Tourism Resource Evaluation Integrating FNN and AHP-FCE: A Case Study of Guilin

Abstract

With the rapid development of the tourism industry, scientific evaluation of tourism resources is crucial to realize sustainable development. Especially how to quantify resource advantages in international tourism cities has become an important basis for tourism planning and policy making. However, the limitations of traditional evaluation methods in the allocation of indicator weights and nonlinear data processing make it difficult to meet the development needs of international tourism cities. Therefore, this study takes Guilin, an international tourist city, as the research object and proposes a hybrid framework integrating fuzzy neural network (FNN) and analytic hierarchy process-fuzzy comprehensive evaluation (AHP-FCE). Based on 800 questionnaire data covering tourists, practitioners, and local residents, the study constructed a multilevel evaluation system (containing 12 specific indexes in the three dimensions of nature, service, and culture) using the Delphi method of expert interviews. It is found that AHP-FCE can effectively analyze the hierarchical relationship of evaluation indexes, but it is easily affected by the subjective judgment of experts. In contrast, FNN can effectively improve evaluation accuracy through the adaptive learning mechanism, and it especially shows significant advantages in dealing with tourists’ perception data. The empirical analysis shows that Guilin has obvious room for improvement in “environmental friendliness” and “cultural communication effectiveness”. The integration framework proposed in this study aims to enhance the scientific validity and accuracy of the assessment results, and provides reference and inspiration for the sustainable development of Guilin international tourism destination.

1. Introduction

Evaluation of tourism resources constitutes a core research topic in tourism geography and resource management, with its theoretical framework having evolved from singular economic value assessment toward multidimensional sustainability evaluation [1]. Early studies predominantly employed qualitative descriptions, exemplified by Gunn’s (1972) “core-periphery” theory of tourist attractions [2], which emphasized the spatial heterogeneity of resource endowments. With the application of systems theory, scholars began developing comprehensive evaluation frameworks: Butler’s (1980) tourism area life cycle theory integrated resource development with market [3], while Ceballos-Lascurain (1996) advocated for incorporating environmental carrying capacity indicators in ecotourism assessments [4]. Since the 21st century, UNESCO’s evaluation criteria for the “Outstanding Universal Value (OUV)” of World Heritage Sites have further advanced paradigm innovation in evaluation through dual cultural–ecological perspectives [5]. Nevertheless, traditional evaluation methods still exhibit significant limitations in multisource data integration [6], highlighting an urgent need for methodological breakthroughs.

The evaluation of tourism resources has evolved from qualitative analysis to quantitative modeling [7]. The analytic hierarchy process (AHP) and fuzzy comprehensive evaluation (FCE) have been widely adopted due to their structured decision-making advantages. Proposed by Saaty, AHP effectively addresses weight allocation in multi-criteria decision-making by constructing hierarchical models and performing consistency checks [8]. For instance, Bo Huang et al. employed the analytic hierarchy process (AHP) and fuzzy comprehensive evaluation (FCE) to formulate the University Historical Building Protection Evaluation Framework (UHBPEF) [9]. Subsequently, the introduction of fuzzy mathematics further addressed uncertainties in the evaluation process. Zadeh’s fuzzy set theory was employed to handle the ambiguous boundaries of tourism resource attributes [10], while Deng Binhua developed a fuzzy comprehensive evaluation model to assess tourism attractiveness in Nyingchi City [11]. However, traditional AHP-FCE methods still suffer from two major limitations: on the one hand, the weight matrix relies heavily on expert judgment, making it difficult to adapt to changes in indicator importance under complex environments; on the other hand, the rigid design of membership functions fails to capture nonlinear relationships in multi-source heterogeneous data. To address these issues, scholars have explored machine learning techniques for optimization, such as support vector machine (SVM) regression for short-term tourist flow prediction and random forest (RF) for risk assessment in water supply networks [12,13]. Nevertheless, these models still exhibit deficiencies in interpretability and fuzzy rule extraction.

In this context, the research motivation of this paper is to construct a tourism resource evaluation model that simultaneously possesses structured modeling, fuzzy processing, and adaptive learning capabilities, in order to enhance the scientificity, flexibility, and interpretability of tourism resource assessment. Specifically, this paper hopes to fill the following research gaps. First, to break through the technical bottleneck of the traditional AHP-FCE method in terms of subjective weight setting and affiliation function curing. Second, to introduce trainable models to fuse unstructured and perceptual data. To this end, this study proposes a comprehensive evaluation framework for tourism resources that integrates FNN and AHP-FCE models, with the goal of establishing a set of tourism resource evaluation systems that can adapt to multi-source data inputs, dynamically adjust weights, and mine fuzzy rules. In addition, this study focuses on the following core issues. First, how to go about constructing a scientific, multilevel tourism resource evaluation index system that covers the dimensions of nature, service, and culture? Second, how to design a tourism resource evaluation framework that integrates AHP-FCE and FNN to enhance the scientificity and accuracy of tourism resource assessment?

This study makes three primary contributions: First, we propose a novel hybrid model integrating AHP-FCE with fuzzy neural networks (FNNs), which synergistically combines the hierarchical analysis of AHP, fuzzy evaluation of FCE, and dynamic optimization capabilities of FNN. This integration achieves dual accommodation of both subjective and objective data. Compared with conventional AHP-FCE methods, our model employs FNN’s self-learning mechanism to refine the AHP weight matrix while utilizing membership functions to quantitatively enhance indicator weighting—significantly improving the scientific rigor of evaluation outcomes. This advancement provides more reliable scientific references for developing Guilin as an international tourism destination.

