity given by Equation (1) serves as a choice to decide the split-up of trees into branches for a CART regression issue.
U r and U l are training data sets pointing to the right and left child nodes, respectively, and Nis the total number of training samples. The output values of the right and left nodes are represented by O and O l respectively.

Random Forest
The random forest classifier comprises many tree classifiers, each of which is created using a random vector sampled separately from the input vector, each of which casts a unit vote for the most popular class to categorize an input vector [33]. The random forest classifier employed in this work grows a tree by randomly selecting or combining features at each node. For each feature/feature combination selected, bagging, a method for generating a training dataset by randomly drawing with replacement N instances, where N is the size of the original training set, was employed. Any instances are categorized by selecting the class with the highest votes from all the tree predictors in the forest (Breiman). The decision tree design process necessitated the selection of an attribute selection measure and a pruning technique. There are several ways to select characteristics for decision tree induction, and most approaches explicitly give a quality measure to the attribute.
The information gain ratio criteria and the Gini index are the most often utilized attribute selection metrics in decision tree induction.
The random forest classifier employs the Gini index as an attribute selection metric, quantifying an attribute's impurity about the classes. For a given training set T, the Gini index can be expressed as: where f (C i , T)/|T| is the probability that the selected instance belongs to class C i . Each time, a tree is built to its maximum depth utilizing a mix of features on new training data. These mature trees are unpruned, which is a significant benefit of the random forest classifier over other decision tree techniques.
The findings indicate that the pruning strategies used, rather than the attribute selection criteria, affect the performance of tree-based classifiers [34].
3. Support Vector Machine The support vector machine (SVM) is a machine learning algorithm, first introduced by Vapnik et al. [35]. SVM is a supervised classification method that reshapes the nonlinear environment into the linear and makes it a processable and straightforward class through the generation of hyperplane. The kernel function is the mathematical function that is utilized for data transformation. SVM translates the original input into a high-dimensional feature space using the training dataset. A separating hyperplane is produced between the points of various tree classes in the original space of n coordinates. SVM calculates the maximum separation between classes and creates a classification hyperplane in the center of the maximum margin. If the point is above the hyperplane, it is categorized as +1; otherwise, it is classified as −1. Then, the properties of new data can be used to forecast which group a new record should belong to. Support vectors are the training points closest to the hyperplane. Following the acquisition of the decision surface, new data can be classified. After obtaining the decision surface, it can be used to categorize additional data. The technique is specified over a vector space. The decision surface for linearly separable space is a hyperplane, which can be represented as (Equation (3)): The vector w and constant b are learnt using a training set of linearly separable items, where x is an arbitrary object to be categorized. SVM was suggested to handle a linearly restricted quadratic programming problem such as Equation (4), with the result that the SVM solution is always globally optimum.
min ω 1 2 with constraints The original input data is transformed into a higher dimensional space using a nonlinear mapping for linearly inseparable objects, and the linearly separating hyperplane can also be found in the new space without increasing the quadratic programming problem's computation complexity by using the kernel function [36]. To put it another way, to compute the similarities between the vectors in higher dimensions space for the linearly inseparable issue, the kernel function is utilized to deduce these similarities in the original lower dimensional space.

Model Evaluation
Contamination between the data types can be verified and analyzed from the confusion matrix table, which is the table that brings the results from data type classification from any method to overlap with the area with the related fact data, and then compares the accuracy between them. The multi-class confusion matrix includes the following Figure 6 [37]:

