Understanding wetland park feature influence through cross-regional multimodal analysis and interpretable modeling

Wetland parks serve both ecological conservation and social service functions, and the mechanisms through which their spatial characteristics influence public perception have attracted wide attention. Given the ecological and socio-economic differences between China and the United States and the availability of social media data, this study focuses on 147 wetland parks in both countries. It collects image-text reviews and ratings, builds a labeling system covering ecological environment, infrastructure, and user experience, and develops a unified feature framework based on multimodal data. The study extracts text, image, and fused features for predicting sentiment values and scores, and applies the SHAP method to explain feature contributions and their interactions. On this basis, high-contribution features identified by SHAP analysis are selected as optimization variables. A multi-objective optimization model is constructed with scores, sentiment values, and resource costs as objective functions. The NSGA-III algorithm is then applied to obtain Pareto-optimal solutions and explore balanced pathways for enhancing perception under limited resource input. The results show that text features better capture subjective perceptions, while image features reflect objective landscapes. Multimodal fusion outperforms single modalities in predictive performance, enhancing label coverage and model interpretability, with a clear advantage in compensating for the limitations of image features. Perceptions of Chinese wetland parks rely more on the combined influence of ecological and service features, while U.S. wetland parks exhibit a more balanced multi-feature synergy. Further comparison reveals significant differences between the two countries in feature contribution rankings, feature interaction effects, and perception patterns. This study proposes an analytical framework that integrates perception modeling and spatial optimization. It not only reveals the central role of ecological and service features in shaping sentiment values and scores but also verifies balanced pathways between perception improvement and resource control through multi-objective optimization.