Second, focusing on tourism resource potential evaluation in Guilin, we developed a scientifically robust assessment index system. The FNN-based iterative optimization of AHP-FCE weights and fuzzy rules addresses critical limitations in traditional models, particularly the rigidity of fixed weights and crude handling of fuzzy boundaries, thereby substantially enhancing both the valuation and predictive capacity for tourism resources.

Finally, in comparison with AHP-BP models—where BP neural networks require precise numerical inputs and additional fuzzification preprocessing—our FNN-based approach directly handles fuzzy indicators in tourism evaluation through built-in membership functions and fuzzy rule bases. This effectively resolves uncertainties in tourism resource assessment. Comparative experiments demonstrate that the FNN-AHP-FCE model outperforms AHP-BP with the following advantages: reduced mean absolute error (MAE) and root mean square error (RMSE), along with a 1.30% higher prediction accuracy. These findings strongly suggest that the FNN-AHP-FCE integration offers greater applicability for tourism resource evaluation tasks characterized by inherent data ambiguities.

The remainder of this paper is organized as follows: Section 2 reviews the research status and summarizes previous findings; Section 3 describes the data sources and methodology employed in the study; Section 4 presents the evaluation indicators and framework for tourism resources; Section 5 provides the results analysis and policy recommendations; Section 6 concludes with research findings and prospects for future studies in this field.

2. Literature Review

Guilin occupies a strategically significant position in China’s national initiative to develop world-class tourist cities, benefiting from its unique endowment of natural landscapes and cultural resources, as well as its advantageous geographical location and transportation accessibility for international tourism [14]. This gives the evaluation of Guilin’s tourism resources both special theoretical and practical significance. Early research primarily focused on the design of tourism landscapes, analyzing Guilin’s scenic resources from the perspective of rural tourism landscape design. These studies proposed scientific planning solutions and design concepts through normative recommendations for rural tourism landscape development in Guilin [15]. In recent years, scholars have shifted attention toward balancing development and conservation: Tan Yingying et al. employed equivalent modification methods and grid analysis to examine spatiotemporal changes in ecosystem service values across Guilin’s six urban districts [16], while Huang Yanling et al. applied grounded theory to collect and code web data from Guangxi’s “Chinese Ethnic Minority Villages,” constructing a conceptual model for transforming ethnic cultural resources into cultural capital [17]. Xu Yan adopted a combined qualitative–quantitative approach to evaluate sports tourism resources, establishing an evaluation system for Guilin using the analytic hierarchy process (AHP) to promote the development of sports cultural tourism [18]. However, existing evaluation models predominantly rely on subjectively predefined indicator systems, where weight allocation and parameter settings are often constrained by expert judgment or historical data biases, making them inadequate for accommodating unexpected disturbances from complex environmental variables.

At the methodological level, existing research predominantly relies on conventional AHP-FCE approaches. For instance, Tao Chen et al. applied the analytic hierarchy process (AHP) and the fuzzy comprehensive evaluation (FCE) method to comprehensively assess the sustainable development status of the Qianfeng Community; yet their weight matrix adjustments required manual intervention [19]. Similarly, Zhang Haibing et al. constructed an evaluation system for soundscapes in rural tourism areas using AHP, but the weight allocation remained subjective and lacked data-driven calibration [20]. Limited attempts to incorporate intelligent algorithms exhibit contextual constraints: Peng Ying applied a TCN-LSTM hybrid model for daily tourist flow prediction in scenic areas [21], but failed to integrate it with resource evaluation models. Tian Huimin et al. proposed a fidelity assessment method for watermarked vector maps combining fuzzy neural networks with AHP, enhancing evaluation accuracy and objectivity through optimized subjective–objective weights via fuzzy comprehensive evaluation and BP neural networks [22]. However, the absence of scientifically justified thresholds may compromise the result’s interpretability. Lin S et al. developed an AHP-BP-based model to assess agritourism-integrated rural environments under the “dual-carbon” policy in Zhejiang [23], effectively combining expert judgment with neural network prediction. However, their study did not incorporate fuzzy comprehensive evaluation, limiting its ability to handle uncertainty in expert input. The summarizing table of technical methods is shown in

Table 1. The summarizing table of technical methods.

Author(s)	Year	Research Focus	Methodology	Key Findings/Limitations
Tao Chen et al. [19]	2023	Sustainable development evaluation of Qianfeng Community	AHP + FCE	Enabled comprehensive evaluation, but weight adjustment required manual intervention, reducing objectivity.
Zhang Haibing et al. [20]	2021	Soundscape evaluation in rural tourism	AHP	Constructed evaluation system, but weight allocation was subjective and lacked data-driven refinement.
Peng Ying [21]	2022	Tourist flow prediction in scenic areas	TCN + LSTM	Demonstrated forecasting ability, yet failed to integrate predictions with tourism resource evaluation frameworks.
Tian Huimin et al. [22]	2024	Fidelity evaluation of watermarked vector maps	AHP + FCE + BP	Improved objectivity and accuracy, but lacked the ability to model multi-source perceptual information and data.
Lin S et al. [23]	2025	Assess agritourism-integrated rural environments	AHP + BP	Combines expert judgment with neural networks, but struggles with uncertain expert input.

. These limitations highlight that developing a tourism resource evaluation model with scenario adaptability, multi-source data integration capacity, and interpretable fuzzy rules constitutes the key scientific challenge in advancing evaluation methodologies.

