Abstract
Abstract Parks are an essential component of urban green infrastructure, and their service capacity and residents’ perceptions directly influence the equity and overall effectiveness of green spaces. Existing studies often rely on a single data source or limited indicators, making it difficult to fully reveal the spatial variations in park perceptions and their underlying mechanisms. Based on 168 parks in Beijing, classified into five types, this study integrates social media text and image information, remote sensing imagery, POI distribution, and socioeconomic data. It establishes an analytical framework that includes 26 influencing factors and five perception dimensions. Semantic modeling of residents’ perceptions is achieved through multimodal label extraction and fine-tuning of the RoFormer model. By using XGBoost and SHAP methods, the study identifies the mechanisms through which multi-source features influence each perception dimension and uncovers the nonlinear interactions among variables. At the same time, an improved 2SFCA method is used to evaluate the service capacity of parks. The study finds that Beijing’s parks exhibit significant spatial differences in both perception experience and service capacity. Peripheral areas have substantial potential for improvement in accessibility and usability. Social media variables make notable contributions in multiple models. Ecological and service conditions are the core drivers of perception. In contrast, the influence of POI facilities and spatial structure variables depends more on specific park types and spatial contexts. Based on these results, this study develops a dual-path optimization strategy. It covers the enhancement of internal perception features and the optimization of external spatial structures. With dual optimization, park service capacity increased by 13.80, 14.33, and 14.54% under walking, cycling, and driving travel modes, respectively. This study provides a data foundation and methodological reference for identifying urban green space inequality and comparing optimization directions.