Objective:　In　response　to　the　challenges　posed　by　mental　health　issues　among　college　students　and　the　declining　quality　of　campus　environments,　this　study　aims　to　reveal　the　complex　mechanisms　underlying　the　relationship　between　campus　audiovisual　environments　and　the　quality　of　students＇　attention　recovery.　It　further　explores　campus　landscape　optimization　pathways　driven　by　multi-source　data,　providing　scientific　basis　for　sustainable　campus　planning.　Methods:　Taking　Fuzhou　University　Town　as　a　case　study,　this　study　integrates　machine　learning　technology　with　multi-source　data　(street　view　images,　social　media　text,　and　PRS-11　questionnaires)　to　construct　a　“multi-modal　perception　mechanism　analysis-dynamic　evaluation　iteration”　framework.　The　CNN-BiLSTM　model　was　used　to　predict　attention　recovery　quality,　combined　with　HRNet　semantic　segmentation,　GBRT　soundscape　prediction,　and　CSV-T4SA　sentiment　analysis　models　to　quantify　audiovisual　elements.　XGBoost　models　and　SHAP　interpretability　analysis　were　employed　to　reveal　the　effects　and　interaction　mechanisms　of　variables.　Results:　(1)　Attention　recovery　quality　is　significantly　higher　in　liberal　arts　and　agricultural/forestry　universities　than　in　science　and　engineering　universities,　with　boundary　effects　and　the　synergistic　design　of　humanistic　soundscapes　being　key　factors;　(2)　SHAP　analysis　identifies　humanistic　soundscapes,　natural　soundscapes,　and　color　complexity　as　core　influencing　factors,　with　their　effects　exhibiting　significant　threshold　characteristics;　(3)　Linear　interaction　mechanisms　among　audiovisual　elements　are　discovered,　such　as　the　interaction　between　vegetation　density　and　building　enclosure　degree　enhancing　recovery　efficacy,　and　the　synergistic　design　of　musical　soundscapes　and　paving　materials　can　optimize　perceptual　experiences.　Conclusion:　By　innovatively　integrating　multi-source　data　and　machine　learning　techniques,　this　study　systematically　analyzes　the　relationship　between　campus　audiovisual　environments　and　attention　recovery,　breaking　through　the　limitations　of　traditional　linear　analysis.　The　proposed　“threshold　response　design”　and　“cross-modal　collaborative　optimization”　strategies　provide　a　new　paradigm　for　campus　planning,　validate　the　scientific　value　of　multi-sensory　interaction　design　for　mental　health　promotion,　and　offer　a　transferable　methodological　framework　for　global　university　environmental　upgrades.　Copyright　©　2025　Chen,　Chen,　Hong,　Zhuang,　Su　and　Ding.