Fuzzy neural networks (FNNs), as an interdisciplinary product of connectionism and fuzzy logic, offer a novel solution to address the static limitations of conventional methods. By simulating the human brain’s fuzzy reasoning mechanisms, FNN integrates the self-learning capability of neural networks with the semantic interpretability of fuzzy systems [24]. In research applications, FNN has primarily been employed for time-series forecasting and complex system modeling—for instance, predicting tourist volumes in scenic areas [25]; Yang Qijun et al. used a T-S fuzzy neural network to construct a model for predictive evaluation of the ecological compensation effect of land consolidation [26]. Notably, studies on the integration of AHP-FCE remain exploratory [27]. Existing attempts predominantly focus on weight optimization, exemplified by Yang Xueping’s use of the analytic hierarchy process (AHP) to develop an evaluation index system for theme park visitor experiences, calculating indicator weights and rankings, followed by constructing an importance-performance analysis (IPA) model for Splendid China’s visitor satisfaction using survey data [28]. However, research targeting complex geomorphic regions remains scarce, where the interplay between ecological fragility and cultural landscapes imposes unique adaptability requirements on evaluation models [29].

The advancement of multi-source data fusion technologies has driven innovation in tourism resource evaluation methodologies. Geographic information systems (GISs), with their spatial visualization and quantitative analysis capabilities, play a pivotal role in studying resource spatial heterogeneity [30]. For instance, Li Xue et al. developed a mathematical model for optimizing tourism routes based on spatial heterogeneity of attractions, incorporating attraction appeal and visitor preference indices as objective functions, and proposed an improved ant colony algorithm for solution [31]. Meanwhile, multi-modal social media data processing techniques have expanded the dimensions of tourist behavior analysis, as demonstrated by Manju V who enhanced BERT-based semantic intensity using a guided LDA model for aspect term extraction in sentiment analysis [32]. However, current technical frameworks predominantly rely on independent modeling of single-modal data, lacking mechanisms for deep integration of multi-source heterogeneous data [33]. To address these limitations, this study adopts an integrated AHP-FCE and FNN approach. The conventional AHP-FCE method provides a structured framework for tourism resource evaluation, while the FNN component effectively compensates for the inherent shortcomings of traditional methods [34], particularly in handling data complexity and adaptability.

This section conducts a systematic review of existing research on tourism resource evaluation in Guilin, revealing two predominant limitations: insufficient evaluation index systems and excessive subjectivity inherent in conventional AHP-FCE methodologies. While acknowledging the demonstrated efficacy of fuzzy neural networks (FNNs) in temporal prediction and system modeling applications, current integrative approaches with AHP-FCE remain fundamentally superficial. To bridge these critical research gaps, we propose an innovative, deeply integrated AHP-FCE-FNN evaluation framework. The framework not only addresses the “shallow integration” problem prevalent in existing hybrid models but also establishes a new paradigm for complex tourism resource evaluation.

3. Data Sources and Research Methods

3.1. Data Sources

3.1.1. Overview of the Research Area

Situated in southern China’s Guangxi Zhuang Autonomous Region (abbreviated as “Gui”, historically known as Guizhou, Jingjiang, and Shi’an), Guilin is a prefecture-level city located in northeastern Guangxi, renowned for harboring the world’s most typical karst landscape system. Officially designated by China’s State Council as an internationally renowned scenic city, a national pilot zone for tourism innovation, and an international tourism transportation hub, Guilin serves as a strategic nexus between the Pan-Pearl River Delta Economic Zone and ASEAN Free Trade Area. The city’s iconic attractions—including the Yulong River in Yangshuo, Longji Terraced Fields in Longsheng, and Lingqu Canal in Xing’an —exemplify the harmonious integration of spectacular natural features (peak clusters, karst caves, terraces, subterranean rivers, and hot springs) with rich ethnic cultural resources (Zhuang, Yao, and Miao minorities). With 89 nationally A-grade or above tourist attractions (including four 5A-grade sites), Guilin ranks among China’s prefecture-level cities with the highest density of premium tourism resources, thereby providing diversified assets for tourism development.

3.1.2. Data Origins

From June to December 2024, researchers conducted randomized questionnaire surveys across major tourist attractions in Guilin (see Appendix A), targeting visitors, scenic area staff, and local residents. The survey evaluated satisfaction levels regarding landscape quality, service provision, and cultural experiences. Primary data collection sites included the Yulong River, Longji Rice Terraces, Lingqu Canal, and adjacent leisure zones. A total of 848 questionnaires were distributed, with 807 retrieved (95.16% retrieval rate), after eliminating invalid questionnaires such as incorrect, missing, and arbitrary filling, a total of 800 valid questionnaires were obtained. This study used 800 valid questionnaires as the sample population, which met the requirements of the AHP-FCE evaluation model for questionnaire quality. Analysis of valid questionnaires revealed distinct tourist preference distributions: cultural experience-oriented visitors accounted for 27.125%, landscape sightseeing-focused tourists represented 38.5%, and leisure recreation-preferring travelers comprised 34.375%. The predominance of landscape-oriented tourists (38.5%) demonstrates that natural scenery constitutes the primary attraction for visitors to Guilin, which aligns precisely with the destination’s iconic positioning as possessing “the most spectacular mountains and waters under heaven.” Subsequently, the significant proportion of leisure-seeking tourists (34.375%) reflects strong market demand for comfort facilities and recreational activities, indicating that Guilin’s tourism service infrastructure has achieved considerable competitiveness. Notably, although cultural-experience tourists represent the smallest segment (27.125%), this still exceeds one-quarter of respondents, revealing substantial development potential in cultural resources (particularly ethnic minority customs and historical heritage). However, this comparatively lower percentage suggests possible deficiencies in current cultural product design or marketing strategies that may require optimization to fully realize this potential.