Class2
Class n Figure 6. The multi-class confusion matrix.
The confusion elements for each class are shown by (Equations (6)-(9)): True Positive (TP) was the actual value, and the predicted value should be the same. False Positive (FP) was the sum of the corresponding rows values except for the TP value.
False Negative (FN) was the sum of values of corresponding column except for the TP value.
True Negative (TN) was the sum of values of all columns and rows except the values of that class for which we calculate the values.
After knowing the TP, FP, FN, and TN values, check a p-value and Kappa coefficient index.
A p-value was used to evaluate the significance of the results in comparison to the null hypothesis when carrying out statistical tests. The null hypothesis asserts that no link exists between the two variables under investigation (one variable does not affect the other). The alternative hypothesis asserts that the independent variable did influence the dependent variable, and the results are important in terms of corroborating the theory under investigation. Statistical significance is defined as a p-value less than 0.05 (usually 0.05). It shows significant evidence against the null hypothesis since the likelihood of the null being true is less than 5%. As a result, the null hypothesis is rejected, and the alternative hypothesis is accepted.
Cohen's Kappa coefficient, which is frequently used to measure reliability, may be used for training-testing reliability. The Kappa coefficient reflects the degree of agreement between the frequencies of two sets of data obtained on two distinct occasions in trainingtesting. The scale of Kappa value interpretation was shown in Table 4. [38].

Kappa
Interpretation <0% No agreement 0.01%-20% Slight 21%-40% Fair 41%-60% Moderate The confusion elements for each class are shown by (Equations (6)- (9)): True Positive (TP) was the actual value, and the predicted value should be the same. False Positive (FP) was the sum of the corresponding rows values except for the TP value.
False Negative (FN) was the sum of values of corresponding column except for the TP value.
True Negative (TN) was the sum of values of all columns and rows except the values of that class for which we calculate the values.
After knowing the TP, FP, FN, and TN values, check a p-value and Kappa coefficient index. A p-value was used to evaluate the significance of the results in comparison to the null hypothesis when carrying out statistical tests. The null hypothesis asserts that no link exists between the two variables under investigation (one variable does not affect the other). The alternative hypothesis asserts that the independent variable did influence the dependent variable, and the results are important in terms of corroborating the theory under investigation. Statistical significance is defined as a p-value less than 0.05 (usually 0.05). It shows significant evidence against the null hypothesis since the likelihood of the null being true is less than 5%. As a result, the null hypothesis is rejected, and the alternative hypothesis is accepted.
Cohen's Kappa coefficient, which is frequently used to measure reliability, may be used for training-testing reliability. The Kappa coefficient reflects the degree of agreement between the frequencies of two sets of data obtained on two distinct occasions in trainingtesting. The scale of Kappa value interpretation was shown in Table 4. [38]. The accuracy of the model is based on the following Equations ((10)- (14)): Positive prediction value = TP i TP i + FP i (12) The ratio of correctly identified positives to the total number of positive samples is known as sensitivity, or actual positive rate or recall. Sensitivity is a crucial metric to assess and compare classifiers with since it indicates the correct classification rate of the class.
The ratio of properly categorized negatives to the total number of negative samples is specificity. Specificity is similar accuracy because the number of occurrences is less.
The ratio of correctly identified positives to the total number of samples labelled as positives is the precision or positive predictive value.
The ratio of correctly identified negatives to the total number of samples classified as negatives is called fallout, sometimes known as the false-positive rate. It is a rate that complements specificity and shows the percentage of "false alarms." The capacity of a classification test to accurately identify or exclude an outcome is measured by accuracy (ACC), and it is the proportion of correct predictions to total samples. When the dataset is significantly skewed, overall accuracy is insufficient to describe the model's performance because overall accuracy might be more significant when most samples are categorized into the majority class.
In addition, the average absolute percentage error (MAPE) was used to assess the effectiveness of the machine learning model in terms of land value price. The mathematical formula of MAPE is Equation (15): where N is the amount of predicting periods, V t is the actual value at period t and P t is the predicting value at period t.

GINI Index
The GINI index, created by Breiman, determines the purity of a specific class following a split along with a particular characteristic [22]. The best split increases the purity of the sets resulting from the split. If L is a dataset with j different class labels, GINI is defined as Equation (16): where p i is relative frequency if class i in L. If the dataset is split on attribute A into two subsets L1 and L2, with sizes N1 and N2 respectively, GINI is calculated as Equation (17): Reduction in impurity is calculated as Equation (18):

