An Approach for Mapping Ecotourism Suitability Using Machine Learning: A Case Study of Zhangjiajie, China

Abstract

The assessment of ecotourism suitability is crucial for sustainable regional development and is seen as an effective strategy to achieve both environmental protection and economic growth. One of the key challenges in land research is effectively identifying potential ecotourism resources while balancing regional protection and development. This study mapped the suitability of ecotourism in Zhangjiajie, China, using a combination of various geospatial data sources and four machine-learning techniques. Additionally, an indicator system was developed, covering the ecological environment, geological geomorphology, socioeconomics, and resource availability. The prediction results for suitability classified the area into four categories: highly suitable, moderately suitable, marginally suitable, and unsuitable; based on the ensemble results generated by the four algorithms, these categories accounted for 19.34%, 28.78%, 23.87%, and 28.01% of the total area, respectively. This study’s findings illustrate the spatial distribution of ecotourism suitability in Zhangjiajie, providing valuable insights for identifying potential ecotourism resources as well as informing regional planning and policy-making.

1. Introduction

Tourism is one of the fastest-growing sectors worldwide [1,2] and plays a vital role in national economies. However, its growth has led to serious environmental problems and increased socioeconomic costs. The most significant issue is the environmental damage caused by tourism development, especially in ecologically sensitive and fragile natural areas [2]. Therefore, it is crucial to adapt the tourism industry to the demands of the new era and promote sustainable tourism development. Globally, there is an increasing emphasis on ecotourism as a new tourism model that balances economic development and environmental protection, in accordance with the United Nations Sustainable Development Goals (SDGs) [3].

The relationship between ecotourism and the UN SDGs can facilitate sustainable regional development [4,5]. Ecotourism, defined by the International Ecotourism Society (TIES) as ‘responsible travel to natural areas that conserves the environment, sustains the wellbeing of the local people, and provides a means to promote sustainable development’ [4], emphasizing the moderation of natural and humanistic tourism activities under the premise of environmental protection [6].

Ecotourism helps minimize damage to and the pollution of the natural environment by offering green, low-carbon, and recycled tourism products and services. This promotes harmonious coexistence between humans and nature [7]. The development of ecotourism not only protects biodiversity and ecological balance but also provides residents with more employment opportunities and economic benefits and promotes the sustainable development of the economy [5]. Ecotourism is the most promising form of tourism for the future because of its focus on conservation, communities, and interpretation [8], which can increase local economic income while safeguarding the environment [9]. According to UN Tourism, global tourism revenue will reach USD 3.3 trillion in 2023, accounting for 3% of global GDP.

Assessing an area’s suitability for ecotourism is an important first step in developing ecotourism resources. It is essential to use scientific rigor and accuracy to achieve sustainable development. Suitability analysis involves evaluating how well a specific land use type is suited to a particular region [10]. Ecotourism suitability evaluation takes into account various aspects, such as the natural environment [11] as well as socioeconomic [12], resource availability [6], and cultural connotations [13]. Multicriteria approaches that integrate multiple sources of spatial data and corresponding GIS spatial analysis techniques have been widely used to assess the ecotourism potential of various locations [9]. In terms of research methods, most existing ecotourism suitability evaluations have relied on GIS and remote sensing (RS) [1,3,10,14,15,16], analytic hierarchy process (AHP) [12,15], multiple criteria decision analysis (MCDA) [1,3,14,16], multiple-attribute decision making (MADM) [9,12], strengths, weaknesses, opportunities, and threats (SWOT) [6], principal component analysis (PCA) [9], and Delphi [17,18], among others. These methods typically involve determining factor weights for comprehensive consideration and integrated evaluation of various influencing factors.

When assessing regional ecotourism potential, previous studies have combined various methods with GIS, the most common combination being GIS with AHP [1,11]. For instance, the scholarly literature [15] used fuzzy AHP (F-AHP) alongside GIS to determine feature weights, while the scholarly literature [12] used AHP and ordered weighted averaging (OWA) for MADM modeling and a generic grading scale in a GIS environment to assess suitability across the study area. Research frameworks that combine GIS and MCDA for ecotourism suitability assessments have also been widely used [3]. For example, the scholarly literature [16] combined GIS and MCDA to assess the suitability of ecotourism in mountainous landscapes [16]. Some researchers have combined these methods with other analytical methods. For example, the scholarly literature [6] integrated SWOT with MCDM to assess the potential of ecotourism areas with natural and cultural features.

The scholarly literature [9] used PCA combined with clustering algorithms to identify potential sites for ecotourism development. They revealed a significant contribution of waterbody variability to ecotourism development. The scholarly literature [17] used the fuzzy Delphi method and expert consultation to establish a cross-dimensional indicator system for the evaluation of sustainable ecotourism, offering valuable insights into the complex dynamics of ecotourism in the Belize region. Integrating GIS has become a key method for evaluating the suitability of regional ecotourism. For example, the scholarly literature [14] employed a GIS overlay to create ecotourism suitability maps, and their classification method provided important insights. Some researchers have also integrated spatial information into interactive platforms. For example, the scholarly literature [5] developed a geospatial decision-support system, which serves as an interactive platform for planners, government agencies, and tourists. However, such platforms have limitations in providing relevant geospatial information and are not conducive to mapping ecotourism suitability.

