Page 151 - Digital Analysis of Remotely Sensed Imagery

120 Cha pte r F o u r

(ITT, 2007). Other special filters such as Sobel, Roberts, dilation, and
erosion are also available, along with adaptive filters such as Lee,
Frost, Gamma, and Kuan. Image texture is described using data
range, mean, variance, entropy, skewness, variance, homogeneity,
contrast, dissimilarity, entropy, and correlation. Filtering in the fre-
quency domain (e.g., fast Fourier transformation or FFT) or inverse
FFT can be easily carried out in ENVI, as well.
Although ENVI does not have any generic radiometric correction
functions, special tools are available for radiometric processing of
Landsat data, such as destriping, atmospheric correction, and calibra-
tion of reflectance using prelaunch parameters. A specific set of tools
are designed for displaying ephemeris data, radiometric calibration
and geometric rectification, and calculation of sea surface tempera-
ture from AVHRR data. ENVI also encompasses tools for calibrating
thermal infrared data to emissivity.

4.3.3 Image Classification and Feature Extraction
There is a comprehensive set of image classifiers in the Classification
toolbox of ENVI. Remotely sensed data may be classified unsuper-
vised (K-means and ISODATA) or supervised (parallelepiped, mini-
mum distance, maximum likelihood, and mahalanobis distance)
(Fig. 4.3). Moreover, ENVI offers three more nonconventional classi-
fiers: binary encoding, neural network, and Spectral Angle Mapper
(SAP). In the binary decision tree classification, pixels are grouped
into classes via a series of binary decisions in multiple stages. The
neural network classifier makes use of standard backpropagation for
supervised learning and a layered feed-forward network for classifi-
cation. Images may be classified at the subpixel level using the Sub-
pixel module. In the SAP spectral classifier, the spectra of input pixels
are compared to those of reference pixels. Their similarity is measured
by the angle between them. All classified results may be spatially fil-
tered during postclassification processing. Indices of classification
accuracy, such as confusion matrix and Kappa coefficient, can all be
generated in ENVI.
In addition to these classifiers, ENVI also contains image process-
ing functions, including anomaly detection, feature extraction, pan
sharpening, and vegetation suppression. Of particular notice is the
feature extraction tool, an object-oriented add-on module that is
designed to quickly, easily, and accurately extract features from high
resolution imagery. Its wizard makes use of both spectral and spatial
properties of pixels. Features are identified based on their spectral
and physical characteristics such as structure and shape. With the use
of such characteristic attributes, these features can be expected to be
classified more accurately. Moreover, features as small as buildings
and vehicles can be extracted from hyperspatial imagery using this
tool. Libraries containing the features of interest can be built over
time, making this tool quite valuable in automating the workflow.