Abstract
With the increasing application of artificial intelligence in environmental design, computer vision and deep learning have emerged as crucial tools for understanding human visual perception. This study focuses on rural landscapes and proposes a visual comfort evaluation model that combines structural visual attributes with deep semantic representations. By employing Residual Network (ResNet101) and eXtreme Gradient Boosting (XGBoost), the model enables quantitative prediction of aesthetic preferences for landscape images. The dataset combines original field photographs with screened samples from the Places365 dataset, categorized and preprocessed under unified visual standards. Multiple visualization methods are also incorporated to enhance model interpretability. The results indicate that green ratio, sky openness, and edge density are among the most influential factors associated with visual comfort scores, although the overall predictive strength remains moderate. These results reflect a preference pattern characterized by brightness, naturalness, openness, and simplicity, although the overall predictive strength remains moderate. Furthermore, the study establishes a feedback framework of "score-structure-strategy" and presents four categories of rural landscape design recommendations, bridging perceptual modeling with spatial design. This research contributes to the quantification of perceptual aesthetics and offers a perception-oriented data framework for rural landscape design in China.