These methods have significantly contributed to the field of ecotourism suitability assessment; however, the potential lack of objectivity in factor weighting of the current evaluation system may result in various contradictions and challenges in ecotourism planning, implementation, and management. During the planning phase, the subjective or unilateral determination of certain key factor weights may overemphasize economic benefits at the expense of ecological conservation [19], which could adversely impact the long-term health and sustainability of ecotourism, irreparably harming local ecosystems. In the implementation phase, non-objective factor identification may result in inefficient resource allocation. Biases in factor inputs can result in resource wastage and misallocation during the construction of ecotourism facilities, thereby reducing the overall benefits of ecotourism [20]. During the management phase, conflicts of interest and social contradictions may arise because of the involvement of multiple stakeholders, including the government, local residents, businesses, and tourists [12,20]. For example, some residents may lose traditional livelihoods or living environments due to the development of ecotourism, whereas some enterprises may cause social dissatisfaction due to the excessive benefits they gain.

Artificial intelligence (AI) technology has facilitated the development of novel approaches and methodologies for ecotourism suitability evaluations, thereby offering substantial benefits in the processing and analysis of ecotourism data. The scholarly literature [21] leveraged machine learning and social media big data to identify patterns of cultural ecosystem services and tourist emotions in Hong Kong’s coastal areas. The scholarly literature [22] leverages social big data to effectively elucidate spatial access patterns and hotspots pertaining to nature tourism, thereby furnishing novel insights into the management of protected areas and augmenting participatory methodologies in practice. However, despite their effectiveness in analyzing macro-level causes, predicting ecotourism market trends, and exploring tourists’ emotional patterns, these methods fail to provide a conclusive map of ecotourism suitability that meets the demand for regional planning, development, and conservation. In this study, to address the challenges and issues faced, we used a machine-learning-based framework to assess the suitability of ecotourism. This framework considered the ecological environment, geology and landforms, humanities and society, and resource availability as the main factors influencing ecotourism suitability mapping. This allowed the evaluation process to gather information, recognize patterns, establish regularities, and make precise predictions based on a wide range of geographical features. By combining AI technology with real-world research, our aim was to achieve a comprehensive and accurate assessment of ecotourism suitability.

2. Study Area and Data

2.1. Study Area

Zhangjiajie, situated in the northwestern part of Hunan Province, China, boasts a unique geographical location with a range of 109°40′–111°20′ E longitude and 28°52′–29°48′ N (Figure 1). The region predominantly features mountainous terrain and is globally acclaimed for its distinctive Zhangjiajie landforms, including the peak sandstone forest landforms. The area experiences a humid subtropical monsoon climate characterized by ample precipitation and simultaneous rain and heat, fostering lush vegetation growth and a diverse range of flora and fauna. Zhangjiajie is home to several rare wildlife species and, as an internationally recognized ecotourism destination, is graced with the distinctions of being a World Geopark, a World Natural Heritage site, and a National Forest Park. The exceptional natural scenery and cultural resources attract a substantial annual influx of domestic and international tourists. In 2023 alone, the city welcomed 36.274 million visitors, generating a total tourism revenue of RMB 51.46 billion, accounting for 83.8% of the region’s gross product.

2.2. Data Source

This study used multiple sources of geospatial data to evaluate ecotourism suitability, as outlined in

Table 1. Data source and details.

Data	Year	Resolution	Source	Description
DEM	2023	30 × 30 m	https://www.usgs.gov/ (accessed on 15 April 2023)	Basic data for terrain analysis.
Landsat 8	2023	30 × 30 m	https://www.usgs.gov/ (accessed on 8 March 2024)	Used to calculate remote sensing indices such as the NDVI.
Geology	1957–1995	1:200,000	https://www.cgs.gov.cn/ (accessed on 12 March 2024)	Used to characterize the regional geological environment.
Nighttime light	2019	130 × 130 m	http://59.175.109.173:8888/index.html (accessed on 13 March 2024)	The resolution of other data sources was too low; however, the Luojia1-01 data were only updated to 2019.
Meteorology	1980–2022	1 × 1 km	http://loess.geodata.cn/data/ (accessed on 14 March 2024)	Used to calculate the multi-year average temperature and precipitation.
Land use [23]	2022	30 × 30 m	https://doi.org/10.5281/zenodo.8239305 (accessed on 16 March 2024)	The status of land use is a prerequisite and basis for ecotourism planning.
OSM	2023	-	http://www.openstreetmap.org/ (accessed on 1 May 2024)	Road data were extracted to calculate features such as road network density.
WorldPop	2020	100 × 100 m	https://hub.worldpop.org/ (accessed on 17 March 2024)	Global high-resolution population distribution dataset.
POI	2023	-	https://lbs.amap.com/ (accessed on 31 December 2023)	Crawled using Python web crawling techniques.
Ecological value	2000–2020	1 × 1 km	https://www.resdc.cn/ (accessed on 18 March 2024)	Five years of data (2000, 2005, 2010, 2015, and 2020) were averaged.
Forest age [24]	2020	30 × 30 m	https://doi.org/10.5281/zenodo.8354262 (accessed on 20 March 2024)	To obtain the 2023 data, a value of 3 was added to the existing raster data.

