Comparison of snow water equivalent retrieved from SSM/I passive microwave data using artificial neural network, projection pursuit and nonlinear regressions

Abstract

The snow water equivalent (SWE) for the Red River basin of North Dakota and Minnesota was retrieved from data acquired by passive microwave SSM/I (Special Sensor Microwave Imager) sensors mounted on the US Defense Meteorological Satellite Program (DMSP) satellites, physiographic and atmospheric data by an artificial neural network called Modified Counter Propagation Network (MCPN), a Projection Pursuit Regression (PPR) and a nonlinear regression. The airborne gamma-ray measurements of SWE for 1989 and 1997 were used as observed SWE, and SSM/I data of 19 and 37 GHz frequencies, in both horizontal and vertical polarization, were used for the calibration (1989 data from DMSP-F8) and validation (1997 data from DMSP-F10 and F13 of both ascending and descending overpass times were combined) of the models. The SSM/I data were screened for the presence of wet snow, large water bodies like lakes and rivers, and depth-hoar. The MCPN model produced encouraging results in both calibration and validation stages ( R 2 was about 0.9 for both calibration ( C) and validation ( V)), better than PPR ( R 2 was 0.86 for C and 0.62 for V), which in turn was better than the multivariate nonlinear regression at the calibration stage ( R 2 was 0.78 for C and 0.71 for V). MCPN is probably better than the linear and nonlinear regression counterparts because of its parallel computing structure resulted from neurons interconnected by a parallel network and its ability to learn and generalize information from complex relationships such as the SWE-SSM/I or other relationships encountered in geosciences.