3.2. Research Methods

This study proposes an innovative tourism resource evaluation methodology that effectively integrates fuzzy neural networks (FNN) with the analytic hierarchy process-fuzzy comprehensive evaluation (AHP-FCE) framework, representing a significant advancement beyond conventional evaluation systems through its multi-level modeling architecture and adaptive optimization mechanism. The methodological framework consists of three key components: first, a rigorously validated three-tier hierarchical index system (objective-criteria-scheme) constructed via AHP with expert-weighted indicators that have undergone comprehensive consistency verification (CR < 0.1); second, a sophisticated fuzzy processing stage employing FCE with customized membership functions to quantitatively characterize qualitative assessment boundaries; and third, a novel FNN optimization module that dynamically adjusts weight allocations through machine learning of membership rules, thereby transforming the traditional static model into an adaptive, data-driven system. The proposed hybrid approach uniquely combines fuzzy logic’s inherent tolerance for uncertainty with neural networks’ powerful nonlinear learning capabilities. While maintaining the systematic advantages of AHP-FCE, the incorporation of FNN’s parameter optimization significantly enhances both the objectivity and timeliness of evaluation results. Consequently, this methodology provides a robust analytical tool for multidimensional tourism resource assessment that successfully bridges theoretical sophistication with practical applicability.

The main contributions of this section are to systematically reveal the structural dimensions of Guilin’s tourism attractiveness through comprehensive data collection and multidimensional analysis, as well as to innovatively propose an integrated AHP-FCE and fuzzy neural network (FNN) hybrid methodology.

4. Tourism Resources Evaluation

In this section, we present the procedural steps of the adopted methodologies (AHP-FCE and FNN), as illustrated in Figure 1. Additionally, key components are elaborated, including computational formulas and flowcharts of the algorithmic processes. The primary contribution of this work lies in the proposal of a novel evaluation framework integrating FNN with AHP-FCE to achieve a comprehensive assessment of tourism resources. To the best of our knowledge, this represents the first application of such a hybrid approach in this domain, establishing a new indicator evaluation system by leveraging their complementary strengths. Specifically, in contrast to conventional AHP-FCE—where AHP and FCE are separately employed to determine indicator weights and construct the evaluation matrix—our method further refines the weight assignments through FNN while enhancing the processing of fuzzified data.

4.1. Establishment of Evaluation Index System

4.1.1. Preliminary Construction of Evaluation Indicators

To comprehensively assess the multidimensional value of tourism resources in Guilin, this study establishes an evaluation system structured around three core dimensions: natural landscape resources, management and service resources, and community cultural resources. Natural landscape resources, serving as the primary attraction of Guilin’s tourism, are evaluated based on natural landscape quality and scenic area planning rationality. The corresponding indicators are constructed with reference to the UNESCO World Natural Heritage evaluation criteria. Management and service resources integrate tourism infrastructure and public service systems. The indicators for this dimension are derived from established tourism city standards and relevant literature. Community cultural resources emphasize the dynamic preservation of ethnic culture and innovation in cultural-tourism integration. The indicators for this dimension are formulated based on cultural tourism research and related studies. Through this framework, a preliminary evaluation system comprising 3 criteria and 14 indicators has been developed.

4.1.2. Expert Consultation

To further refine the evaluation indicators, this study employed a rigorous two-round Delphi method involving a carefully selected panel of 25 experts (60% university scholars and 40% technical professionals) with extensive experience in tourism planning and ecological conservation [35]. The anonymous questionnaires were distributed to the selected experts for rating each indicator, followed by systematic analysis of the returned questionnaires to verify compliance with Delphi method requirements. Through two rounds of Delphi consultations, the final indicators were systematically refined. In the initial round, the indicators “ecological carrying capacity dynamic monitoring” and “cultural creative product diversity” exhibited coefficient of variation (CV) values exceeding 0.30 and coordination coefficients below 0.50, necessitating a second expert consultation. Specifically, some experts noted that “ecological carrying capacity dynamic monitoring” was defined too broadly without clear monitoring targets, while others observed that “cultural creative product diversity” overemphasized quantitative aspects while neglecting qualitative dimensions, as mechanically reproduced souvenirs might diminish cultural value connotations. After consolidating redundant indicators and clarifying definitions, a revised questionnaire was circulated for the second consultation. The results demonstrated significant improvement, with all indicators in the second round showing CV values below 0.25 and coordination coefficients above 0.50. The enhanced consensus among experts and strong alignment between the refined indicators and predetermined objectives indicated that the evaluation system had achieved sufficient convergence, eliminating the need for a third consultation. This rigorous two-round Delphi process ultimately yielded a comprehensive tourism resource evaluation framework comprising 3 criterion layers and 12 indicators (

Table 2. Evaluation index system for tourism resources in Guilin.

$\mathbf{Target} \mathbf{Layer} \mathit{A}$	$\mathrm{Criterion} \mathrm{Layer} \mathit{B}$	$\mathrm{Indicator} \mathrm{Layer} \mathit{C}$
Evaluation Index System for Tourism Resources in Guilin City ($A$)	natural landscape resources ($B_1$)	natural landscape quality ($C_1$)
		visual experience optimization ($C_2$)
		rationality of scenic area planning ($C_3$)
		science popularization tour service ($C_4$)
	management service resources ($B_2$)	recreation facility completeness ($C_5$)
		service comfort level ($C_6$)
		transportation accessibility ($C_7$)
		environmental friendliness ($C_8$)
	community cultural resources ($B_3$)	cultural heritage depth ($C_9$)
		experiential activity diversity ($C_{10}$)
		cultural communication effectiveness ($C_{11}$)
		community participation level ($C_{12}$)

).

