Research on Visual Preference Prediction and Design Guidance for Urban Waterfront Scenery by Integrating Street View Image Data and Machine Learning Methods

To address the issues of high subjective evaluation reliance and insufficient design transformation in the research on urban waterfront landscape visual preferences in the SichuanChongqing region, this study uses street scene images and machine learning as technical support to construct a visual preference prediction model and extract design guidance strategies. The study selected typical waterfront areas in the Sichuan-Chongqing region (such as the Jialing River waterfront area in Chongqing and the Jinjiang River waterfront area in Chengdu) to collect street scene images (860 valid samples), and obtained preference score data through public surveys in the Sichuan-Chongqing region (620 valid samples). The ResNet50 was used to extract image features, and XGBoost was combined to build the prediction model. The results showed that the determination coefficient $\mathbf{R}^{\mathbf{2}}$ of the model on the test set reached 0.82, and the generalization verification (selecting another waterfront area in the Sichuan-Chongqing region) had an $\mathbf{R}^{\mathbf{2}}$ of 0.78; SHAP value analysis indicated that five types of elements such as water visibility and vegetation layering had significant impacts on preferences (for example, for every $10 \%$ increase in water visibility, the average preference score increased by 0.8 points). Based on the quantitative correlation patterns, a typespecific waterfront landscape design guidance strategy adapted to the regional characteristics of Sichuan-Chongqing was formed, providing an operational method for data-driven landscape design in the Sichuan-Chongqing region.