Jc. Sandidge et Rj. Holyer, COASTAL BATHYMETRY FROM HYPERSPECTRAL OBSERVATIONS OF WATER RADIANCE, Remote sensing of environment, 65(3), 1998, pp. 341-352
Water depth, bottom reflectance, inherent optical properties of the wa
ter column (scattering, absorption, and fluorescence), and illuminatio
n conditions combine to determine the upwelling spectral radiance of c
oastal waters. If these complex optical relationships could be quantif
ied, it would be possible to extract coastal information from spectral
radiance data. We use data from the Airborne Visible/Infrared Imaging
Spectrometer (AVIRIS) in a neural network system to establish quantit
ative, empirical relationships between one of these parameters, depth,
and remotely sensed spectral radiance. Data are analyzed for two area
s: the western coast of Florida in the Tampa Bay area and the Florida
Keys between Upper Matecumbe and Plantation Keys. The neural network a
pproach results in retrieval of reasonable depths from spectral radian
ce in both cases over a depth range of 0 to 6 m. Retrieved depths for
Tampa Bay are accurate to a RMS error of 0.84 m relative to depths in
the National Ocean Survey (NOS) Hydrographic Database, and the Keys re
trievals have an RMS error of 0.39 m relative to a bathymetric survey
conducted to support this study. A neural network trained on a combina
tion of the two data sets results in a combined RMS error of 0.48 m, n
early the same performance as neural networks trained individually. Th
e ability of the neural network to generalize, producing algorithms wi
th some degree of universality among diverse coastal environments is,
thereby, demonstrated. The result of the generalization analysis is of
practical importance because it indicates that the neural network may
not require an extensive training set of water depth data in order to
the ''tuned'' for each location where depth retrievals are desired. W
hile empirical the neural network is in some sense a model of the inve
rsion of the radiative transfer problem within the marine environment.
The neural network approach, therefore, operates on a higher level th
an more traditional statistical curve fitting solutions for retrieval
of remotely sensed information. Published by Elsevier Science Inc.