variance in spectral density but those in the latter very small 
) ; Table 2 Categories used in the processing. 
one. ltis known that larger variance category often appears 
  
  
on smaller variance category areas in results processed by No. Category No. Category 
MLH. In order to evaluate the performance of MLDF for oT = g 
such images, we apply the three methods to a coastal region : Pins (medium) 4 18 : Zin S01 (pink) 
LANDSAT/TM image having 256 columns, 256 lines and 3 2 : Bare soil (light) | 9 : Pine (dark) 
spectral bands, and compare their results. Figure 6 shows 3 : Field (radish) 10 : Swimming pool 
original image (a), result processed by MLH (b), one by BDT 4 Rice field 11 | Zinc roof (green) 
(c) and by MLDF (d). In these processes, we use 9 5 OIN TO 12 : Pine (light) 
categories; urban area, sea, vegetation, river, and others. T : e 
Processing times are 23 seconds by MLH, 3 seconds by 6 : Field (cabege) | 13 : Bare soil (dark) 
BDT and 5 seconds by MLDF. Several pixels in river are 7 ‘Slate : 
  
  
  
  
miss-classified into urban area in the results processed by 
MLH and by BDT, but MLDF does not yield such error. 
AGRICULTURAL REGION 
  
€ ERE x n ios 
Fig. 7 Training areas (a) and test areas (b). 
numerical evaluation. Figure 8 shows original image (a), 
result processed by MLH (b), one by BDT (c) and by MLDF 
(d). Result of numerical evaluation of accuracy is shown in 
Table 3, where upper, middle and lower lines indicate correct 
classification rates in MLH, BDT and MLDF, respectively. 
Mean correct classification rate (MCCR) and processing time 
are also listed in Table 4 with results for three spectral band 
image for the same region. We chose spectral bands 3, 4 
  
(c) (d) 
Fig. 6 Original Kawasaki image (a), result processed by 
MLH (b), one by BDT (c) and by MLDF (d). 
We have had a data set acquired on Dec. 1977 at Fukue, 
Aichi prefecture in Japan. The data set consists of airborne 
MSS images and ground truth. We applied the three 
methods to the data set. We use 512 x 512 pixel image 
with 5 spectral bands whose wavelength regions are shown 
in Table 1. Table 2 indicates 13 categories we used, and 
Fig. 7 indicates training areas (a) and test areas (b) for 
Table 1 Observation band of multispectral image. 
  
  
  
Band No. : Wavelength [um] 
DT i035 0A 
i 047 - 040 
3 0.54 - 0.56 
4 0.66 - 0.68 Fig. 8 Original Fukue image (a), result processed by MLH 
5 0.80 - 0.90 (b), one by BDT (c) and by MLDF (d). 
  
  
330 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996 
  
  
ER 
Area 
  
  
and 5 
using 
that IV 
to rec 
that p 
numb 
depei 
inform