. The dataset includes a wide range of geospatial information. Electronic distance measurement (EDM) data, primarily used for analyzing digital terrain, accurately represent the area’s topographic variations. RS data can be used to calculate ecological indices such as the NDVI, providing insights into the area’s ecological characteristics. The nighttime light data show a strong correlation with regional economic development, offering valuable information on the economic aspects of the study area. Meteorological data, including temperature and precipitation, help characterize the region’s climate conditions and are particularly important for understanding the local climate. Geospatial data sources have varying resolutions and applications. To address this, we combined these data sources into a 30 × 30 m raster format for further modeling and analysis.

It is important to emphasize that ensuring the reliability of our model’s training required precise geographic location data for ecotourism attractions. These data were sourced from the esteemed Third Survey and Mapping Institute of Hunan Province. By meticulously extracting feature values for each data point, we ensured that this information effectively contributed to the machine learning model’s training process, thereby enhancing the credibility of our predictive analytics.

3. Methods

3.1. Framework

This study introduces a research framework for assessing the suitability of ecotourism using various geospatial data sources and four machine learning algorithms (Figure 2). The framework consisted of three main components: data acquisition and preprocessing, model training, and model evaluation and prediction. The data acquisition and preprocessing phases, outlined in Section 2.2 and Section 3.3, involved consolidating geospatial data from different sources and creating feature matrices for subsequent modeling. Model training, outlined in Section 4.2, included processing the sample points and training the machine learning model. The data were used to train the machine learning model by extracting the values of each raster layer from the sample points. Model evaluation and prediction are detailed in Section 4.2 and Section 4.3. The results include a comprehensive assessment of the trained machine learning model to ensure its robustness and generalizability. Finally, the prediction results of the four machine learning models were obtained and visualized by inputting all the data into the model. To mitigate the risk of relying on a single model, the predictions from the four machine learning models were combined to produce a final prediction.

Central to ecotourism is the principle of sustainable development. Therefore, constructing an assessment framework for the ecotourism suitability requires a comprehensive consideration of the multifaceted impacts of tourism activities on the local ecological environment. Key factors include soil and water conservation potential and the NDVI, which reflect the ecological environment’s carrying capacity, as well as limiting factors such as slope and relief. These elements are crucial in shaping the evolution of ecotourism. This approach highlights the importance of not only evaluating the economic benefits of ecotourism but also ensuring environmental protection and fostering positive impacts on the local community. Ecotourism aims to mitigate environmental degradation by promoting a harmonious relationship between tourism and the natural world. Our framework utilizes suitability probabilities applied to individual raster cells, offering a high-resolution approach to local ecotourism planning and avoiding overly broad planning strategies. This method clearly distinguishes between conservation and development areas, helping to further minimize the environmental impacts and damage caused by tourism activities.

3.2. Machine Learning

3.2.1. Random Forest (RF)

The RF algorithm, an ensemble learning technique that leverages decision trees to produce accurate predictions [25], integrates the results of multiple decision trees and assigns the category with the highest number of votes as the final output.

$$f\left(x\right)=\underset{Y}{\mathrm{a}\mathrm{r}\mathrm{g}\mathrm{m}\mathrm{a}\mathrm{x}}\sum _{i=1}^k I(h_i(x,{\theta }_i)=Y)$$

(1)

where $f\left(x\right)$ is the final classification result, $h_i(x, {\theta }_i)$ is the decision trees, ${\theta }_i$ is a random variable following independent homogeneous distribution, $I\left(. \right)$ is the indicator function, and $Y$ is the objective variable.

The RF algorithm benefits from the bagging technique, which improves prediction accuracy and increases model diversity using random sampling. RF randomly selects a subset of features for splitting, which introduces an element of randomness that reduces the correlation between features and mitigates the risk of overfitting.

3.2.2. Xtreme Gradient Boosting (XGBoost)

XGBoost is a machine learning algorithm that employs gradient boosting trees. This model was developed by Tianqi Chen [26]. It iteratively adds new decision trees to optimize the objective function. Each added tree focuses on reducing the prediction residuals of the previous tree, thereby improving the overall prediction accuracy. The objective function of XGBoost consists of a loss function and regularization term.

$$\mathcal{L}\left(\varphi \right)=\sum _{i=1}^{I}l\left({\widehat{y}}_i{,y}_i\right)+\sum _{k=1}^K\Omega \left(f_k\right)$$

(2)

where $\mathcal{L}\left(\varphi \right)$ is the objective function, $I$ is the number of training samples, $i$ is the $i$ sample in the sample set, $K$ is the number of boosted trees, $k$ is the $k$ tree in the boosted tree, $l\left({\widehat{y}}_i{,y}_i\right)$ is the loss function that measures the difference between the true label $y_i$ and predicted label $\widehat{y_i}$, and $\Omega \left(f_k\right)$ is the regularity term used to control the model complexity to prevent overfitting and improve the model generalization ability.

$$\Omega \left(f_k\right)=\gamma T+{\displaystyle \frac{1}{2}}\lambda {\left|\right|w\left|\right|}^2$$

(3)

where $\Omega \left(f_k\right)$ is the regularization strength, $T$ is the total number of leaf nodes, $\gamma$ and $\lambda$ are the regularization parameters, and $w$ is the leaf node weight vector.

