The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part BL Beijing 2008
942
2.2 Method of image quality appraisal
Image quality appraisal qualitatively analyzed and described
“Beijing 1” image character by visual detection from some
aspects, such as texture structure, data continuity, noise
characteristic, radiation precision as well as interpretation
ability and so on.
2.3 Method of land use classification
Supervised and unsupervised classification methods were used
to extract the sample. First Initial classification was carried on
in order to get the training area. Simultaneously, "the same
thing with different spectrum, the different thing with same
spectrum" category and region were founded for revision
classification result. Then by selecting training sample and
classification function, supervised classification was made,
finally revised classification result by interpreter and made
classification precision appraisal. During the supervised
classification process, in order to contrast classification
precision, three classification functions were selected. The
maximum likelihood formula was as follows:
determining accurate reference category of the sampling point,
classification precision appraisal was finished. Appraisal results
showed in confusion matrix. Overall precision and Kappa
coefficient were extracted from the matrix. The value was
bigger, classification precision was higher. The detailed flow
was as follows:
[Micro-satellite Data Processing!
lUnsupervised
Sample Extract
Building Classification Machine
1
1
[Maximum Likelihood -|—1~Mahalanobis Distance h-l-Minimum Distance
Supervised Classification
[Accuracy Assessment I
D = ln( a e ) - [ 0 .5 ln(| Cov c |)] -
[ 0 .5 ( A - M C )T (Cov ; 1 )( X - M
)
Fig.l Image land use classification flow
in the formula, D : Weighting distance, C : Some
characteristic type, X : Element survey vector, M c :Sample
average vector of type C, Cl c : Percentage probability of any
element belonging to type C , CoV c : Element covariance
matrix of type C ,T : Transposition function. The minimum
range formula was as follows:
d ( x , M , ) = [ £ ( x k - m ik ) 2 ]
k = 1
in the formula: n : Wave band number, ^ :Some characteristic
t i / n M- j
wave band, 1 :Some cluster center, 1 : sample L mean,
Mtk . 7h e ith cen t r e and ^th jonj element value,
d(x,M„ , , ¡fL
1 : Distance of the element to the U L
kind
M ;
central( 1 ). The Mahalanobis distance formula was as
follows:
D = (X - M C )T (Cov ;')(! - M )
in the formula: D • Mahalanobis distance, c : Some specific
kind, X : Element survey vector, ^ c : Type c template
Cov r
average vector, c : Element covariance matrix in type
c template, T : Transposition function.
2.4 Method of Classification Precision appraisal
Using ERDAS software, classification precision appraisal was
finished. Stochastically 500 spots were extracted. After
3. RESULTS AND DISCUSSION
3.1 Visual analysis result
By visual detection, “Beijing 1” image spectrum information is
rich. The texture structure is clear. Escaping belt and leaking
belt phenomenon does not exist. Data continuity is good. There
is no partially geometry distortion. The forest land, the lawn,
the urban land, the countryside residential area, the reservoir
and the pit pond can be distinguished clearly from the image.
The mountain takes on seal brown (crag bare) and red, scarlet
red (covered by vegetation). Water body outline is obvious,
Because of different silt content, takes on different colour. The
tiny drainage ditch is green. The reservoir and pit pond wer
dark blue or black along with different depth. In the image
Beijing built-up urban area outline is clear, but the city street
texture is fuzzy. The whole interpret situation is good.
3.2 Spectrum information analysis result
In order to providing reference for land use computer automatic
classification research, Get terrain feature spectrum diagram of
curves (Figure 2), extracted different land spectrum information,
analyzed image spectral signatures. From the chart, we can see:
the water spectrum value in the wave band 1 was lower than the
spectrum value in the wave band 2 and 3. Farming land, forest
land and lawn were most obvious in wave band 1. The wave
band 1 was green vegetation's high echo area. The wave band 2
was vegetation's low echo area. It was sensitive to the plant
chlorophyll's absorption. Spectrum value of the Sand beach and
construction land was high and the spectrum information was
easy to get. But the urban land and the countryside residential
area's curve characteristic were similar, and the forest land and
the lawn curve type were similar. When making computer
automatic classification, it was not easy to differentiate the
information.