4.2. Construction of Judgment Matrix and Calculation of Weights

Four judgment matrices were established, $B_1$–$C_n$ (n = 1, 2, 3, 4), $B_2$–$C_n$ (n = 5, 6, 7, 8), $B_3$–$C_n$ (n = 9, 10, 11, 12) and $A$–$B_n$ (n = 1, 2, 3).

Based on the evaluation indicator system, expert scoring was conducted to perform pairwise comparisons of the indicators’ relative importance. For instance, if factors A and C are considered equally important, they are assigned a value of 1; if factor A is x times more important than factor C, the reciprocal value 1/x is assigned to the inverse comparison. As a specific example, when factor A is moderately more important than factor C, it receives a score of 3, while the inverse comparison (C versus A) is assigned a value of 1/3. The complete scoring criteria are detailed in

Table 3. Scale and corresponding interpretation.

Scale	Corresponding Interpretation
1	The two factors are equally important
3	One factor is moderately more important than the other
5	One factor is strongly more important than the other
7	One factor is very strongly more important than the other
9	One factor is extremely more important than the other
2, 4, 6, 8	Intermediate values between adjacent judgments
$\mathrm{If} \mathrm{the} \mathrm{scale} \mathrm{value} \mathrm{of} \mathrm{factor} i$$\mathrm{relative} \mathrm{to} \mathrm{factor} j$$\mathrm{equals} a_{ij}$$, \mathrm{then} \mathrm{the} \mathrm{scale} \mathrm{value} \mathrm{of} \mathrm{factor} j$$\mathrm{relative} \mathrm{to} \mathrm{factor} i$$\mathrm{equals} 1/a_{ij}$

.

Based on the judgment matrix derived from pairwise comparisons of factors, the weights of individual indicators can be computed and subsequently normalized. The specific computational procedure is as follows:

• First step; column-wise normalization: each element in the matrix is divided by the sum of its corresponding column.
• Second step; row summation: the normalized columns are aggregated by summing each row.
• Third step; weight vector generation: each element in the resultant row-sum vector is divided by n (the matrix dimension) to obtain the final weight values.

Assuming the judgment matrix is as follows:

$$A=\left[\begin{array}{ccc}{a}_{11}& \dots & a_{1n}\\ ⋮& \ddots & ⋮\\ a_{n1}& \dots & a_{nn}\end{array}\right].$$

(1)

Since the obtained weight vector is as follows:

$$w_i={\displaystyle \frac{1}{n}}{\sum }_{j=1}^n{\displaystyle \frac{a_{ij}}{{\sum }_{k=1}^n{a}_{kj}}}(i=1,2,3,\dots ,n)$$

(2)

To ensure the validity of the analytic hierarchy process (AHP) outcomes, consistency verification must be rigorously performed on the judgment matrix, as inconsistency levels may vary under different evaluation conditions. The calculation formula is as follows:

$${\lambda }_{max}={\displaystyle \frac{1}{n}}\sum _{i=1}^n{\displaystyle \frac{{\left(AW\right)}_i}{W_i}},$$

(3)

$$CI={\displaystyle \frac{{\lambda }_{max}-n}{n-1}},$$

(4)

$$CR={\displaystyle \frac{CI}{RI}},$$

(5)

In the consistency verification process, ${\lambda }_{max}$ represents the maximum eigenvalue of the judgment matrix, where $n$ denotes its dimension (order), ${\left(AW\right)}_i$ corresponds to the $i$-th element obtained from the matrix multiplication of the judgment matrix and weight vector, $CI$ denotes the consistency index, and $RI$ refers to the average random consistency index for varying matrix dimensions (as presented in

Table 4. Average random consistency index.

$\mathit{n}$	1	2	3	4	5	6	7	8	9	10
$\mathit{R}\mathit{I}$	0	0	0.52	0.89	1.12	1.26	1.36	1.41	1.46	1.49

). Following established AHP conventions, the judgment matrix is considered to exhibit satisfactory consistency when the calculated consistency ratio ($CR$) is less than 0.10, thereby validating the reliability of the derived weights for subsequent analysis.

4.3. Establishment of Fuzzy Comprehensive Evaluation Model

This study adopts the fuzzy statistical method combined with the analytic hierarchy process to construct the fuzzy comprehensive evaluation model of tourism resources.

First, we define the fuzzy comprehensive evaluation set $D_j$ (j = 1,2,3,4) with four satisfaction levels: very satisfied (score = 100), relatively satisfied (80), neutral (60), and very dissatisfied (40).

The second step involves conducting single-factor fuzzy evaluations to establish a comprehensive fuzzy judgment matrix, where public assessments of individual factors in Guilin’s tourist attractions are systematically analyzed. For each evaluation index, the public’s scoring opinion is collected through a questionnaire survey, and the proportion of each grade is taken as the index’s affiliation on the fuzzy hierarchy (which is actually the frequency-type affiliation), and the affiliation degree of Guilin’s tourism resources evaluation index is obtained (see

Table 5. Affiliation degree of evaluation indicators for tourism resources in Guilin City.

