International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
   
  
  
  
  
  
  
  
  
TM Band-5 | TM Band-4 | TM Band-3 | SPOT 
Entropy (bits) 6.48 6.01 5.00 6.17 
Information amount (bits) 4.38 3.40 1.60 3.44 
; Average code length 6.51 6.04 3.97 6.20 
Huffnan Coding, | conpressionate 1.23 1.32 2.01 1.29 
Uni-directional Average code length 5,22 4.85 3.07 4.50 
differential coding Compression rate 1.53 1.65 2.01 1.78 
Bidirectional Average code length 5.12 4.77 3.88 3.66 
differential coding Compression rate 1.56 1.68 2.06 2.19 
  
  
Table 2. Comparisons between Huffman coding and differential coding 
  
ce 
© 
eS. 
u» 
> 
cs A LL 
557 | | 
E i B Std-2305| 
o i i 
- 
& 
ce 
& .l 
e 
> 
o - infame À— À— 
  
  
40 80 80 100 120 140 
Grayscale 
Figure 3. Grayscale histogram of the initial image 
  
  
  
  
  
  
  
© 
e 
& L 
© 
> 
os 
ss i 
38 7 m Std.=3.25 
u 
ce 
e 
i. | 
1 
ys 
eu Lb. 
40 20 0 20 40 
Grayscale 
Figure 4. Grayscale histogram of the differential image 
3.2 District Forecast Differential Coding 
Four types of representative land covers are selected and shown 
in Figure 5. All of them are SPOT-5 panchromatic images with 
a 2.5 meter spatial resolution, consisting of 1024 by 1024 pixels. 
Differential encoding algorithm was applied to these four 
images, with results shown in Table 3. The results accord with 
what is expected, i.e., the compression rates of the images are in 
reverse proportion to the information amounts. The land covers 
with larger information amounts, e.g., urban areas and farm 
lands, would achieve smaller compression rates, and vice versa 
(e.g., mountain areas and water bodies). Therefore, according to 
the negative correlation between information amount and 
compression rate, a district forecast differential encoding 
method is proposed to further improve the compression rates of 
remote sensing images. 
First, images may be categoried by the texture complexities, 
e.g., built-up areas, farmlands, forests, deserts, water, etc. For 
different land cover types, differential encoding algorithm is 
applied. And the more effective one between unidirectional or 
bidirectional differential encoding is selected for each land 
cover type according to the compression efficiencies. At last, 
the corresponding compression modules would be built for each 
type. 
Moreover, due to the definitive orbits of satellites, land cover 
types inherent to remote sensing images are relatively certain. 
As is mentioned above, the information amounts inherent to 
images are different because of different land cover types, 
which results in different compression rates. Hence, historic 
remote sensing images may be firstly collected to judge the land 
cover types when satellite flying over certain districts. In other 
words, the orbit is divided into districts according to land cover 
types. As the satellite flying over oceans, the differential 
encoding module for this type may be executed to compress the 
image data. Similarly, the differential encoding module for 
forest may be called when the satellite flying over a forested 
area. In this way, the differential coding algorithms may be 
applied to images according to different districts so as to obtain 
higher compression rates and sufficiently and effectively usage 
of the limited storage space on-board. 
3.3 Near Lossless Compression 
To testify, the operator is applied to the SPOT-5 image in 
Figure 2(b) one time, two times and three times, respectively. 
The result images are shown in Figure 6, and the corresponding 
compression results of Huffman coding and differential 
encoding are listed in Table 4, where the times of edge 
protection smoothing set to naught equals to directly applying 
the encoding algorithms to the original image. 
According to Table 4, with the times of edge protection 
smoothing increasing, the compression rates of Huffman coding 
are seldom improved. However, significant improvements in 
both the average code length and compression rate may be 
achieved after the application of differential coding. Moreover, 
considering different uncertainties of different land cover types, 
the same comparative experiments are applied to the four 
images shown in Figure 5, the results of which are listed in 
Table 5. It is obvious that different land cover types will obtain 
different compression rates. Furthermore, the lower the 
information amount is, the higher the compression rates would 
be achieved.