Many sensor systems are available for sensing the earth surface from satellites as well as airborne and mobile
platforms. Thus, fusing data from multiple sensors is becoming a common theme in earth remote sensing. A major
goal of remote sensing image fusion is resolution enhancement. In this paper, optimization techniques are presented
and discussed in order to help make an image fusion process a practical method for not only spectral signature based
image analysis but also for algorithm development in remote sensing of water. The technique described and
explored in this paper includes the identification of feature areas, stratified random pixel selection, singular value
decomposition model building for synthetic image generation, and optimization of the 2D Butterworth filter cutoff
and order coefficients in a spectral and spatial resolution enhancement protocol. The process is also called spatial
sharpening of hyperspectral imagery as presented in this paper. Examples of methods for estimating errors in the
data fusion process are also described using coastal littoral zone remote sensing imagery with an emphasis on
weathered oil scenes. The optimization and testing of a data fusion methodology or protocol described utilizes image
to image georeferencing methods, nearest neighborhood and linear remapping of multi-resolution spatial and
spectral imagery. The central optimization procedures entails random selection of pixels from feature areas in
simultaneously acquired multispectral and hyperspectral scenes in order to build multiple "SVD" singular value
decomposition models and optimized selection of these image models for each hyperspectral channel based upon the
non-parametric K-S p-statistical test. The model synthetic imagery is then used with the 2D discrete cosine and
inverse cosine filters, a 2D Butterworth filter. Optimization of the 2D Butterworth filter cutoff and order coefficients
are conducted for each hyperspectral band and these coefficients are optimized using the same K-S based tests. The
above optimization protocol results in synthetic reflectance hyperspectral cube where minimization between
observed and synthetic hyperspectral signatures has been performed for each hyperspectral channel. Results indicate
the synthetic hyperspectral resolution enhancement methodology is most sensitive to (a) the pixels selected (from
feature areas) for use in the SVD model building process and (b) the 2D Butterworth cutoff frequency selected.
|