HR: 0830h
AN: H51D-1119 [PDF]
TI: Supervised Classification of Surficial Geology Units Using Local Statistics from Digital Orthophoto
Quadrangles as Input into an Artificial Neural Network
AU: * Phelps, G A
EM: gphelps@usgs.gov
AF: U.S. Geological Survey, 345 Middlefield Rd MS 989, Menlo Park, CA 94025 United States
AU: Miller, D M
EM: dmiller@usgs.gov
AF: U.S. Geological Survey, 345 Middlefield Rd MS 989, Menlo Park, CA 94025 United States
AB:
In the arid southwestern United States a U.S. Geological Survey project is underway to create 100,000 scale surficial
geologic maps. Surficial geologic deposits are classified according to process of deposition using geomorphic and
sedimentologic features. Geomorphic position, surface roughness, pavement maturity, pedogenesis, and inset relations are used
to establish relative deposit ages. Geologic mapping is conducted using field methods and interpretation of remote-sensing
images, primarily stereo-pair aerial photographs viewed through a stereoscope. Recently the project has been investigating
methods of classification based on remote sensing data to improve the mapping process. A supervised classification method has
been developed to distinguish alluvial geomorphic units ranging in age from active to Pleistocene. The remote sensing
technique makes use of spatially autocorrelated information contained in digital orthophoto quadrangles (DOQs) and the
classification power of a multilayer feedforward artificial neural network. The DOQs are processed by examining local
statistics of the DOQ gray-scale values in moving-window neighborhoods across the image (mean, median, range, and standard
deviation) as a function of distance for each point on the image. For example, at a given point the mean is examined in a 3x3
square neighborhood, a 5x5 square neighborhood, and so on. An empirical function of the mean vs. the window length is then
generated. The shapes of the resulting functions for points within a given geomorphic unit, when considered together, are
distinct for each generalized geomorphic unit. Each unit is represented not by a single function, however, but rather by a
family of functions. The family of functions generated is complex, yet almost every function belonging to a given geomorphic
unit is at least 5% different, most much more so, than those belonging to any other unit. A sample set of functions, 1000
for each of three geomorphic units, was used to train a three-layer multilayer feedforward network. Multilayer feedforward
networks are reasonably immune to noise and correlation of input variables, and so are ideal for distinguishing inputs that
interact in complex ways. The method produced results consistent with recent mapping. The method was further tested by
applying it to a test area, roughly 20 km southwest of the training area. This test also produced results consistent with
recent mapping. The method has several advantages compared to training using other common remote sensing data sets: (1) it
uses 1-m resolution DOQs, so classifies at higher resolution than possible with many data sets; (2) it examines neighbors to
classify pixels, thus honoring spatial patterns, an advantage for mapping spatially-dependent features such as geomorphology
and vegetation, and (3) it allows for both mixed classification and rejection of classification of the given classes on an
individual pixel level, which allows for the identification of mixed geomorphic units and new (unclassified) geomorphic
units.
DE: 1625 Geomorphology and weathering (1824, 1886)
SC: Hydrology [H]
MN: 2003 Fall Meeting