<p>Driven by the concepts of smart cities and humanized design, the static evaluation and optimization methods of traditional garden landscapes can no longer meet the dynamic and personalized demands of residents for high-quality living Spaces. This study innovatively proposes an interactive landscape optimization framework integrating multimodal visual feature temporal alignment, cross-level emotional dependence modeling, and emotion-spatial potential correlation mining, breaking through the limitations of traditional methods in multimodal fusion, spatial correlation analysis, and model robustness. This method captures the behaviors and microscopic expressions of tourists through the visual sensor network in the landscape space, solves the problem of asynchronous data by using the multimodal time series alignment mechanism, and extracts highly robust behavior data based on the improved DCNN. Construct a hierarchical emotional model with bidirectional feedback to quantify the dynamic emotional evolution of tourists based on the V-A dimension; By controlling confounding factors through the emotion-spatial potential correlation mining model and combining with the Bayesian optimization algorithm enhanced by ecological constraints, an optimization scheme that takes into account emotional experience, ecological protection and functional practicality is generated. Under the experimental settings and datasets of this study, the F1-score of emotion recognition is stable at 92.7% ± 0.8%. The improvement rate of emotional income of the Bayesian optimization scheme is 12.3%-18.7% higher than that of MOEA and GA. The optimization scheme increases the average tourist satisfaction by 30.3% ± 2.1%, and it is stable and effective in different types of gardens. This research provides theoretical innovation and engineering support for data-driven precise, humanized and sustainable landscape design, and clarifies the implementation path and ethical norms.</p>