In: Wagner W., Szekely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B
550
1) The same as steps 1~3 in training;
2) Detect linear targets with fusion operate of D1 and D2
in testing images;
3) Utilize AdaBoost classifier in training stagey to get
Pu\
4) Construct MRF model as Eq.l and get optimal
solution K(0);
5) Get Pi.TPi with ^-th iterative solution Y (t) ;
6) Construct MRF model as Eq.4 and get optimal
solution
7) Repeat steps 4 and 5 until little changes existing in
y(*+i) #
4. EXPERIMENTS
4.1 Experiments setup
Experiments are done on SAR image datasets. The datasets and
parameters are illustrated as following.
4.1.1 Data: The SAR datasets contains a 1500 X 1200 pixels
image that are selected from VV polarization SAR images of
Guangdong Provinces of China in May 2008 of TerraSAR
satellite. The spatial resolution is 1,25m* 1,25m. Each image of
the SAR datasets has a ground truth getting from manual
labeling under ArcGIS software. Our experiments consist of 4
classed: farmland, woodland, building, water. Half of this
image is used for training, the remaining for testing.
4.1.2 Parameters: In linear target detection, the template is
selected with 15 pixels high, 13 pixels width and 3 pixels centre
region. The threshold of Dl, D2 and fusion operate are 0.35
0.45 and 0.35 individually. The minimum region area of
superpixels in Meanshift based over-segmentation is 400 pixels,
with spatial bandwidth and range bandwidth are both 3 pixels.
Features used here are gray histogram and SoftLBP 0. The
length of sub-lines is 50 pixels and the width of sub-line regions
is 20 pixels.
4.2 Classification Performance
The classification results of the proposed approach in this paper
are shown in fig.7. Fig.7.c is the beginning of iteration result
where Pj,TPi = 0, that is without LTP. And there are some
isolated points in the classification map. Moreover, there are
many indented edges along the linear targets. In the fig.7.d, e
and f, isolated points and indented edges decrease gradually
since the addition of LTP.
Compared with groundtruth data labeled artificially,
classification accuracies are listed in table. 1. It shows that the
average accuracy has been improved only a little from ieration-
0 to ieration-3, but the overall classification performance has
large improvement.
Fig. 7. Experimental Results, (a) original image; (b) groundtruth data with linear targets detected with fusion operate of Dl and D2 operates; (c)
classification results in iteration 0 (without LTP); (d) classification results in iteration 1 (with LTP); (e) classification results in iteration 1 (with
LTP); (f) classification results in iteration 1 (with LTP);