This dissertation classifies trees, snow, and clouds using multiparameter millimeter-wave radar data at 35, 95, and 225 GHz. Classification techniques explored include feedforward multilayer perceptron neural networks trained with standard backpropagation, Gaussian and minimum distance statistical classifiers, and rule-based classifiers. Radar data products, serving as features for classification, are defined, radar and in situ data are presented, scattering phenomenology is discussed, and the effect of data biases are analyzed.; A neural network was able to discriminate between white pine trees and other broader-leaved trees with an accuracy of 97% using normalized Mueller matrix data at 225 GHz; wet, dry, melting, and freezing snow could be discriminated 89% of the time using 35, 95, and 225 GHz Mueller matrix data; and metamorphic and fresh snow could be differentiated 98% of the time using either the copolarized complex correlation coefficient or normalized radar cross section at three frequencies.; A neural network was also able to discriminate ice clouds from water clouds using vertical and horizontal 95 GHz airborne reflectivity measurements with a success rate of 82% and 86% when viewing the clouds from the side and below respectively. Using 33 and 95 GHz data collected from the ground, a neural net was able to discriminate between ice clouds, liquid clouds, mixed phase clouds, rain, and insects 95% of the time using linear depolarization ratio, velocity, and range. As a precursor to this classification, a rule-based classifier was developed to label training pixels, since in situ data was not available for this particular data set. Attenuation biases in reflectivity were also removed with the aid of the rule-based classifier. A neural network using reflectivity in addition to other features was able to classify pixels correctly 96% of the time.
展开▼
