BBTGA Progressing

11/4/99 Finish implementation of Bayesian BTGA (BBTGA). Use a serial of proportions of scene 33 to train and use remained pixels to test. Table 1 shows the results. Each training data set and testing data set pair is used for training and testing for three times. The listed date are the average.


%	0.5	1	5	10	20	50
Training size	168502	337992
Testing size	33615853	33446363
Tree size	8034	16165
N_impurity	449	869
P_impurity	124626	250404
N_size	3568	7213
P_size	43876	87588
Time(Hour)	0.266	0.569
A_training	88.3	88.5
A_testing	79.9	80.1

Table 1. Test on the newly implemented BBTGA. %: the proportion of scene 33 data used for training. Training size: the training data set size. Testing size: testing data set size. Tree size: the size of the generated tree. N_impurity: the terminal node number from criterion of impurity reduction. P_impurity: the population of training data in the terminal node number from criterion of impurity reduction. N_size: the terminal node number from criterion of terminal node size. P_size: the population of training data in the terminal node number from criterion of terminal node size. Time(Hour): the time needed for training. A_training: the classification accuracy when applied the tree on the training data set. A_testing:the classification accuracy when applied the tree on the testing data set.

11/4/99 Test BBTGA on ground observation data set. The training data is saved in file "/usr/mvl2/feng/data/ground.bin.char.data". Training data set size is 2700, testing data set is from 99.5% of the scene 33 and the size is 33615853. The test is repeated for three times and Table 2 shows the result.

	1	2	3	Average
Tree size	115	147	113	125
N_impurity	9	12	6	9
P_impurity	2067	1981	1991	2013
N_size	49	62	51	54
P_size	633	719	709	687
Time (second)	7	9	7	8
A_training	90.41	90.67	89.78	90.29
A_testing	72.89	72.03	73.55	72.82

Table 2. Test on the newly implemented BBTGA with ground observation. N_impurity: the terminal node number from criterion of impurity reduction. P_impurity: the population of training data in the terminal node number from criterion of impurity reduction. N_size: the terminal node number from criterion of terminal node size. P_size: the population of training data in the terminal node number from criterion of terminal node size. Time(Hour): the time needed for training. A_training: the classification accuracy when applied the tree on the training data set. A_testing:the classification accuracy when applied the tree on the testing data set.