USING ARTIFICIAL NEURAL NETWORKS TO EVALUATE THE EFFECT OF LANDCOVER IN ESTIMATING SNOWPACK PROPERTIES WITH SSM/I DATA

摘要

This study focuses on to improving the estimation of snow depth, SWE, in Great Lakes area of the United States. This research consists of two major parts. In the first part, three successive winter seasons (2001-2004) were investigated for behavior of SSM/I channels vs. snow depth for various points in the study area. We looked into the correlation between various "signatures" of SSM/I channels such as 19V-37V, 19H-37H, and 22V-85V and SWE and snow depth. The second part is focused on using an Artificial Neural Network (ANN) model to identify snow covered area. The model inputs are the SSMI channels and the output is snow cover. To account for the variation of land cover, Normalized Difference Vegetation Index (NDVI) was introduced as additional input to the ANN model. The ground truth data were obtained from two different sources. The first one, National Climate Data Center (NCDC) snow monitoring section, provides gauge measurements for snow depth as point measurements. The second, SNODAS dataset is produced by NOHRSC using gamma radiations and gauge measurements combined with a physical model. The results show up to 85 percent correlation between SSMI channels and snow depth and SWE in high latitudes. This tends to decrease towards to the lower latitudes. The introduction of NDVI as an additional input increases the overall accuracy of estimated SWE between 10 to 15 percent