Safety Risk Assessment of Cave Adventure Tourism Projects: A Case Study of Shierbeihou, Guizhou, China

2.1 Research design

This study combined indicator development, expert-based weighting, and empirical case evaluation in three sequential phases.

First, a preliminary indicator pool was developed from the literature, field observations, and expert input. Second, the Delphi method was used to screen and revise the indicators, and AHP was used to determine their relative weights. Third, the framework was applied to the Shierbeihou cave adventure tourism project using questionnaire data and fuzzy comprehensive evaluation.

2.2 Study area and data sources

The case study was the cave adventure tourism project at Shierbeihou Scenic Area in Suiyang County, Zunyi City, Guizhou Province, China. The project is closely related to Shuanghe Cave National Geopark and has typical karst cave characteristics. Its operation involves tourists, specialist guides, equipment and facilities, route organization, emergency rescue, medical support, and environmental risk factors, making it an appropriate case for applying the proposed framework.

Three data sources were used. Literature and policy materials on cave tourism, adventure tourism, tourism safety, and scenic-area management informed the initial risk pool. Field observations at Shierbeihou Scenic Area from 1 to 7 April 2026 covered the cave route, staff arrangements, equipment condition, warning signs, site management, and emergency arrangements. The empirical questionnaire survey was conducted on 5 and 6 April 2026.

Prior to data collection, research permission and informed consent documentation were obtained from Universitas Negeri Jakarta. Participation was voluntary and anonymous, and respondents were informed that the data would be used solely for academic research. The research permission number was 1724/UN39.6SP/sLT/2026.

2.3 Indicator development and Delphi consultation

The first indicator pool was established in the context of the operational characteristics of cave adventure tourism projects. Four general categories of risk emerged: personnel risk, facility risk, management risk and environmental risk. The first-level indicators, second-level indicators, and third-level candidate indicators were collected from the literature review, field observations, and experts' suggestions as an initial pool of indicators, with 4, 10, and 53 indicators, respectively.

Following the Delphi consensus approach, 11 experts were purposively recruited for three rounds [18]. Eligibility required more than 5 years of relevant experience; university academics had to hold at least an associate professor appointment, scenic-area personnel at least a deputy manager position, and technical specialists had to be certified national mountain-rescue trainers. The panel included scenic-area managers, a project manager, mountain-rescue trainers, and university researchers with 7-31 years of relevant experience; individual qualifications are reported in Table A1. All 11 experts completed all three anonymous rounds (100% response in each round; no attrition; Table A2). In each round, experts rated indicator importance and applicability and provided written comments.

Indicator screening combined two quantitative criteria with written expert comments. With 11 experts, a full-score ratio of 0.19 or lower corresponds to no more than two experts assigning the highest importance rating and was used as a study-specific low-recognition threshold. A coefficient of variation greater than 0.25 was treated as substantial dispersion, consistent with previous Delphi indicator-development research [19]. Indicators meeting either criterion were flagged for deletion or review, and final decisions to remove, merge, revise, or retain items also considered the experts' written comments. The principal first-round changes are summarized in Table A3. After three rounds, the final system comprised four primary, nine secondary, and 46 tertiary indicators.

2.4 Weight determination using the analytic hierarchy process

After the indicator system was finalized, the 11 experts rated the relative importance of indicators on a 1-9 scale. The expert ratings were analyzed using AHP software, and judgment matrices were constructed for indicators within each hierarchical group [19]. Local weights were derived from these matrices. The complete group-level judgment matrices are provided in Tables A4-A17.

The global weight of each secondary indicator was calculated by multiplying its local weight by the corresponding primary weight. The global weight of each tertiary indicator was calculated by multiplying its local tertiary weight by the corresponding secondary and primary weights.

For each judgment matrix, the maximum eigenvalue, consistency index, and consistency ratio were calculated. A consistency ratio below 0.10 was considered acceptable. Observed values ranged from 0.0000 to 0.0422 (Table 1), and matrix-specific statistics are reported in Tables A4-A17. Kendall's W, rather than the consistency ratio, was used to describe agreement across the Delphi panel.

Local and global weights for all 46 tertiary indicators are provided in Table A18.

