Feature contributions and predictive modeling of aeolian sand transport detection in the atmospheric surface layer

• An interpretable machine-learning framework is developed for aeolian sand transport detection. • Streamwise velocity is identified as the dominant predictor of saltation occurrence. • Rolling-window statistics at an optimal 30-second scale markedly improve prediction skill. • Near-surface turbulent sweeps play a key role in saltation initiation. • Friction velocity governs sand saltation onset, with modulation by turbulence intensity. Predicting aeolian sand transport in the atmospheric surface layer (ASL) is difficult because turbulence-driven intermittency produces hysteresis and burst-like saltation. This study develops an interpretable machine learning framework for aeolian transport event detection using high-frequency wind and saltation flux measurements. The analysis compares raw velocity components, first-order statistics, fluctuation velocities, and second-moment derivatives including friction velocity ( u ∗ ) and turbulence intensity (TI). Model performance was benchmarked across averaging windows and heights and interpreted using SHapley Additive exPlanations (SHAP) to quantify the contribution of individual features to model predictions. Streamwise velocity emerged as the strongest predictor, with importance peaking near the surface and reemerging aloft where reduced dissipation provided cleaner signals. Rolling-window aggregates of first-order statistics achieved superior accuracy to raw inputs and the Time Frequency Equivalence Method, with optimal performance at a 30 s window that matched the velocity decorrelation timescale. Fluctuation velocities alone yielded poor results but enhanced predictive skill when combined with first-order statistics, with sweeps and outward interactions dominating saltation initiation. Models based on u ∗ and TI achieved comparable accuracy but required longer windows (300 s), and SHAP analysis indicated that features derived from near-surface winds, including u ∗ and TI as well as fluctuation components, ranked highest in importance. Overall, the framework improves predictive accuracy beyond traditional methods, preserves physical interpretability, and offers valuable insights for feature selection, measurement strategies, and atmospheric monitoring of aeolian processes.