stanbul 2004 
n forestry 
)/. 
orthophoto 
phy data", 
, Slovenia, 
cedure for 
|] Airborne 
Sco, CA. 
and digital 
»photomap 
ie. Remote 
cisco, CA. 
ser locator 
Proc. 31d 
mation in 
I use of 
based on 
n, In Proc. 
w-Delhi. 
AN INTEGRATION OF WAVELET ANALYSIS AND NEURAL NETWORKS 
IN SYNTHETIC APERTURE RADAR IMAGE CLASSIFICATION* 
Qiming Qin* , Robert R. Gillies ^ . Rongjian Lu*, Shan Chen* 
* Institute of Remote Sensing and GIS, Peking University, China, 100871 gmgin(@pku. edu.cn 
? Department of Geography and Earth Resources, Department of Plants, Soils Biometeorology, Utah State University 
* 
Logan, UT 84322-4820 USA, voice: (435) 797 2664 email: rgilliesi@cnr.usu.cdu 
* Institute of Remote Sensing and GIS, Peking University, China, 100871 
KEY WORDS: Image Classification, Wavelet Analysis, Neural Networks, Texture and Structural Characteristics, SAR 
ABSTRACT: 
In this paper. the concepts of wavelet analysis and neural networks are applied to the classification of shuttle imaging radar 
experiment C (SIR-C) synthetic aperture radar (SAR) data from a location in northwest China. Initially, the paper presents the visual 
elements of tone, texture and structural features on SAR imagery as important bases for image classification and target recognition. 
The wavelet analysis is used as a method to extract elements of texture and structural features; it involves deriving the energy of 
sub-image blocks through wavelet decomposition. A improved backpropagation neural network was applied to a multiresolution 
representation of six images comprising reflectance SAR data and those obtained by the wavelet transform. A simple scene was 
classified, vielding poplar trees and bushes. Where they were well differentiated the probability of yielding the correct classification 
was found to be 100%. Erroneous classification occurred in transition areas between cover types where the percentage of correct 
classification fell slightly. The results suggest that such an integrated approach to classification is applicable for SAR data that 
involves regular textures and structures with rather strong orientation of land features. 
1. INTRODUCTION 
Maximum the often quoted advantage of radar, i.e. being 
unaffected by cloud cover, is paramount to its use in certain 
regions of the world. Radar images are however characterized 
by distinctive properties that are in part present due to an image 
formation process that is quite different from that of 
conventional optical images, i.c. SPOT or Landsat TM imagery. 
In principle. many of the conventional algorithms that are 
applied in the analysis of multispectral data can be applied in 
the analysis of Synthetic Aperture Radar (SAR) data (Simard et 
al. 2000). There are no theoretical limitations to the number of 
features (or bands) used by any of these algorithms. On the 
other hand, due to the nature of SAR data, transformation tools 
are often required to provide collateral information to assist in 
the process of image classification. 
Each homogencous region of a SAR image contains certain 
characteristics that are important bases for target recognition 
and image classification. The significance of certain image 
interpretation elements is particularly useful to establish 
coherent information set that permits a robust classification of a 
SAR image. This study utilizes three elements for SAR image 
interpretation - namely tonc, texture and pattern (i.e. combining 
elements of structure, orientation or direction). 
Tone refers to the relative brightness of the pixel elements and 
represents a qualitative measure of microwave backscatter 
strength. Tone on SAR imagery mainly relies upon the 
backwards scattering character of the terrain object. For natural 
objects, rough surfaces such as mountains and agricultural 
fields produce powerful backscattered returns and in doing so, 
form a variety of textural and structural features. Smooth 
surfaces such as calm water and flat land surfaces act as 
  
specular reflectors so that most of the energy is reflected away 
from the imaging SAR. Differentiating between such specular 
reflectors will generally depend more upon recognizing pattern. 
Texture from a SAR image is spatial information of the image 
tone variety repeated with a certain rule. That is to say it is the 
arrangement of tone and is manifested by an arrangement of 
variation in brightness. This variation. in brightness has an 
important intrinsic property as a frequency of tonal change and 
it is this that is particularly useful in the discrimination. of 
different areas of SAR illuminated areas. 
Pattern presented on a SAR image is the composition formed 
regularly by some spatial characters of target objects in an area. 
Such spatial arrangements of objects on the ground may be 
systematic or random. They may exhibit structure in space, e.g., 
they may lie north — south and be parallel or form more intricate 
patterns. Nevertheless orientation and direction, whether part of 
a systematic pattern or not, can serve to varying degrees as a 
useful basis for the interpretation of SAR images. 
The combination of such visual elements as a means to 
classification is by no means straightforward. For example, 
man-made target objects (i.c. urban) generally have regular 
geometrical character. However, the tones and textural features 
in the SAR imagery will vary along with antenna look direction. 
Brvan (Bryan 1979) found that different orientation angles 
formed between cultural targets and antenna look direction 
could produce dramatic differences in image gray tone: streets 
paralleling with radar track direction took on light lines tones, 
because the buildings alongside the streets play a major role in 
determining echo strength. On the other hand. the streets across 
track direction appeared as dark tone lines or had no 
presentation on the SAR image. 
This research is supported by the Special Funds for Major State Basic Research Project (Grant No: G2000077900) and supported 
by Chinese National Natural Science Funds (Grant No: 40071061).