| The thesis concerns multiple sensor processing and the application of an image processing technique to remote sensing data. Considered are radiometric images of land surfaces in the northern Great Plains generated at different (center) frequencies by the Scanning Multichannel Microwave Radiometer (SMMR) of the Nimbus-7 satellite. Through comparisons of these multispectral images, surfaces are segmented into regions and the terrain of each region is classified as frozen or thawed. Because all frequency channels share a common aperture, the images generated at different frequencies are of different resolution. To avoid mis-classification, images at each frequency are processed to compensate for resolution differences prior to classification. With no a priori scene information available, resolution compensation is performed by synthesizing images at all frequencies at the resolution of the lowest frequency (coarsest) image used in classification. Thus, fine-resolution information is lost.;The thesis presents a technique, derived from scale-space filtering of computer vision, to recover fine-resolution information in surface classification. Specifically, if the spatial transfer functions of the sensor is approximately Gaussian at all frequencies, then scale-space techniques can be used to estimate boundaries between different surface regions at fine-resolution. First, surfaces are segmented and classified using multispectral processing at coarse-resolution. Second, boundaries between different surface regions, as derived from multispectral processing, are registered to boundaries on high-frequency images that have been resolution compensated (i.e., coarse-resolution images derived from high-frequency, fine-resolution images). Third, boundaries on compensated, high-frequency images are tracked from coarse-to-fine resolution as resolution compensation is reduced. Estimates of surface boundaries are the resulting boundary locations in uncompensated, fine-resolution images.;The effects of using sensors with non-Gaussian spatial transfer functions are analyzed, and signal models are described for which this resolution recovery process performs exactly. In addition, the Freeze Indicator (FI) is presented as a parameter for the classification of frozen terrain. |
