|dc.description.abstract||Accurate areal measurements of snow-cover extent are important for hydrological and climate modeling. The traditional method of mapping snow cover is binary where a pixel is approximated to either snow-covered or snow-free. Fractional snow cover (FSC) mapping achieves a more precise estimate of areal snow-cover extent by determining the fraction of a pixel that is snow-covered. The two most common FSC methods using Moderate Resolution Imaging Spectroradiometer (MODIS) images are linear spectral unmixing and the empirical Normalized Difference Snow Index (NDSI) method. Machine learning is an alternative to these approaches for estimating FSC, as Artificial Neural Networks (ANNs) have been used for estimating the subpixel abundances of other surfaces. The advantages of ANNs over the other approaches are that they can easily incorporate auxiliary information such as land-cover type and are capable of learning nonlinear relationships between surface reflectance and snow fraction. ANNs are especially applicable to mapping snow-cover extent in forested areas where spatial mixing of surface components is nonlinear.
This study developed an ANN approach to snow-fraction mapping. A feed-forward ANN was trained with backpropagation to estimate FSC from MODIS surface reflectance, NDSI, Normalized Difference Vegetation Index (NDVI) and land cover as inputs. The ANN was trained and validated with high spatial-resolution FSC derived from Landsat Enhanced Thematic Mapper Plus (ETM+) binary snow-cover maps.
ANN achieved best result in terms of extent of snow-covered area over evergreen forests, where the extent of snow cover was slightly overestimated. Scatter plot graphs of the ANN and reference FSC showed that the neural network tended to underestimate snow fraction in high FSC and overestimate it in low FSC. The developed ANN compared favorably to the standard MODIS FSC product with the two methods estimating the same amount of total snow-covered area in the test scenes.||en