Non-stationary turbulence induced by sub-mesoscale disturbances can substantially bias flux estimates, yet is largely overlooked in existing evaluation frameworks. Using turbulence measurements over the heterogeneous Loess Plateau, we found that classical stationarity tests classify >60% of records containing clear submesoscale disturbances as stationary, highlighting critical limitations over complex terrain. To address this limitation, a separation method for Sub-Mesoscale and Turbulent motions (SMT) is applied to synthetic non-stationary time series, accurately reconstructing prescribed sub-mesoscale structures and demonstrating high fidelity. Then, we develop multi-variable non-stationarity indices based on the SMT-derived components. Controlled sensitivity experiments with synthetic data demonstrate that these indices track sub-mesoscale intensity clearly and monotonically, providing a more objective diagnostic than traditional metrics. Anchored by thresholds derived from these experiments, we propose a nine-grade classification framework. This framework offers a robust diagnostic tool for flux evaluation over complex surfaces, with implications for weather forecasting, carbon cycle assessments, and climate model evaluation.