Nonlinear and spatially differentiated assessment of visually perceived cultural ecosystem services in living heritage riverfronts

Large-scale living heritage landscapes require a CES perspective to understand the tension between heritage conservation and other cultural services. However, existing studies often represent heritage-related CES through recognized heritage resources, tourism hotspots, or aesthetic proxies. Taking the Grand Canal riverfront in southern Jiangsu as a case, this study examines whether ordinary municipal riverfront interfaces can provide heritage-related CES through visual perception. It evaluates two visually grounded CES dimensions—landscape aesthetic preference (HLP) and visual heritage value perception (HVP). A GPS-referenced image dataset was constructed from boat-based surveys, producing 1312 valid north-bank riverfront images with an average spacing of 31.4 m. OpenCLIP-based clustering and stratified sampling were used to select representative samples for face-to-face perception evaluation, and a fine-tuned deep-learning scoring model was then used to predict landscape aesthetic preference (HLP) and visual heritage value perception (HVP) across the full sample. XGBoost, SHAP, PDP, and GeoSHAP were further combined to identify key variables and interpret numerical and spatial nonlinearities. The results display that ordinary riverfront interfaces can provide heritage-related CES: 38% of samples outside heritage-concentration sections achieved above-average HVP scores. HLP and HVP are moderately correlated ( r = 0.675, p < .05). However, clear divergence emerges in industrial landscapes with limited greenery and in greenery-dominated scenes; With XGBoost models of both perceptions (validation R 2 values of 0.937 for HVP and 0.872 for HLP), SHAP results reveal that visual configuration variables dominate both perceptions (over 60%), while element proportions contribute more to HVP and landscape pattern indices contribute more to HLP. Nonlinear analyses further indicate that HLP is more sensitive to preferred ranges and compositional balance, whereas HVP responds more strongly to the visual prominence of specific elements. GeoSHAP results also reveal broader spatial variation in HLP effects, while HVP-related spatial effects are more concentrated in the center-eastern section. These findings suggest that heritage-related CES in living heritage landscapes cannot be inferred solely from heritage proximity or aesthetic preference, and that CES evaluation should shift toward a perception-based, nonlinear, and spatially differentiated understanding of heritage-related services.