Under non-stationary conditions over complex terrain, turbulent motions are often superimposed on low-frequency processes such as sub-mesoscale disturbances and quasi-periodic oscillations, leading to pronounced scale mixing in flux estimation. Here we exploit the adaptive multiscale decomposition capability of complementary ensemble empirical mode decomposition (CEEMD) to develop a multiscale flux decomposition framework. High-frequency fluctuation signals are first decomposed into a set of intrinsic mode functions (IMFs); morphological criteria based on extrema-point density and the coefficient of variation (CV) are then introduced to automatically identify and separate small-scale turbulence from low-frequency sub-mesoscale components. Using a synthetically constructed non-stationary dataset together with power spectral density analyses, we demonstrate the physical consistency and robustness of the proposed scale attribution. Applying the framework to high-frequency observations from the Semi-Arid Climate and Environment Observatory of Lanzhou University (SACOL) over the Loess Plateau shows that, as non-stationarity intensifies, the fractional contribution of the low-frequency component to the total flux increases. Relative to total fluxes from conventional eddy covariance, the isolated small-scale turbulent flux exhibits fewer extremes, a more distinct diurnal cycle, and generally lower monthly means. For CO2 flux in particular, the monthly mean contribution from small-scale turbulence is about 25% lower than the conventional estimate, indicating that under strongly non-stationary conditions low-frequency motions have a significant impact on ecosystem-atmosphere carbon exchange. Overall, this study provides a new diagnostic tool for quantifying the relative contributions of multiscale processes to land-atmosphere exchange under non-stationary conditions.