Understanding the spatiotemporal dynamics of soil redistribution is essential for quantifying watershed-scale soil erosion and regional carbon cycling. Previous research has predominantly emphasized soil detachment from "source" areas, while sediment re-deposition and the regulatory role of non-photosynthetic vegetation (NPV) remain insufficiently accounted for, leading to biased estimates of net erosion fluxes. To address this limitation, this study developed a novel multi-model coupled framework based on Linear Spectral Unmixing (LSU) and geomorphological source-sink theory to quantitatively characterize long-term erosion-transport-deposition dynamics. Within this framework, partial correlation analysis and an XGBoost-SHAP approach were integrated to disentangle the relative importance and interactions of dominant drivers. Simulations based on the Revised Universal Soil Loss Equation (RUSLE) coupled with a Transport Limited Sediment Delivery (TLSD) model (R-2 = 0.80) revealed an overall declining yet fluctuating trend in net erosion rates from 1975 to 2023 (63.86-302.64 t & centerdot;km(-2)& centerdot;a(-)(1)), with localized rebounds after 2005 driven by climate variability. Incorporating NPV into C-factor estimation improves the representation of vegetation controls and uncovers a unimodal, nonlinear threshold response between vegetation cover and net erosion. A "sensitive erosion transition zone" (16% < f(NPV) < 27% and 14% < f(PV) < 32%) is identified, reshaping the linear paradigm linking vegetation cover to erosion. Overall, net erosion dynamics arise from nonlinear interactions among topography, vegetation, climate, and human disturbance. The proposed LSU-RUSLE-TLSD framework grounded in "erosion source" and "deposition sink" theory provides a robust and transferable tool for assessing watershed-scale soil redistribution and its underlying driving mechanisms.