XGBoost uses a second-order Taylor expansion of the loss function to accurately approximate the true loss function.

$${\tilde{\mathcal{L}}}^{\left(t\right)}=\sum _{i=1}^I\left[g_i{f}_t\left(x_i\right)+{\displaystyle \frac{1}{2}}{h}_i{f}_t^2\left(x_i\right)\right]+\Omega \left(f_k\right)$$

(4)

where ${\tilde{\mathcal{L}}}^{\left(t\right)}$ is the optimized objective function, and $g_i={\partial }_{{\widehat{y}}_i^{\left(t-1\right)}} l\left(y_i,{\widehat{y}}^{\left(t-1\right)}\right)$ and $h_i={\partial }_{{\widehat{y}}_i^{\left(t-1\right)}}^2 l\left(y_i,{\widehat{y}}^{\left(t-1\right)}\right)$ are the first- and second-order gradient statistics of the loss function, respectively.

The XGBoost algorithm demonstrates outstanding predictive performance, scalability, interpretability, and support for parallel computing. It also automatically handles missing values. However, it is essential to fine-tune the algorithm parameters to prevent overfitting and underfitting.

3.2.3. Support Vector Machine (SVM)

The SVM algorithm is a popular supervised machine learning technique that classifies data by identifying an optimal hyperplane [27]. When the data are linearly separable, the problem can be transformed into an optimal solution by solving the Lagrange duality; however, when the data are not linearly separable, a kernel function is introduced to map them to a higher-dimensional space, making them linearly separable.

$$f\left(x\right)=w^T\Phi \left(x\right)+b$$

(5)

where $f\left(x\right)$ is the distance from the sample $x$ to the hyperplane, $w$ is the feature weight vector, $w^T$ is the transpose of the weight vector $w$, $\Phi \left(x\right)$ is the feature space transformation function, and b is the bias term.

SVM is well equipped to handle high-dimensional data and can effectively handle outliers and noise. Additionally, it has strong generalization capabilities because of its ability to convert nonlinear problems into linear ones using kernel functions. However, its performance can be sensitive to parameters such as the kernel and penalty coefficient C. Thus, it is essential to use techniques like cross-validation and grid search to identify the best parameters during experimentation.

3.2.4. Adaptive Boosting (AdaBoost)

The AdaBoost algorithm, originally proposed by Freund and Schapire in 1997 [28], is an iterative approach that combines multiple weak classifiers into a single robust classifier. Its underlying principle is to adjust the weights assigned to both samples and weak classifiers, enabling the classifiers to progressively focus on instances that are difficult to classify, thereby enhancing the overall accuracy of the classification task [29].

$$f(x)=sign\left(\sum _{n=1}^N {\alpha }_n\cdot G_n(x)\right)$$

(6)

where $f\left(x\right)$ is the trained strong learner, $N$ is the number of weak learners, ${\alpha }_n$ is the weight corresponding to the $n$ weak learner, and $G_n\left(x\right)$ is the $n$ weak learner.

$${\alpha }_n={\displaystyle \frac{1}{2}}log{\displaystyle \frac{1-e_n}{e_n}}$$

(7)

The AdaBoost algorithm exhibits robustness, adaptability, and generalizability while maintaining its capacity for effective data fitting even in scenarios with numerous features and intricate data distributions.

3.3. Feature Processing

In our study, we utilized scientific rigor, reliability, and accessible data to classify the factors affecting the assessment of regional ecotourism suitability into four components: ecological environment, geology and landforms, humanities and society, and resource availability. These components were derived from previous studies, which provided valuable insights for our research. Each component was then classified into five distinct characteristics for analysis, as outlined in Table 2.

Table 2. Feature building and information on relevant details.

	Feature	Processing	Description	Reference
Ecological environment (Figure 3)	NDVI	$NDVI={\displaystyle \frac{Red-Green}{Red+Green}}$	$Red$$\mathrm{and} Green$ denote the red and green bands in remote sensing data, respectively.	[9]
	Temperature	Kriging interpolation and masking	Affects the variability of natural landscapes.	[2]
	Precipitation	Kriging interpolation and masking	Affects vegetation ecology and scenic safety.	[30]
	Humidity	Kirchhoff transform	Affects vegetation growth and distribution, which directly affect landscape ecology.	[10]
	Land use	Resample and masking	Affects the scale and quality of ecotourism.	[10]
Geology and landforms (Figure 4)	Geology	Resample and masking	Reflects the region’s geological environment.	[31]
	Elevation	Resample	Reflects surface altitude conditions.	[13]
	Slope	Surface analysis	Reflects surface microstructure.	[13]
	Aspect	Surface analysis	Reflects surface mountain ecology.	[32]
	Relief	$R=H_{max}-H_{min}$	$H_{max}$$\mathrm{and} H_{min}$ denote the maximum and minimum elevation per unit area, respectively.	[15]
Humanities and society (Figure 5)	Population	Resample and masking	Influences the layout of ecotourism development and resource allocation.	[33]
	Nighttime light intensity (NILI)	Resample and masking	It is highly correlated with economic development.	[31]
	Road network density (ROND)	$D_r={\displaystyle \frac{L_r}{S_r}}$	$L_r$$\mathrm{and} S_r$ denote the road lengths and unit sizes, respectively.	[34]
	Distance from roads (DISO)	Euclidean distance	Used to measure regional transport accessibility.	[30]
	Distance from settlements (DISS)	Euclidean distance	It serves as a crucial support node and stakeholder in the advancement of ecotourism.	[2]
Resource availability (Figure 6)	Aesthetic landscape value (ALSV)	Kriging interpolation and masking	Comprehensively reflects the region’s natural and human landscapes.	[35]
	Biodiversity	Kriging interpolation and masking	A prerequisite for the development of ecotourism.	[35]
	Soil conservation capacity (SOCC)	Kriging interpolation and masking	Its strength directly affects the sustainable development of ecotourism.	[32,36]
	Forest age	Resample and masking	Reflects vegetated landscape resources.	[24]
	Distance from river (DISI)	Euclidean distance	Provides water and ecological support.	[9]

