Study on Nonlinear Driving Mechanisms of Spatiotemporal Evolution in Sanjiang Plain Wetlands Based on Explainable Learning Methods

Against a backdrop of global climate fluctuations and intensifying human activities, wetlands are undergoing severe degradation. Understanding the mechanisms that govern wetland evolution is essential for the sustainable development of wetland ecosystems. However, wetland evolution is highly heterogeneous across space and time, and existing studies have generally paid insufficient attention to nonlinear effects and interactions among driving mechanisms, limiting a comprehensive understanding of its intrinsic processes. We integrated the Light Gradient Boosting Machine model with SHapley Additive exPlanations to explain the nonlinear driving mechanisms of spatiotemporal wetland evolution across the Sanjiang Plain using multitemporal remote sensing data from 1990 to 2023. Wetland dynamics exhibited a stage-dependent pattern characterized by substantial natural-wetland loss in the early period, followed by partial marsh-wetland recovery and continued artificial-wetland expansion. Artificial wetlands expanded continuously, whereas marsh wetlands declined markedly before 2005 and showed partial recovery thereafter. From 1990 to 2005, wetland evolution was mainly controlled by topographic and climatic factors. From 2005 to 2023, the influence of socioeconomic development and proximity to the road network on wetland change intensified. Nonlinear interactions among driving factors shifted from synergistic promotion by natural factors in the early stage to inhibitory effects among natural, socioeconomic, and locational factors in the later stage. This stage-specific change in driving mechanisms was crucial to the shift in dominant drivers of wetland evolution. By characterizing single-factor nonlinear effects and multifactor interactions, this study reveals the stage-specific driving mechanisms of wetland evolution in the Sanjiang Plain and provides scientific support for regional wetland management and remote sensing monitoring.