CR 75:181-199 (2018)  -  DOI: https://doi.org/10.3354/cr01518

ABSTRACT: Snow cover modelling is primarily focussed on snow depletion in the context of hydrological research. Degree-day or temperature index models (TIMs) as well as energy balance models (EBMs) are conventional to quantify catchment runoff. Whereas the former exploit relationships between snow (and/or ice) melt and air temperatures, the latter rest upon quantifying melt as the deviation from heat balance equations. However, the 2 approaches contain distinct drawbacks. For example, increasing temporal resolution decreases the accuracy of TIMs, and no spatial variability is provided, whereas EBMs have large dataset requirements for climate, landscape and soil properties. Nevertheless, detailed knowledge about shifts in seasonal ablation times and spatial distribution of snow cover is crucial for understanding hydrological systems, plant distribution and various other research interests. Therefore, we propose a statistical model based on a combination of high resolution spatio-temporal climate datasets and climate-related topographic data, which were obtained by meteorological network stations, remote sensing and GIS analysis. The main objectives were to identify suitable inputs and to develop a robust binary snow distribution model that enables the mapping of major physical processes controlling snow accumulation, melt and stagnation in a high mountain environment in the Gaurishankar Conservation Area in Nepal. We used the random forest technique, which represents a state of the art machine learning algorithm. The snow distribution was predicted very accurately with high spatio-temporal resolution (daily on 0.5 × 0.5 km), with hit rates of around 90% and an overall model accuracy of 90.8% compared to independent Moderate Resolution Imaging Spectroradiometer (MODIS) observations.

KEY WORDS: Snow cover · Remote sensing · MODIS · Statistical modelling · Himalaya · Random forest