Indicator	Very Satisfied	Relatively Satisfied	Neutral	Very Dissatisfied
natural landscape quality	0.36	0.2	0.329	0.111
visual experience optimization	0.295	0.233	0.35	0.122
rationality of scenic area planning	0.128	0.334	0.397	0.141
science popularization tour service	0.188	0.27	0.4	0.142
recreation facility completeness	0.209	0.245	0.411	0.135
service comfort level	0.159	0.296	0.409	0.136
transportation accessibility	0.088	0.236	0.502	0.174
environmental friendliness	0.25	0.236	0.361	0.153
cultural heritage depth	0.22	0.236	0.369	0.175
experiential activity diversity	0.204	0.216	0.42	0.16
cultural communication effectiveness	0.134	0.266	0.43	0.17
community participation level	0.134	0.259	0.47	0.137

). Taking “natural landscape quality” as a representative example, analysis of collected valid questionnaires revealed the following satisfaction distribution among respondents: 36% expressed being “very satisfied”, 20% reported being “relatively satisfied”, while 32.9% and 11.1% indicated “neutral” and “very dissatisfied” evaluations, respectively. Consequently, the fuzzy vector of the indicator is =(0.36, 0.2, 0.329, 0.111). The fuzzy judgment matrix $\mathrm{R}$ for all indicators can be constructed in this way.

$$\begin{gathered} R=\left[\begin{array}{cccc}{r}_{11}& r_{12}& \dots & r_{1n}\\ r_{21}& r_{22}& \dots & r_{2n}\\ ⋮& ⋮& \ddots & ⋮\\ r_{m1}& r_{m2}& \dots & r_{mn}\end{array}\right], \\ {\mathrm{r}}_{\mathrm{ij}}={\displaystyle \frac{\mathrm{Number} \mathrm{of} \mathrm{respondents} \mathrm{rating} \mathrm{the} \mathrm{i}\text{-}\mathrm{th} \mathrm{evaluation} \mathrm{indicator} \mathrm{as} \mathrm{Dj}}{\mathrm{The} \mathrm{total} \mathrm{number} \mathrm{of} \mathrm{participants} \mathrm{in} \mathrm{the} \mathrm{evaluation}}} \end{gathered}$$

(6)

In the third step, the evaluation model is established by determining the single-factor evaluation matrix R and incorporating the factor weights obtained through the analytic hierarchy process (AHP), followed by performing a fuzzy transformation operation to derive the final weight vector V, which comprehensively integrates both the qualitative expert judgments from AHP and quantitative public evaluation data through rigorous mathematical processing to ensure scientifically sound and operationally reliable assessment results for tourism resource evaluation.

$$V=WR,$$

(7)

The fourth step involves determining indicator weights using the Delphi method; a structured forecasting technique originally developed by the RAND Corporation in the 1960s that systematically aggregates expert knowledge and subjective judgments [36,37]. In this study; 25 domain experts independently completed questionnaires through iterative rounds of anonymous feedback until consensus was achieved. Following two deliberation rounds;

Table 6. The impact of the criterion layer on the target layer.

	${\mathit{B}}_1$	${\mathit{B}}_2$	${\mathit{B}}_3$	$\mathit{w}$
${\mathit{B}}_1$	1	3	5	0.6334
${\mathit{B}}_2$	1/3	1	3	0.2605
${\mathit{B}}_3$	1/5	1/3	1	0.1061
$CR$ = 0.0897 < 0.10, the results passed the consistency verification.

presents the resulting judgment matrix of relative importance among the three criterion layers; using the same methodology; the weight values for each individual indicator were determined (see

Table 7. The impact of the indicator layer on the criterion layer.

	${\mathit{C}}_1$	${\mathit{C}}_2$	${\mathit{C}}_3$	${\mathit{C}}_4$	$\mathit{w}$
${\mathit{C}}_1$	1	4	6	3	0.512
${\mathit{C}}_2$	1/4	1	2	1/2	0.198
${\mathit{C}}_3$	1/6	1/2	1	1/3	0.114
${\mathit{C}}_4$	1/3	2	3	1	0.176
$CR$ = 0.087 < 0.10, the results passed the consistency verification.
	${\mathit{C}}_5$	${\mathit{C}}_6$	${\mathit{C}}_7$	${\mathit{C}}_8$	$\mathit{w}$
${\mathit{C}}_5$	1	2	3	5	0.5225
${\mathit{C}}_6$	1/2	1	2	3	0.2919
${\mathit{C}}_7$	1/3	1/2	1	2	0.1709
${\mathit{C}}_8$	1/5	1/3	1/2	1	0.0961
$CR$ = 0.094 < 0.10, the results passed the consistency verification.
	${\mathit{C}}_9$	${\mathit{C}}_{10}$	${\mathit{C}}_{11}$	${\mathit{C}}_{12}$	$\mathit{w}$
${\mathit{C}}_9$	1	2	4	5	0.452
${\mathit{C}}_{10}$	1/2	1	3	4	0.312
${\mathit{C}}_{11}$	1/4	1/3	1	2	0.163
${\mathit{C}}_{12}$	1/5	1/4	1/2	1	0.073
$\mathit{C}\mathit{R}$ = 0.085 < 0.10, the results passed the consistency verification.

).

4.4. Establishment of Fuzzy Neural Network Model

In tourism resource evaluation, the fuzzy neural network (FNN) establishes an assessment framework that combines robustness with adaptive capability by integrating the semantic processing capacity of fuzzy logic with the nonlinear learning characteristics of neural networks (the FNN architecture is illustrated in Figure 2). The implementation process consists of three key phases: First, the indicator weights generated through the analytic hierarchy process (AHP) are incorporated as input variables alongside empirical survey data. Second, the input data were subjected to fuzzification processing employing membership functions, while initial fuzzy rules were extracted from both the AHP hierarchical structure and fuzzy comprehensive evaluation (FCE) results, effectively combining expert knowledge with data-driven methodology. Finally, the iteratively optimized training outcomes are processed to generate outputs, thereby establishing the complete FNN model. This hybrid approach systematically integrates qualitative expert judgment with quantitative data analysis through its unique fusion of fuzzy inference systems and neural network learning mechanisms.

