High Spatial Resolution Digital Imagery
69
⎧
⎫
nb
∑ ti ri
⎪
⎪
⎪
⎪
α = cos- 1⎨
i=1
1/ 2 ⎬
1/ 2
⎛ nb 2 ⎞ ⎛ nb 2 ⎞ ⎪
⎪
⎪ ⎜ ∑ ti ⎟ ⎜ ∑ ri ⎟ ⎪
⎩ ⎝ i=1 ⎠ ⎝ i=1 ⎠ ⎭
where
α = spectral angle
ti = spectral vector of unclassified pixel
ri = mean spectral vector of training (reference) signature
nb = number of bands
The spectral angle is then calculated 37 times for each image pixel (ti) be-
cause there are 37 ri reference signatures or classes. The pixel is then assigned to
the class where the spectral angle is the smallest. If all 37 angular estimates are
greater than the maximum specified angle (i.e., 0.08 radians or 4.58 degrees), the
pixel is assigned to an "unclassified" group.
The results of the initial classification using the SAM algorithm showed
some confusion among the vegetation classes. Also, some pixels (approximately
10% in both mosaics) were not classified. Additional SAM runs using larger
spectral angles (0.090.12) successfully increased the number of pixels that were
classified. However, the confusion among classes was also increased. Therefore,
the results from the 0.08 spectral angle were retained for the final class maps.
5.3.3.2 Maximum-Likelihood Classifier. A maximum-likelihood classifier was
then applied to the unclassified image pixels. This algorithm uses the class-
conditional probability density functions to calculate the likelihood that a given
pixel, with its unique spectral vector (i.e., variable ti from the SAM equation),
belongs to each of the reference classes (i.e., variable ri). Every pixel is then as-
signed to the class with the maximum probability of membership. This classifi-
cation technique is used extensively with both multispectral and hyperspectral
imagery.
5.3.3.3 Minimum Mapping Unit Filter for Wetland Vegetation Mapping Over
Blackwater Wildlife Refuge. A minimum mapping unit (MMU) filtering routine
was also applied to the full thematic resolution class maps. This filtering
algorithm first delineates all of the raster polygons throughout the thematic
image. A raster polygon is defined as a group of adjacent (i.e., connected) pixels
with the same thematic class value. The adjacency criterion defines the polygon
using pixels joined along the four flat edges of the square and those pixels joined
Previous Page