Table 2. Feature building and information on relevant details.

	Feature	Processing	Description	Reference
Ecological environment (Figure 3)	NDVI	$NDVI={\displaystyle \frac{Red-Green}{Red+Green}}$	$Red$$\mathrm{and} Green$ denote the red and green bands in remote sensing data, respectively.	[9]
	Temperature	Kriging interpolation and masking	Affects the variability of natural landscapes.	[2]
	Precipitation	Kriging interpolation and masking	Affects vegetation ecology and scenic safety.	[30]
	Humidity	Kirchhoff transform	Affects vegetation growth and distribution, which directly affect landscape ecology.	[10]
	Land use	Resample and masking	Affects the scale and quality of ecotourism.	[10]
Geology and landforms (Figure 4)	Geology	Resample and masking	Reflects the region’s geological environment.	[31]
	Elevation	Resample	Reflects surface altitude conditions.	[13]
	Slope	Surface analysis	Reflects surface microstructure.	[13]
	Aspect	Surface analysis	Reflects surface mountain ecology.	[32]
	Relief	$R=H_{max}-H_{min}$	$H_{max}$$\mathrm{and} H_{min}$ denote the maximum and minimum elevation per unit area, respectively.	[15]
Humanities and society (Figure 5)	Population	Resample and masking	Influences the layout of ecotourism development and resource allocation.	[33]
	Nighttime light intensity (NILI)	Resample and masking	It is highly correlated with economic development.	[31]
	Road network density (ROND)	$D_r={\displaystyle \frac{L_r}{S_r}}$	$L_r$$\mathrm{and} S_r$ denote the road lengths and unit sizes, respectively.	[34]
	Distance from roads (DISO)	Euclidean distance	Used to measure regional transport accessibility.	[30]
	Distance from settlements (DISS)	Euclidean distance	It serves as a crucial support node and stakeholder in the advancement of ecotourism.	[2]
Resource availability (Figure 6)	Aesthetic landscape value (ALSV)	Kriging interpolation and masking	Comprehensively reflects the region’s natural and human landscapes.	[35]
	Biodiversity	Kriging interpolation and masking	A prerequisite for the development of ecotourism.	[35]
	Soil conservation capacity (SOCC)	Kriging interpolation and masking	Its strength directly affects the sustainable development of ecotourism.	[32,36]
	Forest age	Resample and masking	Reflects vegetated landscape resources.	[24]
	Distance from river (DISI)	Euclidean distance	Provides water and ecological support.	[9]

4. Results

4.1. Feature Correlation Analysis

We created a correlation heatmap (Figure 7) to visualize the connection between different geographic features. The correlation between precipitation and elevation was remarkably strong, with a value of 0.87. This indicated that the topography of the Zhangjiajie Mountain Area significantly influences rainfall patterns. In the summer, the southeastern monsoon brings moisture from the Pacific Ocean, leading to topographic rain, which results in abundant water for vegetation growth, promotes biodiversity, and enhances the scenic beauty of the region. The correlation between slope and relief was 0.91, which aligns with the fundamental geographic principles. Biodiversity and ALSV had the highest correlation (0.94), suggesting that areas with rich biodiversity tend to have higher ALSV values.

To apply the geographic features to ecotourism suitability assessment modeling, we integrated the variance threshold and mutual information methods for feature selection. The variance threshold test revealed no redundant features with a variance of zero. Subsequently, we utilized the mutual information method to analyze the relationship between the features and labels and showed that all the features exhibited a positive mutual information estimate, suggesting their relevance to the labels. Based on both analyses, we decided to retain all the features for subsequent modeling, as they contribute to preserving the original information of the data while also ensuring efficient modeling.

4.2. Model Evaluation

A thorough evaluation is necessary during the modeling process to ensure the model’s generalizability and robustness. In this study, we selected five evaluation metrics, namely, accuracy, precision, recall, F1, and area under the curve (AUC), to assess the model’s classification performance under ten-fold cross-validation (

Table 3. Evaluation and comparison of model performance metrics through ten-fold cross-validation.

Metrics	Accuracy	Precision	Recall	F1	AUC
RF	0.8937	0.8959	0.8757	0.8828	0.9610
SVM	0.8717	0.8678	0.8614	0.8626	0.9355
XGBoost	0.9207	0.9176	0.9173	0.9154	0.9721
AdaBoost	0.8938	0.8752	0.9130	0.8909	0.9603

). Furthermore, we visualized the confusion matrix and ROC curves (Figure 8 and Figure 9) to provide a comprehensive understanding of the model’s performance.