Table 1. Consistency statistics for analytic hierarchy process (AHP) judgment matrices

2.5 Questionnaire survey and fuzzy comprehensive evaluation

The empirical survey was conducted on 5 and 6 April 2026 at Shuanghe Cave in Shierbeihou Scenic Area. Tourists participating in the cave adventure project were recruited on site by convenience sampling, while project managers and specialist guides involved in project operations were recruited purposively. The project served approximately 1,000 visitors per day, and the scenic area employed 334 staff across operational and support roles. A total of 458 questionnaires were returned. Eleven were excluded because of uniform response patterns (n = 9) or incomplete responses (n = 2), leaving 447 valid questionnaires (360 tourists and 87 practitioners; valid response rate, 97.5%). Responses were collected across four-time blocks on each survey day (Table A19), and respondent characteristics are summarized in Table A20.

Tourist and practitioner responses were pooled to estimate a mixed-stakeholder, perception-based risk profile. Because ordinary tourists may have limited access to internal operational information, scores for management, maintenance, professional qualifications, and medical support are interpreted as perceived risk signals rather than objective measures of operational performance. No respondent-specific weighting or subgroup analysis was applied.

The questionnaire contained the 46 final tertiary risk indicators. Each was rated from 1 (very low risk) to 5 (very high risk). Trained investigators explained the survey items before respondents rated each indicator based on direct participation, observation, or work experience. Response frequencies were converted to membership degrees to form the fuzzy evaluation matrix. The AHP weight vector was then combined with the matrix to obtain secondary, primary, and overall perception-based risk scores [21, 22].

The evaluation grade vector was defined as:

$V=\left(v_1, v_2, v_3, v_4, v_5\right)=(1,2,3,4,5)$               (1)

where, 1 represents very low risk, and 5 represents very high risk. The membership degree of indicator i belonging to grade j was calculated as:

$r_{i j}=\frac{n_{i j}}{N_i}$              (2)

where, n_ij represents the number of respondents selecting grade j for indicator i, and N_i represents the total number of valid responses for indicator i. The fuzzy evaluation matrix was expressed as:

$R=\left[\begin{array}{ccccc} r_{11} & r_{12} & r_{13} & r_{14} & r_{15} \\ r_{21} & r_{22} & r_{23} & r_{24} & r_{25} \\ \ldots & \ldots & \ldots & \ldots & \ldots \\ r_{m 1} & r_{m 2} & r_{m 3} & r_{m 4} & r_{m 5} \end{array}\right]$              (3)

The comprehensive membership vector was calculated as:

$B=W \times R=\left(b_1, b_2, b_3, b_4, b_5\right)$             (4)

where, W is the AHP-derived weight vector, R is the fuzzy evaluation matrix, and B is the comprehensive membership vector. The final comprehensive evaluation score was calculated as:

$S=B \times V^T$             (5)

Because the indicator system had a multi-level structure, the evaluation was conducted layer by layer, from tertiary indicators through secondary and primary indicators to the overall risk score.

2.6 Internal consistency and data suitability

Cronbach's alpha was used to assess internal consistency, while the Kaiser-Meyer-Olkin measure (KMO) value and Bartlett's test of sphericity were retained as data-suitability checks. For the 447 valid responses across 46 items, Cronbach's alpha was 0.948, KMO was 0.958, and Bartlett's test was significant (p < 0.001).

This study developed and applied a project-level safety risk assessment framework for cave adventure tourism. The framework integrates personnel, facility, management, and environmental dimensions, with facility, personnel, and management risks carrying the greatest relative weight.

In the Shierbeihou case, the overall fuzzy evaluation score was 2.3338, corresponding to the low-risk category on the study's five-level perception-based scale. Elevated local scores remained for facility inspection and maintenance, accidental injury, tourist health, guide emergency response, project-manager safety training, response timeliness, and accident handling. The overall category should therefore be interpreted as a summary of perceived risk, not as an objective safety certification.

The study's contribution is the integration of Delphi-based indicator refinement, AHP weighting, and fuzzy comprehensive evaluation into a project-level pathway for identifying and prioritizing cave adventure tourism risks.

Several limitations should be considered. The empirical application involved one site, and the two-day on-site convenience recruitment of tourists and purposive recruitment of practitioners may limit representativeness. Most evidence came from questionnaires and expert judgment. Tourist and practitioner responses were pooled, although some management-related indicators may be less observable to tourists; investigator explanations may also have influenced ratings. The same expert panel contributed to indicator development and weighting, so the study represents framework development and single-case application rather than independent validation. No objective accident, inspection, environmental monitoring, or rescue-performance records were available for triangulation. Future studies should use role-stratified samples, independent validation panels or sites, and objective operational data.