Using Fisher’s linear classification procedure, a particular sample, say k, is then assigned to the first group if a’ k is closer to the mean of z1, or to the second group if a’ k is closer to the mean of z2. The procedure described above is extended discriminate between 12 land cover classes using nine independent polar metric radar measurements. The training data is then re-arranged according to polarization and radar frequency and same technique was applied to the new data. This is done in order to determine the significance of radar polarimetry and frequency in land cover feature detection. The results are then compared with that obtained using the “conventional” maximum likelihood algorithm.

3.Results and discussion of results

3.1 Discriminate analysis of all variables
A forward stepwise selection of input variables was performed in order to determine which variables contribute significantly to group separation as well as to eliminate variables that are redundant. However, it was found that all nine variable contributed significantly to group separation. This may be due to the fact that the different polarizations and radar frequencies are important in the detection of different land cover types, such as L-band for agricultural cover, P-band for soil cover and cross-polarized L-or P-band for forest cover. The variables PHH, CHH, CVV, LHH, CHV, PVV, PHV, LHV and LVV were entered in that order. The standardized coefficients of the discriminant functions are shown in table 1. the contribution of a variable to group separation may be inferred from the magnitude of the coefficients. The first function, which accounts for 70 per cent of group separation is dominated by CHH, PHH, CVV and CHV. The second function, which accounts for 8 per cent of group separation, is dominated by PHH, LHH, LHV and PVV. The first three discriminate functions already account for 91 per cent of the group separation. In the first three discriminate functions, all variables, except for LHH, contribute significantly to group separation. LHH actually dominates the higher-numbered functions.

Table 1 Standardized coefficients of derived discriminate functions

	f1	f2	f3	f4	f5	f6	f7	f8	f9
CHH	0.42	-0.09	-0.30	-0.62	-0.06	0.60	0.05	0.14	-0.51
CHV	0.25	.1	-0.09	-0.50	-0.26	-0.64	-0.12	-0.19	0.62
CVV	0.32	0.39	0.13	0.40	-0.05	0.53	0.00	-0.54	0.28
LHH	0.14	-0.03	-0.55	0.68	0.58	-0.39	0.19	0.04	-0.14
LHV	0.04	-0.15	-0.45	0.51	-0.59	0.06	-1.15	0.55	0.01
LVV	0.10	0.09	-0.02	0.08	-0.07	0.21	0.44	0.74	0.44
PHH	0.32	-0.23	0.66	0.19	-0.37	-0.23	0.97	-0.13	-0.74
PHV	0.4	-0.85	0.32	-0.35	0.60	0.14	0.05	- 0.38	0.65
PVV	0.05	0.51	045	-0.20	0.62	-0.03	-0.33	0.39	-0.28
Discriminating power	60.63	11.19	6.83	4.28	1.78	1.14	0.26	0.26	0.01
Per cent	70.40	12.99	7.94	4.97	2.02	1.33	0.30	0.04	0.01
Cumulative per cent	70.40	83.39	91.33	96.30	98.32	99.65	99.95	99.99	100.0

The image defined by the first three discriminate functions are shown in figures 1a to 1c. In the first discriminate function, the 12 land cover classes appear distinctly. The water areas are clearly differentiated from the non-water areas. Wet and dry fish ponds are also differentiable. Urban areas appear bright in contrast with other land features which appear in several gray tones. Mature and young rice paddies are shown in bright and dark patches respectively into upper left portion. The lower right portion of the image represents the forested forested areas in Guimaras province which have a textured appearance. The second discriminate function is observed. The image has a coarse appearance which obscures the water-land boundary. The land cover features are no longer easily detectable. Nevertheless, dark patches appear in the upper left quadrant of the image representing rice paddies, grassland and bare soil. The third discriminate function is again dominated by negative loadings. Other than the water body, urban areas appear very bright in contract with the other land cover types. The individual fish ponds are still recognizable unlike other linear features such as roads, the outline of rice paddies and the river.

Fisher’s linear classification procedure is applied to the test data. The classification procedure yielded an overall accuracy rate of 84.56 per cent, or kappa coefficient of 0.83 . The result is shown in table 2. the classified image is shown in figure 1d. A high accuracy rate (>85 per cent) was achieved for the classification of rough water surface, mature and young rice, other built-up areas, grass, mangrove and bare soil. Most calm water pixels were classification as rough water and wet fish pond while pixels corresponding to residential areas were classified as other built-up and forested areas. The former may be due to variations in dielectric properties in the individual pixels making up the training areas for calm water while the latter may be due to variations in texture in the individual and neighboring pixels making up the training areas for residential areas.
Table 2 Result of the classification of the test data

	RW	CW	CD	DF	MR	YR	OB	RA	F	C	M	BS
LDA	150	71	124	146	137	145	128	92	113	133	136	147	1522
C	67	87	122	113	126	147	124	83	68	95	81	137	1250
L	150	7	125	96	53	98	38	74	53	115	110	118	1037
P	63	0	56	121	104	70	60	93	108	133	129	131	1068
HH	146	23	129	139	101	139	39	65	45	134	138	141	1229
HV	150	1	53	106	110	95	45	28	43	124	100	142	997
VV	44	101	34	106	64	105	130	82	82	85	38	70	941
MLD	131	128	119	145	142	122	136	81	138	120	126	138	1526

LDA = linear discriminate analysis; C = C-band-grouped; L = L-band-grouped; P=P-band-grouped; HH=HH-polarization –grouped; HV = HV-polarization-grouped; VV=VV-polarization-grouped; MLA = maximum likelihood algorithm; RW = rough water; CW = Calm water, WF = wet fish pond; DF = dry fish pond; MR = mature; YR = young; OB = other build-up areas; RA = residential areas; F = forest; G= grass; M= mangrove; BS = bare soil

3.2 Discriminate analysis of grouped variable
Table 2 likewise shows the result of the classification using only a subset of the variables grouped according to polarization and radar frequency. The most obvious observation is that the number of correctly classified resulting from the discriminate analysis of the grouped variables are considerably les than when all variable are considered. This highlights the importance of polarimetry in land cover discrimination.