Abstract
Polarimetric synthetic aperture radar over the coastal province of Iioilo in the Philippines was acquired by the NASA AIRSAR mission in November 1996. For the study, an image of nine bands, consisting of C-, L- band frequencies at HH, VV and HV polarizations were considered for analysis. An image transformation method was applied to an 800 pixel x 700 line subset of the image for subsequent identification of coastal land cover features. This method, discriminate analysis, was a multivariate statistical technique which maximizes the separation of groups based on a set of measurements. The transformed image was classified and then compared with the result obtained by using the ‘conventional’ maximum likelihood classifier. A classification accuracy rate of 84.56% was achieved using discriminate analysis and 84.78% using the maximum likelihood algorithm alone. The results showed that there was no significant difference in the performance of discriminate analysis and maximum likelihood algorithm in the identification of land cover features.

Introduction
Experience with single band synthetic aperture radar (SAR) data tells us that, despite knowledge of radar backscatter mechanisms involving the radar beam and natural objects on the ground, performing land cover classification is difficult, to say the least. In order to overcome this inherent difficulty, literature suggests that we do integration with optical data, analysis of multi-data SR imagery and derivation of other parameters such as texture and fractal dimension in order to achieve an acceptable rate of classification accuracy. The advent of polarimetric synthetic aperture radar (SAR) presented an opportunity to ground targets using three radar frequencies and four polarizations. In late 1996, the National Aeronautics and Space Administration of the United States deployed the airborne SAR imagery have been acquired for several test sites and are now being utilized for various applications including geologic mapping, volcanic studies, land cover mapping, rice crop monitoring and mineral exploration. This paper reports on our experience in analyzing AIRSAR data for the identification of land cover types in a small coastal area in a rapidly-developing province in the central part of the Philippines called Iioilo.

The general objective of the study is to evaluate a multivariate statistical technique called discriminate analysis in terms of its ability to “discriminate” between several pre-determined land cover classes. In particular, the study seeks to (a) determine which individual radar bands make up the land cover discrimination model and how much each individual band contributes to the model: (b) do the same using radar bands grouped According to frequency and polarization; and 9c) compare the results of discriminate analysis with that obtained using the “conventional “ maximum likelihood algorithm.

2. Methods

2.1 Data preparation
An AIRSAR strip has a size of about 10 km by 60 km. From this strip, an 8 km by 7 km sub-scene was extracted. This sub-scene covered the coastal areas in the provinces of Iiolio and Guimaras, separated by the Iioilo-Guimaras strait. The digital numbers were convered to dBs, after which a 5 x 5 Lee filter was applied. Rectification was done using only in the x and y directions since the study area was relatively flat anyway. From this sub-scene, 12 land cover classes were identified. These were calm water surface, rough water surface, wet fish ponds, dry fish ponds, newly-planted rice, mature rice, residential areas other build-up areas, grasslands, mangroves, forested areas and bare soil. Training patches corresponding to each of the land cover types were defined on the image. After that, 300 points were randomly selected from each class. This was done in order to ensure that the effects of spatial autocorrelation and other effects associated with a pixel-based classification are reduced. Half of the number of points were used to train the classifiers while the remaining half were used as test data. The study assumes “reciprocity”, that is HV-VH, and thus, only nine of the 12 bands are used in the analysis.

2.2 Data analysis
Discriminate analysis is an image transform which has roots in statistics. It is performed in order to describe the inherent separation existing between two or more variable, in this case, land cover types, based on a set of observations (radar backscatter values). It does this by computing a number of discriminate functions equivalent to the number of variables. These functions are essentially are that the data follows a multivariate normal distribution and that the variance-covariance matrices for the number of classes are statistically equivalent. However, violations of these assumptions do not lead to “fatal” results.

The procedure for the two-group case is described. We are given samples x1, x2, … xn and y1,y2…, yn. with group mean vectors of m1 and m2 respectively and a population covariance S. Each vector (x1 and y1) consists of measurements on a given number, say p, of variables. The discriminate function is the linear combination of these p variables that maximizes the distance between the two transformed group mean vectors. These functions take on the from:

Z1 = a’x = a1x1 + a2 x2 + ….. + anxn
Z2 = a’x = a1y1 + a2 y2 + ….. + anyn

Discriminate analysis is concerned with finding the vector a that will maximize the standard difference between the means of z1 and z2. The maximum occurs when :

a = S-1 ( x- y)