As shown in Table 3, XGBoost exhibited superior performance compared to the remaining three types of models under 10-fold cross-validation. AdaBoost and RF followed in second and third place, respectively. In addition, SVM displayed a notable performance gap compared with the other three model types.

To visualize the model’s classification performance, we created a confusion matrix, as shown in Figure 8. XGBoost exhibited the best performance, followed by RF. Although the overall error rate of AdaBoost was slightly higher than that of SVM, the difference was not substantial.

Based on the results of the ROC curve and AUC calculations presented in Figure 9, XGBoost had the best performance, followed by AdaBoost, RF, and SVM. However, the differences between the models were not significant, especially for RF and AdaBoost, whose AUC values differed by only 0.0012. All four models performed well in predicting spatial classifications for the test data features.

4.3. Mapping Results of the Models

The input data were individually fed into four different pre-trained models, which produced the predictions shown in Figure 10. While these models showed a high level of agreement, certain areas had some differences. This suggested that while machine learning techniques can effectively model real-world training data, they may produce varied predictive outcomes in different regions due to differences in the underlying algorithms.

The evaluation results of the models (Table 3 and Figure 9) and the spatial distribution characteristics (Figure 10) indicated that the spatial variability in the model fit was consistent with the actual growth patterns of ecotourism in the region. These four model categories provided insights into the untapped potential of regional ecotourism, which can be visualized using geospatial visualization techniques. Notably, XGBoost exhibited the highest predictive power, followed closely by RF, whereas SVM fell short of the other three types of models, which can possibly be attributed to the fact that tree-based models possess greater learning capacities and are, therefore, better equipped to capture intricate spatial nuances in the data.

4.4. Comprehensive Results

To visualize the predictive performance of the models, two representative regions, namely, the urban area of Zhangjiajie and the Wulingyuan District (an ecotourism hotspot), were selected for comparative analysis (Figure 11).

A comparison of the predictions of the four models in the two selected regions revealed that the urban area of Zhangjiajie (Figure 11a-1–a-4) was not conducive to ecotourism development owing to the prevalence of human activity, with urban construction land and industrial and agricultural land holding leading positions (as indicated by the red color in Figure 11). However, the southern part of the urban area exhibited high suitability for ecotourism development (indicated by the green color in Figure 11), consistent with the current state of ecotourism development in the area. This region encompasses the Tianmen Mountain National Forest Park, which boasts a range of popular tourist destinations, such as Tianmen Cave, Ghost Valley Trail, and Glass Viewing Platform. The area is characterized by mountainous terrain, lush vegetation, rivers, and water systems, with diverse natural scenery, a significant elevation variation, and steep cliffs. The region’s rich ecotourism resources contributed to the formation of the current spatial distribution pattern of ecotourism, which is supported by our predictive findings.

Within the Wulingyuan District, the predictions of the four models exhibited a high degree of consistency, as depicted in Figure 11b-1–b-4. The area is generally well suited for ecotourism, which can possibly be attributed to the unique natural and cultural tourism resources in the region, as well as the protection provided by Zhangjiajie National Forest Park. The region is home to the renowned ecotourism destination, the Wulingyuan Scenic Spot, which features natural wonders such as quartz sandstone peak forest landforms, as well as cultural attractions such as air corridors and the setting of the classic Chinese novel Journey to the West. The spatial distribution of ecotourism attractions in the area aligns closely with the model predictions, further supporting the reliability of the predictive results obtained using the machine learning approach.

4.5. Result Verification

To mitigate the risk of relying on a single model, we computed the average prediction across the four models (Figure 12).

The zones with high suitability were mostly located in the central region of Wulingyuan, southern and southeastern regions of Yongding, northwestern, western, and southwestern regions of Cili, and northern, northwestern, and eastern regions of Sangzhi. These areas were selected to coordinate regional ecotourism resources, including Zhangjiajie National Forest Park, Tianzishan Nature Reserve, Suoxiyu Nature Reserve, and Yangjiajie Nature Reserve. In the southern part of Yongding (south of the urban area of Zhangjiajie), several patches of highly suitable areas were separated from the core urban area. This is related to two factors, first of which is the presence of Tianmen Mountain National Forest Park (Figure 13a), which features a combination of a natural landscape, biological resources, and humanistic history and is rich in ecotourism resources. Second, this area is a transitional zone between the basin topography of the urban area of Zhangjiajie and the southern Tianmen Mountain, resulting in different levels of suitability with clear boundaries. The highly suitable zones of Cili are concentrated primarily in the northwestern (Figure 13h), western, and southwestern regions, which feature mountainous terrain, high vegetation cover, and strong soil and water conservation capacity, coupled with a unique topography and outstanding aesthetic landscape value. In Sangzhi, the distribution patterns of highly suitable zones were similar.

