code

These three scripts together establish the complete technical workflow of this study, from image classification to spatial driver analysis. First, the CNN classification script uses EfficientNetV2-S to build a Cultural Ecosystem Services (CES) image classification model. By incorporating a two-stage training strategy, data augmentation, focal loss, EMA weights, and center-priority test-time augmentation (TTA), the script performs multi-class classification of social media images and outputs the best-performing model together with training diagnostics, thus providing a reliable CES data foundation for subsequent spatial analysis. Second, the GRF bandwidth search script focuses on parameter optimization for the Geographical Random Forest (GRF) model. Using grid-level CES diversity and environmental variables, it conducts a systematic search of key parameter combinations, such as bandwidth, ntree, and mtry, and evaluates model performance using out-of-bag mean squared error (OOB MSE), thereby identifying the optimal settings for robust spatial modeling. Finally, the RF-SHAP-GRF integration script combines Random Forest (RF), SHAP, and GRF analyses into a unified framework. It identifies the global importance of explanatory variables, interprets the direction and magnitude of their effects on CES diversity, and further reveals the spatial non-stationarity of driving mechanisms. At the same time, it generates the core visual outputs of the study, including RF importance plots, SHAP summary plots, RF–SHAP overlay figures, and GRF local importance maps. Taken together, these three scripts correspond to the three major stages of the analytical pipeline—image classification, parameter optimization, and driver interpretation—and form a complete, reproducible, and highly interpretable methodological framework for the study.

本研究的完整技术流程由这三段脚本共同搭建完成，覆盖从图像分类到空间驱动因子分析的全链条环节。首先，卷积神经网络（Convolutional Neural Network, CNN）分类脚本采用EfficientNetV2-S构建文化生态系统服务（Cultural Ecosystem Services, CES）图像分类模型。该脚本整合两阶段训练策略、数据增强、焦点损失、指数移动平均（EMA）权重以及中心优先测试时增强（Test-Time Augmentation, TTA）技术，对社交媒体图像执行多分类任务，并输出性能最优的模型与训练诊断信息，为后续空间分析提供可靠的CES数据基础。其次，地理随机森林带宽搜索脚本聚焦地理随机森林（Geographical Random Forest, GRF）模型的参数优化任务。该脚本以网格级CES多样性与环境变量为基础，系统搜索带宽、ntree、mtry等关键参数组合，并通过袋外均方误差（Out-of-Bag Mean Squared Error, OOB MSE）评估模型性能，最终确定适配稳健空间建模的最优参数配置。最后，RF-SHAP-GRF整合脚本将随机森林（Random Forest, RF）、SHAP与GRF分析整合为统一框架。该脚本首先识别解释变量的全局重要性，解析其对CES多样性的影响方向与强度，并进一步揭示驱动机制的空间非平稳性。与此同时，脚本生成本研究的核心可视化成果，包括RF重要性图、SHAP汇总图、RF-SHAP叠加图以及GRF局部重要性地图。综上，这三段脚本分别对应分析流程的三大核心阶段：图像分类、参数优化与驱动因子解释，共同构成本研究完整、可复现且具备高可解释性的方法论框架。