5. Analysis of Results and Development Strategies and Suggestions

5.1. Analysis of Results

The membership degrees of indicators, 800 valid questionnaire data, and the previously derived indicator weights obtained through the analytic hierarchy process (AHP) were correspondingly input into the fuzzy neural network (FNN). Using the weight values of each indicator from the four judgment matrices as desired outputs, optimized indicator weights were obtained through multiple iterative training sessions (see Figure 3). Comparative analysis of the training and testing results using both backpropagation (BP) neural networks and FNN approaches yielded performance comparison charts (see Figure 4) and error analysis tables (see

Table 8. Comparison of error and accuracy between BP and FNN.

	MAE	RMSE	Prediction Accuracy
BP	17.57	22.1406	71.9907%
FNN	15.6784	20.2931	73.30%

), demonstrating the relative effectiveness of each methodology.

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{1}{n}}\sum _{i=1}^n (y_i-{\widehat{y}}_i{)}^2},$$

(8)

$$\mathrm{M}\mathrm{A}\mathrm{E}={\displaystyle \frac{1}{n}}\sum _{i=1}^n |y_i-{\widehat{y}}_i|,$$

(9)

In the analytical framework, $n$ corresponds to the total sample size, while $y_i$ and ${\widehat{y}}_i$, respectively, represent the observed (actual) value and model-predicted value for the $i$-th sample instance, with the subscript $i$ indexing individual observations across the complete dataset ranging from 1 to $n$.

From the above results, the following can be seen:

• The criterion layer weight for natural landscape resources was determined to be 0.5571, accounting for 55.71% of the total evaluation system, establishing it as the core dimension in assessing Guilin’s tourism resource development potential. Notably, the indicator weight for natural landscape quality reached 0.5847, representing over half of the natural landscape resource layer’s total weight, which underscores the unique market value of Guilin’s karst landforms and Li River landscapes as pivotal elements in tourism resource development. The weighting coefficients for visual experience optimization and Science popularization tour service were calculated as 0.1541 and 0.1957, respectively. Although these represent secondary rather than dominant factors in the evaluation framework, they nevertheless play a critical role in enhancing both visitor immersive experience quality and scientific knowledge dissemination efficacy within the tourism context.
• The management service resources criterion attained a weight of 0.3202, demonstrating the substantial influence of scenic area infrastructure and service quality on development potential. Within this dimension, recreation facility completeness (weight = 0.4658) and service comfort level (weight = 0.2772) emerged as critical limiting factors that directly affect visitor satisfaction. Conversely, transportation accessibility and environmental friendliness received comparatively lower weights, suggesting the need for targeted improvements through ecological retrofitting and intelligent transportation system development to optimize these aspects.
• The community cultural resources criterion was weighted at 0.1227, indicating that the potential of cultural elements in Guilin’s comprehensive tourism development has not yet been fully realized. Within this dimension, cultural heritage depth (weight = 0.5808) and experiential activity diversity (weight = 0.2382) emerged as core driving factors for cultural experiences, requiring enhanced attractiveness through intangible cultural heritage revitalization and interactive program design. In contrast, cultural communication effectiveness (weight = 0.1062) and community participation level (weight = 0.0748) received relatively lower weights, necessitating improvement strategies such as new media marketing and community co-construction mechanisms to boost their impact.
• As visualized in Figure 4, the comparative analysis of prediction results demonstrates that the FNN’s prediction curve (red) closely aligns with the actual values (blue), particularly within the test sample interval (e.g., x-axis 50–100), where its fluctuation amplitude is significantly smaller than that of the BP neural network. This indicates FNN’s superior capability in capturing data trends. Table 8 further reveals that the FNN more accurately identifies inherent data patterns, especially when handling complex nonlinear relationships in tourism resource evaluation, where its fuzzy membership functions and rule base effectively mitigate noise interference. By integrating fuzzy logic’s semantic processing capacity with neural networks’ adaptive learning properties, the FNN exhibits distinct advantages in tourism resource assessment. Compared to the BP neural network, it achieves reduced error rates while maintaining deep compatibility with the AHP-FCE framework, thereby providing a high-precision, robust analytical tool for evaluating and sustainably developing Guilin’s tourism resources.

5.2. Discussion

In the experimental section of this study, several key parameters were carefully selected to enhance the prediction performance and generalization ability of the fuzzy neural network model. Firstly, during the fuzzy inference system initialization, the subtractive clustering method was adopted with a cluster influence range set to 0.5, which is a commonly used value that balances model complexity and computational efficiency while avoiding overfitting due to excessive membership functions. Secondly, during the model training phase, the Epoch Number of ANFIS is set to 10. Considering the small sample size, this value ensures convergence while avoiding noise interference caused by excessive iterations. In addition, to ensure the effectiveness of the input features of the model, we pre-set a variance threshold to exclude feature columns with nearly zero variance, ensuring that the training data has sufficient discriminability. Finally, the data normalization process adopts the mapminmax method to uniformly map the input and output to the [0,1] interval, which helps to improve the stability and numerical accuracy of model training. The above parameter settings are combined with experimental data characteristics and recommendations from existing literature, aiming to construct a stable and efficient fuzzy neural network regression model.

In comparison with existing studies, the integrated AHP-FCE-FNN model proposed in this research exhibits both methodological advantages and certain limitations. For example, Tao Chen et al. utilized the AHP-FCE method to evaluate sustainable regional revitalization strategies, effectively structuring expert judgments but lacking the ability to capture nonlinear relationships in the data, which our model addresses through the inclusion of a fuzzy neural network. Similarly, Peng Ying employed a TCN-LSTM hybrid model for scenic spot passenger flow prediction. While this deep learning approach performs well in temporal modeling, it does not incorporate expert knowledge or a multi-criteria decision-making framework, which reduces interpretability, which our model improves by integrating AHP and FCE.

