Mapping daily snow depth with machine learning and airborne lidar across two contrasting snowpacks

Abstract. Daily, basin-scale snow depth maps are needed for forecasting and operations, yet airborne lidar typically provides only episodic snapshots. We present a portable relative-depth machine-learning framework that converts a small number of lidar acquisitions plus a single daily driver time series (in-situ station or ERA5-Land) into temporally coherent, per-pixel daily snow depth maps. A random forest model is trained on lidar–driver differences where lidar supplies the spatial pattern of departures and the driver supplies temporal evolution; learning is constrained to observed conditions using a valid-pixel mask and synthetic zero-depth maps at season start and end. We evaluate the approach in two contrasting regimes—Mores Creek, Idaho, and Hubbard Brook, New Hampshire—using multi-year lidar records. Across both basins, performance is fit for purpose (R² 0.89–0.90; RMSE 8–28 cm; MAE 5–19 cm; near-zero bias). Mores Creek, a larger heterogeneous western basin benefits more from adding lidar-informed residual maps, than Hubbard Brook, a smaller transitional eastern basins, where the primary value is correcting local departures from the mean and refining melt timing. Spatial diagnostics and Shapley values show that residuals are organized by landscape controls including elevation, aspect/northness, microtopography, slope, and a redistribution proxy.  Lidar-cadence experiments indicate diminishing returns after a few acquisitions: roughly five flights in early season, four in mid-winter, and five in late season recover most skill at Mores Creek, while Hubbard Brook shows a similar pattern with about three flights in early-mid winter and five in mid-late winter, but with greater variability in model skill. The timing of lidar acquisitions also influences model transferability. Models trained on mid-season data generalize well to both early and late season conditions, whereas models trained on late-season data perform poorest when applied to early season dates. ERA5-driven runs closely track in-situ-driven results, indicating the feasibility of using reanalysis datasets where stations are absent. The method is intentionally interpolative and should be applied within its area of applicability, but it offers a practical route from episodic lidar snow surveys to meter-scale, daily, basin-scale products and actionable guidance on survey timing and frequency.