The unsuitable zones were mainly found in the western and northwestern parts of Wulingyuan, northern section of Yongding, central, eastern, and southwestern regions of Cili, and southwestern part of Sangzhi. These areas are relatively small in Wulingyuan and are mostly located in urban areas and settlements on the outskirts of the Wulingyuan scenic area (Figure 14a). They are dominated by residential buildings, water conservation land, and permanent basic farmland, which do not support ecotourism development. The northern part of Yongding is part of the urban area of Zhangjiajie and is characterized by urban land use with a significant amount of farmland on the periphery, including permanent basic farmland essential for food security in China (Figure 14c). Similarly, in Cili, the unsuitable zones are larger due to the area’s topography, resulting in a flat topography that has developed into residential areas and farmland with limited ecotourism resources (Figure 14f,g). In Sangzhi, the unsuitable zones exhibited a northeast–southwest-oriented strip-like pattern influenced by the unique topography created by the local geological environment (Figure 14b).

Zones with moderate suitability exhibited a distribution pattern similar to that of the highly suitable zones, being primarily concentrated around their periphery. For example, the small mountain range in the southern part of the Wulingyuan urban area, Siduping Township and Tianle Mountain Village in Yongding, Wujiadai and Rhinoceros Villages in the southwestern part of Cili, and Linxiyuan and Niudongkou Villages in the northern part of Sangzhi fell into this category. These areas are located at the interface between human-intensive areas and mountainous regions. They offer both protective functions and development potential, with high vegetation cover and excellent soil and water conservation capacity. They also boast unique flora and fauna, making them well suited for future ecotourism activities. The marginally suitable zones, on the other hand, are primarily located in the transition zone between moderately suitable zones and unsuitable zones. These areas are characterized by high human activity and fragile ecosystems and should, therefore, be protected.

4.6. Contributions of Features

Machine learning algorithms rely on the features constructed to learn during the training phase. In tree models, the importance of a feature is determined by the frequency with which it is chosen as a split point or the degree of information gain resulting from the split. In this study, we used the feature importance attributes of the model to calculate and average the feature importance rankings for the RF, XGBoost, and AdaBoost tree models (Figure 15).

According to the model’s average ranking results for feature importance, temperature, precipitation, and soil conservation capacity were the top three features, accounting for 15.44%, 13.69%, and 13.83% of all the features, respectively. These three features are crucial in determining the ecological environment, which in turn affects ecotourism development. Temperature and precipitation are the key factors influencing the attractiveness of ecotourism destinations. Zhangjiajie’s moderate altitude and mountainous terrain create a comfortable environment for tourists, especially during the summer months when the region experiences a large influx of visitors seeking relief from the heat.

Furthermore, precipitation has a significant impact on the regional ecology and feasibility of ecotourism initiatives. Adequate precipitation is essential for vegetation growth and the development of unique ecological landscapes. Soil conservation directly affects a region’s environmental carrying capacity, which is crucial for sustainable ecotourism. Improved conservation capacity minimizes the negative effects of tourism on the environment. Additionally, factors such as relief height and forest age play significant roles, accounting for 10.82% and 9.11%, respectively. Increased relief height contributes to the formation of unique geological features, supporting the development of ecotourism resources with exceptional scenic value. Meanwhile, a higher forest age indicates greater potential for ecotourism resources focused on tall vegetation, offering visitors the opportunity to immerse themselves in nature and explore plant diversity.

The aforementioned findings are substantiated by evidence drawn from existing research [10,13,15,30,31,32,33,34,35,36] and collectively underpin the theoretical framework of the present study. Compared to previous findings, the rankings of the feature importance derived from the machine learning techniques are fundamentally distinct because they are based on the inherent attributes of the analyzed sample data, without relying on human intervention to predetermine factor weights. In contrast, the majority of existing studies employ methodologies such as the expert scoring approach or AHP [15] to assign factor weights prospectively, which culminates in a composite analysis to produce the final ecotourism suitability map. Consequently, the assignment of feature weights in traditional analytical frameworks is susceptible to researcher subjectivity. In contrast, machine learning methodologies, in which are based on data-driven insights, provide a more objective and impartial assessment.

5. Discussion

5.1. Model’s Fractional Anisotropy

Although data-driven machine learning can provide accurate predictions, it requires a substantial number of high-quality samples to perform effectively. Different training samples can lead to variations in models, emphasizing the importance of reliable training data to minimize such differences and improve the model’s strength and applicability. In the present study, we used the training samples from the Professional Surveying and Mapping Department of the Third Surveying and Mapping Institute of Hunan Province to ensure data reliability and reduce model divergence caused by inaccurate training samples. It is crucial to note that a model’s performance is significantly influenced by its parameters, which can also contribute to model divergence. To address these concerns and improve the model’s reliability, we controlled the parameters while maintaining the quality of the training samples. In our pursuit of higher accuracy, we utilized learning curves, grid search, cross-validation, and regularization techniques to prevent overfitting and improve the model’s applicability.

5.2. Impact of Land Policy

The development of ecotourism is closely related to local land policies that regulate it. These policies form a connected system that influences ecotourism. While restrictive land policies may hinder ecotourism to some extent, this is not inherently contradictory because ecotourism is based on environmental conservation. Different countries have different perspectives on land development, leading to variations in land policy formulations. Therefore, assessing the suitability of ecotourism must be based on an understanding of national and local land policy frameworks.

