From correlation to causation: Causal mechanisms, scale dependency, and structure–function mismatch in landscape-water quality responses in plateau lake basins

Deciphering the causal effects of landscape patterns on water quality is critical for watershed management, yet most studies are confined to correlation analyses that fail to provide reliable causal evidence, particularly in sensitive plateau lake basins under compounded anthropogenic pressures. Taking the globally representative Fuxian Lake basin as a case study, this study developed an integrated analytical framework of production-living-ecological spaces, landscape structure, and water quality response (PLES-LS-WQ). Machine learning (XGBoost-SHAP) and causal inference model (Causal Forest) were combined with land-use scenario simulation (PLUS model) to quantify the true causal effects of key landscape pattern factors on water quality (TN, TP, COD) across multiple scales and scenarios. The results challenged several conventional assumptions: (1) Significant context dependency was revealed. The causal effects of landscape patterns were not fixed but determined by the dominant pollution source (mining, urban, or agricultural). For instance, the aggregation of production space improves water quality in mining-dominated areas but degrades it in agricultural zones. (2) The study provided the first causal evidence for a structure–function mismatch. Contrary to the traditional belief that ecological connectivity is beneficial, clustered ecological spaces under high pollution loads (mining and urban) exerted significant negative causal effects on water quality, indicating that their physical structure failed to translate into ecological purification functions and even facilitated pollutant transport. (3) Scale analysis and scenario simulation further indicated that ignoring context dependency and functional mismatch may lead to management failure or even adverse outcomes. By providing robust causal evidence, the study argued for a paradigm shift in watershed management from the pursuit of universal structural optimization towards a context-specific, function-oriented approach.