Validating the Land Value Predictive Value Obtained by Applying the Machine Learning Technique Model
The land value predictive values obtained by applying the machine learning models were validated with the multi-class confusion matrix to evaluate the accuracy of the prediction. Tables 5-7 show the multi-class confusion matrix of each model.  Table 6 shows the CART model, the land value ranging from 0-15,000 THB/Wah 2 row 1 column 1 was true positive (322), the false positive value of the price was 73 (2 + 42 + 12 + 17), the false negative value was 35 (6 + 18 + 11), and true negative was 13,111.
Regarding the accuracy of the confusion matrix accuracy of each model, it was found that the RF model had the highest accuracy of 96.9%, followed by the SVM model with a maximum accuracy of 90.26%, and the CART model had a precision of only 62.56%. For its part, by using Cohen's Kappa statistics or Cohen's Kappa coefficients, it was found that the RF model had the highest such values at 95.66%, followed by 86.87% and 54.3%. The confusion matrix of each model found that RF models had the highest precision and consistency values, followed by SVM and CART, respectively, and all models had p-values below 0.000000000000000000022.   In addition, from the confusion matrix table of all models, sensitivity, specificity, positive prediction value, and negative prediction value can be calculated. In Table 8, the example of calculating the sensitivity of the CART model value of 0-15,000 THB/Wah 2 was 0.902 (322/(322 + 35)). Meanwhile, the specificity of the CART model of the price was 0.99 (13,111 /(13,111 + 73)). The positive prediction of the price's CART model was 0.815 (322/(322 + 73)). Finally, the negative prediction of the CART model of the price was 0.997 (13,111/(13111 + 35)). Then, the sensitivity, specificity, positive prediction value, and negative prediction value of all models and price estimation in these methods can be calculated.  Figure 7 shows the sensitivity values distribution of all models. The table found that RF and SVM models found high sensitivity values at all land value estimations except those at levels 32,001-40,000 THB/Wah 2 , SVM was significantly less than the RF model. The CART model has low sensitivity.       Figure 8 shows the specificity values distribution of all models. According to the table, all models have high specificity values at all land value estimation levels. Specifically, the RF model finds a value equal to 1 at all land value estimation levels.      Figure 10 shows the negative prediction value distribution of all models. In the negative prediction value, all models had high negative prediction values. Specifically, RF found a value equal to 1 at almost every land value distribution. Except for 32,001-40,000 THB/Wah 2 . Negative Prediction Value    Figure 10 shows the negative prediction value distribution of all models. In the negative prediction value, all models had high negative prediction values. Specifically, RF found a value equal to 1 at almost every land value distribution. Except for 32,001-40,000 THB/Wah 2 . Lewis [39] claims that MAPE values of less than 10% indicate a high level of accuracy, while values of 10% to 20% indicate a reasonable level of accuracy. Thus, RF (1.906) and SVM (5.242) are a high level of accuracy. At the same time, CART (20.391) is a reasonable level of accuracy. Table 9 shows the average MAPE value of machine learning models. Negative Prediction Value Lewis [39] claims that MAPE values of less than 10% indicate a high level of accuracy, while values of 10% to 20% indicate a reasonable level of accuracy. Thus, RF (1.906) and SVM (5.242) are a high level of accuracy. At the same time, CART (20.391) is a reasonable level of accuracy. Table 9 shows the average MAPE value of machine learning models. After all models have been validated, each model's predicted probabilities of land value estimation can be removed in a data frame format. Then, the testing feature class and predicted probabilities data frame are joined together using featured ID field by GIS programming. After successful joining, the estimated land value map is created. According to the estimation land value, using points-of-interest data and the machine learning classification algorithms (Figure 11), the RF ( Figure 11C) and SVM ( Figure 11D) maps are the closest to the testing map. Meanwhile, the CART ( Figure 11B) map was less accurate than the testing map ( Figure 11A).

