Integrating heterogeneous user-generated contents into spatial Modeling of agricultural landscape recreational services: A geographically weighted random Forest approach

Agricultural Landscape Recreational Services are crucial for farmland conservation and rural tourism; however, existing studies primarily rely on single-source User-Generated Content, which limits their representativeness. Given the strong spatial heterogeneity of agricultural landscape characteristics, this study develops an evaluation framework that integrates multi-source heterogeneous User-Generated Content with geospatial data. Using Qujing City, China, as a case study, a Geographically Weighted Random Forest (GWRF) model is introduced to predict Agricultural Landscape Recreational Services across the entire region, and Shapley Additive Explanations (SHAP) are used to interpret the driving mechanisms. The results show that, under Spatially Blocked Cross-Validation (SBCV), the model's independent predictive capacity is 0.417, establishing a reliable generalization baseline. The residual Moran's I is significantly reduced, quantitatively confirming the spatial independence and robustness of the evaluation process. Unlike global Random Forest, which averages the impact of features, Geographically Weighted Random Forest uncovers spatial variations in drivers: Agricultural Landscape Recreational Services in urban areas are primarily driven by transport accessibility, while those in rural areas heavily rely on accommodation facilities. SHAP analysis further reveals the dialectical interaction between the built and biophysical environments. Although single-variable analysis shows negative marginal effects from proximity to certain facilities due to commercial or visual disturbance, interaction analysis identifies the synergistic enhancement of a “service cluster” mechanism. The combination of convenient transport and comprehensive accommodation amenities creates a stronger conditional effect, effectively mitigating the negative impacts of individual features. However, this synergy is still constrained by topographic environments, as recreational services are notably limited in complex, high-altitude terrain. This study confirms the feasibility of integrating multi-source heterogeneous User-Generated Content with machine learning for predicting recreational services, and quantifies the generalization boundaries of geographic big data in spatial extrapolation. These findings provide decision-making references for the differentiated management of farmland ecosystem services and sustainable rural tourism development. • Integrate multi-platform Heterogeneous UGC to evaluate agricultural landscape recreational services. • Apply Geographically Weighted Random Forest to capture spatial heterogeneity of agricultural landscape recreational services. • Excessive traffic or commercial facilities may reduce agricultural landscape recreational quality. • “Service Clusters” mitigate the negative visual disturbance caused by individual commercial facilities. • The transition from spatial interpolation to extrapolation explains the decrease in R 2 while ensuring more rigorous out-of-sample validation.