Lin S et al. developed an AHP-BP-based model to assess agritourism-integrated rural environments under the “dual-carbon” policy in Zhejiang, effectively combining expert judgment with neural network prediction. However, their study did not incorporate fuzzy comprehensive evaluation, limiting its ability to handle uncertainty in expert input. In contrast, our model integrates FCE to better process subjective and linguistic information. While Lin S et al. used structured panel data suitable for objective assessments, our questionnaire-based data reflects real tourist perceptions and resource value more effectively. Moreover, the BP neural network they applied is prone to local minima, whereas the FNN in our model offers stronger nonlinearity and generalization.

Finally, Dai Z et al. combined AHP-FCE with GA-BP in smart learning environment assessment, improving optimization through genetic algorithms but increasing computational complexity. In contrast, our AHP-FCE-FNN model balances interpretability, adaptability, and efficiency, although it may still be constrained by the quality of initial expert judgments and the sensitivity of fuzzy rule generation.

5.3. Development Strategies and Suggestions

5.3.1. Government-Led Eco-Tourism Development in Guilin

Our research demonstrates that strengthening governmental leadership in optimizing top-level resource management represents a critical strategy for sustainable tourism development. Natural landscape resources serve as the core driver of Guilin’s tourism development, with landscape quality occupying the dominant position. Implementation priorities should include the following: karst landform restoration, water resource management, and smart scenic area construction. The operational plans for scenic area planning rationality require adjustment to prevent overdevelopment and ecological imbalance, with particular emphasis on avoiding excessive commercialization while prioritizing science popularization tour service to enhance the scientific rigor and interactivity of visitor experiences. These measures constitute not only urgent requirements for local economic development but also practical needs for establishing another flagship brand for the city. Guilin municipal government should proactively exercise its leadership role to create an enabling environment for sustainable tourism growth.

5.3.2. Guilin Cultural Tourism Brand Upgrade Strategy

Guilin should implement a comprehensive brand enhancement strategy centered on establishing a “1 + N” brand matrix, with the iconic “Guilin’s landscape is the finest under heaven” as the core IP complemented by specialized sub-brands including “Zhuang and Yao Ethnic Culture” and “The Ancient Charm of the Lingqu Canal”, which will be achieved through five key interventions: firstly, developing immersive experiential offerings utilizing VR technology to reconstruct historical scenes of Lingqu Canal while introducing innovative night tourism products such as the “Li River Night Light Show”; secondly, executing integrated marketing campaigns across multiple media channels to optimize visual experiences and strengthen cultural communication, concurrently enhancing the “Guilin International Landscape Culture Tourism Festival” by incorporating new thematic sections featuring intangible cultural heritage performances and ethnic culinary markets; thirdly, fostering tourism-city integration through the deployment of smart tourism service centers in strategic locations and the formulation of standardized service evaluation protocols to elevate environmental sustainability standards; fourthly, expediting transportation infrastructure upgrades to improve regional accessibility; and fifthly, revitalizing community cultural assets by establishing heritage workshops that enable tourist participation in traditional artisanal activities with equitable revenue-sharing arrangements, coupled with blockchain-based authentication systems for cultural products—this holistic approach systematically integrates technological innovation, cultural conservation, sustainable practices, and infrastructure modernization to position Guilin as a city that achieves optimal equilibrium between tourism development, cultural authenticity preservation, and ecological protection.

6. Conclusions

The proposed FNN-AHP-FCE hybrid model establishes a novel methodological framework for tourism resource evaluation by synergistically integrating fuzzy logic and neural networks, achieving dynamic coupling of subjective and objective data. This approach establishes a structured decision-making foundation through AHP’s hierarchical architecture (with all judgment matrices demonstrating CR < 0.1), while leveraging FNN’s adaptive learning capabilities for data-driven optimization. Furthermore, it incorporates FCE’s robust fuzzy boundary processing to effectively address uncertainty, collectively serving as a scientific engine to facilitate Guilin’s development as a world-class tourism destination.

Notwithstanding its contributions, this study has two notable limitations that warrant discussion. First, the temporal resolution of our dataset remains constrained by the absence of real-time visitor behavior metrics (e.g., mobile device trajectories, instantaneous visitor flows), which may limit the model’s adaptability in dynamic scenarios—a gap that could be addressed through IoT-enabled data streams in future iterations. Second, while the model demonstrates robust performance in Guilin’s karst-dominated environment, its generalizability to geographically heterogeneous destinations (e.g., UNESCO-listed heritage cities) requires further validation through cross-regional benchmarking studies.

Building upon the current findings, two pivotal research trajectories warrant further investigation. First, the integration of multi-source big data analytics—encompassing social media sentiment mining and visitor emotion recognition—with long short-term memory (LSTM) networks could significantly enhance temporal pattern recognition, thereby establishing a dynamic evaluation framework. Second, the application of transfer learning and domain adaptation techniques shows promising potential to improve the model’s cross-regional generalizability across diverse tourism resource typologies, particularly for UNESCO heritage sites and ecological preservation zones.

These proposed improvements would not only elevate the model’s scientific rigor and universal applicability but also provide a theoretical framework for coordinating 60 human–environment relationships in natural heritage sites under China’s “Beautiful China” initiative. Such advancements would facilitate the realization of synergistic benefits across ecological preservation, economic development, and social welfare.