The evaluation’s findings will have a significant impact on the development of ecotourism policies, directly influencing land use planning. This can be observed through several factors. First, the results of the evaluation will guide the formulation of policies by identifying areas suitable for ecotourism and areas that require special protection, leading to the implementation of more precise and effective land policies such as land use restrictions, protection measures, and ecological compensation mechanisms. For example, at the turn of the century, the local jurisdiction implemented a land management strategy involving the conversion of farmlands back into forested areas. This initiative, known as the ‘returning farmland to forest’ policy, achieved two key objectives. It facilitated the rehabilitation of a degraded ecological environment and instituted an ecological compensation mechanism designed to provide financial support to local farmers, thus fostering a more sustainable relationship between human activities and the environment. Second, the evaluation results can be utilized to improve land use planning, especially in ecologically sensitive areas. By considering the needs of ecotourism activities, land use can be organized in a way that allows tourism to coexist harmoniously with the natural environment. Additionally, targeted land and soil remediation plans should be developed to address illegal land use, thereby restoring and enhancing land quality and productivity. Furthermore, the evaluation contributes to sustainable regional development by promoting scientific planning, the rational use of land resources, and the healthy development of ecotourism, leading to the coordinated development of regional environmental protection and economic growth. For instance, the locale’s extant land use policy, characterized by a dual emphasis on conservation and sustainable development, aims to balance the stringent protection of the ecological environment with the thoughtful expansion of eco-tourism activities. This approach not only preserves the natural habitat but also improves the economic wellbeing of the local population through the promotion of environmentally friendly tourism initiatives. Lastly, the evaluation results can enhance the efficiency of land policy implementation by clarifying policy objectives and standards, thus reducing subjectivity and arbitrariness in the policy implementation process [4]. This provides a basis for monitoring and evaluating policy implementation to ensure that policies are effectively implemented to achieve the desired outcomes. For example, the China Spatial Planning Observation Network (CSPON) provides real-time monitoring and early warning systems for the implementation of both national and regional spatial plans. This advancement significantly enhances the scientific rigor and effectiveness of spatial planning and management across the country.

5.3. Applied Range and Limitations

This study used machine learning and data mining to objectively evaluate the suitability of regional ecotourism. We used accurate and scientific sample data, which resulted in excellent prediction outcomes. Our approach effectively overcomes the limitations of traditional subjective methods and allows for independent research on ecotourism suitability within the study area. However, the effectiveness of our method relies on having a sufficient number of labeled samples, as machine learning models require a certain number of training samples to function optimally. Therefore, the practical application of our method may be limited by the availability of labeled data.

5.4. Future Work

Our experiment confirmed that using machine learning to evaluate ecotourism suitability can effectively identify and map potential ecotourism areas. This method provides more objective evaluation results because it does not rely on the manual determination of factor weights. However, it is crucial to obtain effective training samples with labels for this approach to be effective. In cases where training samples in the study area are insufficient or lacking, the method may have limitations at the algorithmic level. In such situations, future research should explore incorporating transfer learning to reuse models or samples from different domains for a more comprehensive assessment of ecotourism suitability.

Furthermore, a global-scale evaluation of ecotourism suitability remains to be conducted due to variations in the sample data and regional characteristics. In the future, it will be necessary to break down national and regional barriers and integrate land policies, regional cultures, and unique tourism resources for a more comprehensive evaluation.

6. Conclusions

This empirical study conducted in Zhangjiajie, China, integrated multisource geospatial data and machine learning techniques to evaluate the suitability of ecotourism in the region. Our findings demonstrated the effectiveness of using machine learning methods for regional ecotourism suitability evaluation, as they can achieve objective evaluation without human interference in determining factor weights. Furthermore, this study demonstrated the inherent limitations of the traditional method of human interference, which can be addressed by the proposed data-driven approach.

The AUC of the four models (XGBoost, AdaBoost, RF, and SVM) for the training data were 0.9884, 0.9660, 0.9648, and 0.9567, respectively. XGBoost performed the best, followed by AdaBoost and RF, whereas SVM performed the worst. To avoid relying on a single model, the prediction results were averaged. In the combined prediction results, highly suitable, moderately suitable, marginally suitable, and unsuitable areas for ecotourism accounted for 19.34%, 28.78%, 23.87%, and 28.01% of the total area, respectively.

The cartographic representation of the predicted outcomes delineates the spatial configuration of preexisting ecotourism destinations and accurately identifies prospective ecotourism resources. Though currently unexploited as ecotourism attractions, these locations showcase exceptional scenic and landscape appeal and hold significant developmental potential in the future. They frequently face constraints imposed by transportation infrastructure, topography, tourism-related amenities, and policies yet to be formulated; however, this presents an opportunity rather than a hindrance. Such conditions allow for enhanced ecological preservation—a fundamental requisite for the sustainability of ecotourism. Moreover, certain areas previously considered suitable for development have since been deemed inappropriate due to their ecological sensitivity and vulnerability. Consequently, the government has instituted a comprehensive suite of land use policies—including, but not limited to, ecological protection red lines, safeguards for drinking water sources, the delineation of permanent basic farmland, and restrictions on urban expansion—to conserve these regions. In the long term, these measures stand to confer immense benefits.

Future research should explore training samples and model selection. In regions with insufficient training samples or zero samples, transfer learning is a viable option as it enables model transfer and reuse even in the absence of training samples or zero samples in the target domain, thereby reducing the reliance on training samples in traditional machine learning methods.