220.

[Cover] [Contents] [Index]

Page 296

Table 7.6 Confusion matrices generated by the extension of Bayesian theory, evidential reasoning and iterative conditional mode algorithms

No.	1	2	3	4	5	6	7	8
(a) The extension of Bayesian theory
1	280	2548	1173	160	8	1	5	9
2	37	14730	700	1256	62	2	102	26
3	59	3464	3056	237	22	1	34	9
4	9	10432	340	4960	205	0	64	1
5	5	5912	107	1212	1429	104	70	1
6	8	850	39	70	269	595	78	0
7	0	1103	28	31	2	0	2176	0
8	31	4457	364	231	1	0	0	339
(b) Evidential reasoning
1	241	2395	1114	143	II	0	5	7
2	19	11934	535	933	44	2	75	14
3	50	3436	2916	223	19	0	31	7
4	13	10666	385	4860	206	0	83	2
5	5	5996	139	1224	1351	89	49	1
6	43	1579	154	163	358	612	136	0
7	0	1348	39	66	0	0	2150	0
8	58	6142	525	545	9	0	0	354
(c) Iterative conditional mode algorithms
1	409	1247	784	25	0	0	0	0
2	2	19433	786	717	0	0	121	4
3	15	2743	3916	74	0	0	10	0
4	1	10985	111	5983	132	0	77	0
5	0	4697	5	1225	1693	118	65	0
6	1	446	0	25	172	585	105	0
7	0	782	0	0	1	0	2151	0
8	1	3163	205	108	0	0	0	381
(a) Average producer’s accuracy: 67.85%; kappa coefficient: 0.262 (b) Average producer’s accuracy: 65.63%; kappa coefficient: 0.233 (c) Average producer’s accuracy: 79.09%; kappa coefficient: 0.368.

autoregressive random field model (MAR) (Chapter 5). The window size for texture feature extraction is 5×5 pixels, and the neighbour support parameter for MAR is defined as {(1, 0), (−1, −1), (−1, 0)}. Three texture features, the mean value of neighbourhood vector θ (Chapter 5, Equation (5.48)), covariance (Equation (5.49)), and mean (Equation (5.50)) for each radar image are used as inputs.

Class-conditional probabilities were generated for each data source using Equation (2.23), which assumes a Gaussian probability density function. The maximum likelihood method was used to classify each data source individually. The average producer’s accuracy (Section 2.7) for the TM bands alone is 56.58%. Average producer’s accuracy is 50.22% for the SIR-C C band HH and HV images, while the average producer’s accuracy

[Cover] [Contents] [Index]