Spatio-temporal variability of dust on snow: interactions with topography and snowpack dynamics observed with UAVs

Aeolian dust surface deposition on seasonal snowpacks strongly influences snow albedo and melt dynamics, yet the environmental drivers of dust accumulation and redistribution at metre-scale resolution remain incompletely understood. UAV-based multispectral imagery enables detailed mapping of snow surface darkening associated with Light Absorbing Particles (LAP) such as mineral dust, offering new opportunities to investigate spatial distribution patterns in complex alpine terrain. This study examines the potential of UAV multispectral acquisitions to determine dust-on-snow spatial distribution and the relative influence of topographic factors on its variability during the seasonal evolution of the snowpack.

Data were collected in 2025 over a ~0.5 km² alpine study basin in the Spanish Pyrenees using a MicaSense Altum multispectral sensor mounted on a DJI Matrice 300 UAV. Five UAV acquisition campaigns were conducted between initial Saharan dust deposition and snowpack melt-out. Spectral indices sensitive to snow surface darkening by LAP were computed from the UAV imagery. Additionally, from 10 to 20 distributed in situ snow surface samples were manually collected concurrently with UAV acquisition flights to determine surface LAP concentration and close-range spectral response using a hand-held hyperspectral radiometer to calibrate UAV-derived surface LAP concentration.

A suite of potential predictors to represent potential controls on surface LAP redistribution and accumulation were selected: elevation, slope, northness, topographic position index (TPI), maximum upwind slope (Sx), diurnal anisotropic heat index (DAH), snowpack depth and snowpack depth difference. Random forest (RF) models were applied independently to each acquisition date in order to assess how the relative importance of these controls evolved through time considering the different states of the dust layer in the snowpack.

The RF models generally reproduced the spatial variability of the LAP indices well, according to internal out-of-bag evaluation and the RMSE errors remained around low for days with larger LAP concentration variability. Throughout the study period, the state of the snowpack notably influenced the relative importance of the predictors to the response variable. We were able to observe days in which fresh snow partially covered the dust layer, causing predictor variables related to snow accumulation and elevation to show the highest relative importance. Subsequently, after the full surfacing of the dust layer, the largest LAP concentrations were found in concave areas, notably increasing the relative importance of TPI.

The results demonstrate the value of combining multi-temporal UAV multispectral observations with interpretable machine-learning approaches to account for the temporal sequence of dust deposition, burial, re-exposure, and melt to advance understanding of aeolian dust processes in alpine snow-covered environments.