)86b.
. Na-
COMPARISON OF TEXTURE ANALYSIS TECHNIQUES
IN BOTH FREQUENCY AND SPATIAL DOMAINS
FOR CLOUD FEATURE EXTRACTION
ges Nahid Khazenie!?
es in
m of University Corporation for Atmospheric Research,
ages Boulder, CO 80301
l| get
ig to Kim Richardson?
?Naval Research Laboratory, Monterey, CA 93943-5006
ABSTRACT
Identification of cloud types through cloud classification using satellite observations is yet to produce consistent
and dependable results. Cloud types are too varied in their geophysical parameters, as measured by satellite
remote sensing instruments, to provide for a direct accurate classification. To aid in classification, texture
measures are additionally employed. These measures characterize local spectral variations in images. They are
widely used for image segmentation, classification, and edge detection. Numerous methods have been developed
to extract textural features from an image on the basis of spatial and spectral properties of the image. In our effort,
several of these methods are studied for their applicability in cloud classification and cloud feature identification.
The examined texture methods include a) spatial gray-level co-occurrence matrices, b) gray-level difference vector
method, and c) a class of filters known as Gabor transforms. Methods a) and b) are spatial and statistical while
method c) is in the frequency domain. A series of comparative tests have been performed applying these methods
to NOAA-AVHRR satellite data. A discussion as to the suitability of these texture methods for cloud classification
concludes this study.
Key Words: texture analysis, cloud classification, Gabor transforms, spatial gray-level co-occurrence matrices,
gray-level difference vector (GLDV), NOAA-AVHRR.
INTRODUCTION
Identification of cloud types by automated cloud classifiers,
which operate on a pixel by pixel basis, has yet to show
dependable and accurate results. Clouds have geophysical
parameters which are too inconsistent, as measured by satellite
remote sensing instruments, to provide for a direct accurate
classification. No method developed to date provides a
reliable spectral signature which would uniquely identify a
specific cloud type anywhere on the Earth globe during any
season. Cloud types vary in their spectral response at different
latitudinal locations and at different times of the year. These
variations complicate methods required for cloud type
identification using remote sensing techniques.
Surveying the various available statistical, structural, and
frequency domain techniques applied to cloud classification, it
appears that there are not enough parametrization vectors to
uniquely separate any one cloud type. For this reason, texture
analysis methods are drawn upon in addition to aid in this
problem. The use of texture parameters has been reported on
extensively in recent literature (Wechsler, 1980). Texture
techniques used in our study include a) spatial gray-level co-
occurrence matrices, b) gray-level difference vector (GLDV)
method, and c) a class of filters known as Gabor transforms.
Each of these approaches has unique merit for providing
additional information about cloud masses within a scene.
These unique differences are the focus in this study.
Images in this case study are composites of Advanced Very
High Resolution Radiometer (AVHRR) channel one and
channel four. Pixel by pixel classifications of cloud types,
based on the spectral and spatial responses from these
channels, are enhanced with results from the various texture
analysis algorithms. Results of the classifications from the
combined techniques are compared and discussed.
1009
DATA
An image from the Gulf of Alaska was chosen for this work.
This region was selected due to its high latitude which presents
challenging solar zenith angles. It also provides snow within
the scene which tests snow and cloud separation capabilities of
the candidate methods. Furthermore, the general
meteorological activity within this region is high thereby
presenting a continuous varying source of frontal cloud
masses.
The scene selected for presentation is one of eight images used
in this study. It is an AVHRR image from 15 October 1988,
19Z. A full resolution (1.1 km per pixel) sector of 1024 by
1024 ten-bit pixels was extracted from the original 2048 by
2048 data set.
The channel one and channel four radiance images are shown
in Figures 1 and 2. The channel one image is histogram-
equalized for purposes of display. The channel four image is
inverted so as to represent clouds in lighter gray shades.
The large band of clouds in the extreme right of the image is a
frontal cloud mass that has previously moved through the area.
This cloud mass is characterized by high thick cirrus over
cumulus. These clouds are brightened by their height as well
as by the low sun angle which is characteristic for this
northern latitude.
In the lower central portion of the image are well defined cloud
streets. They are trailed by open cell stratocumulus and
altostratus that extend to the left center of the image. The
mixed layered cloud mass in the lower left portion of the image
represents stratus and altostratus with a cover of thick cirrus.
Some closed cell stratocumulus are at the bottom of the image
between the stratus and frontal clouds.