The Factors That Influence the Land Value Parcel
We describe which predictor factors the RF model found to be the most important for predicting land value. Then, factor importance is measured using RF with the mean decrease in Gini method. The most important factor from the algorithm is school, and the commercial group also appears to be important to the model, including a clothing store, pharmacy, convenience store, hawker stall, grocery store, automatic teller machine, supermarket, restaurant, and company, as shown in Figure 12.
value estimation can be removed in a data frame format. Then, the testing feature class and predicted probabilities data frame are joined together using featured ID field by GIS programming. After successful joining, the estimated land value map is created. According to the estimation land value, using points-of-interest data and the machine learning classification algorithms (Figure 11), the RF ( Figure 11C) and SVM ( Figure 11D) maps are the closest to the testing map. Meanwhile, the CART ( Figure 11B) map was less accurate than the testing map ( Figure 11A

The Factors That Influence the Land Value Parcel
We describe which predictor factors the RF model found to be the most important for predicting land value. Then, factor importance is measured using RF with the mean decrease in Gini method. The most important factor from the algorithm is school, and the commercial group also appears to be important to the model, including a clothing store, pharmacy, convenience store, hawker stall, grocery store, automatic teller machine, supermarket, restaurant, and company, as shown in Figure 12.

Discussion
Wen et al. [15] studied the education institutions via the real estate market. They found that education facilities had a very high effect on real estate value, especially when the quality of the school was considered with the real estate value. Many studies [40] have found that house values in South Korea are affected by education facilities.
Hu et al. [10] found that stores and commercial groups had a very high effect on the

Discussion
Wen et al. [15] studied the education institutions via the real estate market. They found that education facilities had a very high effect on real estate value, especially when the quality of the school was considered with the real estate value. Many studies [40] have found that house values in South Korea are affected by education facilities.
Hu et al. [10] found that stores and commercial groups had a very high effect on the land value, especially when the commercial group was within 15 min of walking. Furthermore, the results were also produced using machine learning, and they found that the random forest regressor (RFR) and extra-trees regressor (ETR) produced the best house rent predictions. Wu et al. [41] also found that random forest produced the best classification results related to a real estate asset.
The results mentioned above found that education facilities, stores, and commercial groups affected real estate values. In this study, the predictive results were accurate by using data generated from Google Places API to solve outdated data from the government sector that was prepared the following fiscal year. We used Google Places API to solve this problem, such as a study by Wu et al. [42] that found that using POI data from the private sector produced data analysis results with more accuracy.
The MAPE value of past research is listed in Table 10. The RF model in this work outperforms several prior models in terms of MAPE, as seen in the table. However, Pai et al. [43] models were better than this work, but this work uses Google Places API to solve the problems. Google Places API is convenient and up to date. It is an exciting issue if researchers want to solve the problem without relying on a dataset that requires government permission. In this issue, other researchers need to consider the pros and cons of the methodology between Pai's and this work. Various factors might affect real estate success, including countries, cultures, market trends, and economic situations. As a result, this research presents a viable and comparable option for land value estimation to retain stability and feasibility over time; forecasting models should be modified and enhanced regularly.

Conclusions
Using a machine-learning algorithm produces more accurate study results. Moreover, machine learning can tune parameters to determine the appropriate value via experiment, which allows the method to differentiate the data appropriately.
In this study, we found that machine learning algorithms provided accurate predictive results and displayed the factors that affected the accuracy of the predictive land values for each land parcel. The advantages of this study method are (1) machine learning can predict the land value for each land parcel, (2) the method identifies the factors that affect the land value prediction for each land parcel, (3) machine learning is a helpful tool to support decisions, and (4) machine learning could improve the prediction of the land value on a macro level in the future.
This study's limitation is that the land value data from the Treasury Department is in a static format rather than a dynamic format and must be updated regularly. If the Treasury Department uses API land values and publishes them online, future studies should be beneficial.
The factors used in this study could be improved by using alternative factors, such as the number of transit stations, public facilities, street networks, and walkability that may play essential roles in predicting land value estimations.
This study confirms that machine learning can produce land value prediction results that are more accurate. Furthermore, point of interest data from Place API is useful to estimate land value distribution. The future works, machine learning, and point of interest data from Place API could be applied in other fields of science to produce more information. In the future, such data and techniques can be used to predict the direction of urban expansion, find public utility needs, and predict crime locations.