A COMPARATIVE STUDY ON THE METHODS FOR ESTIMATION OF MIXING RATIO WITHIN A PIXEL
Yasunori Terayama,
1 Honjo, Saga-si,
Yoko Ueda, Kohei Arai
(Dept. Information Science, Science and Engineering Faculty, Saga University)
Saga 840 Japan
and Masao Matsumoto
(Faculty of Computer Science and Systems Engineering. Kyushu Institute of Technology)
680-4 Kawazu, Iizuka-si,
ABSTRACT
Fukuoka 820 Japan
A comparative study has been made with the following three methods for estimation of mixing ratio within a
pixel. The first one is the well known least square method by means of Generalized Inverse Matrix. The second
the Least SQuare method minimizing the square error of
one is Maximum LikeliHood method and
estimated Mixing ratio.
last one is
It was found that the estimation accuracy of the Maximum LikeliHood method is
superior to the other methods in the case for the simulated data with S/N ratio of up to about 28.
KEY WORDS: Mixed Pixel, Estimation of Mixing Ratio,
1. INTRODUCTION
There are many classification methods of remortly
sensed imagery data. Many of them put label in a
pixel basis. Even for the pixels consist of plural
categories, ordinary classification method give one
category to the pixels. Then, it was considered that
information of mixing ratio within a pixel was
taken out without abandoning that information.
In remortly sensed image, estimation of partial
class mixing ratio within a pixel have the
significant role for land coverage. For example,
cloud coverage estimation based on category
decomposition is useful for making products of sea
surface temperature.
Category decomposition give us the way get that
proportion from a mixed pixel (MIXEL).
In this paper, we picked up the three methods for
estimation of category proportion within a pixel
which are proposed. Comparing these estimation
accuracies with simulation data consists of the
mixels added nomal distributed random number.
2. THE METHODS OF CLASS MIXING RATIO ESTIMATION
2.1 Least square method by means of Generalized
Inverse Matrix
Let an observed vector be I with the
dimensionality M, mixing ratio or proportion vector
be B with the number of classes N and the matrix
representing the spectral response of all classes
be A.
I=AB (1)
I: (1.1... In)? (2)
A= {Mo Ao. s Ma (3)
lA d. erm Aon |
benedés endi (dedita |
JA con s Ann |
B= (3.8... Bu)? (4)
Since I is given vector, if matrix A is assumed,
then vector B is determined under constraint that
minimizing norm of estimation error Ei,
| S E;? [--nmin, E- I - A B (5)
986
Inverse Problem Solving, Cloud Coverage.
Bs (A A A I (6)
Then, under constraint that the sum of mixing
ratio has 1.0, equation (6) is solved the following
analyticaly.
1 =u A 71
BA IS: ——————— (ACA UT (7)
ut (AA) I
At this point, u is a vector has factors with all of
1.05
2.2 Maximam LikeliHood method
If independent variables Xi (i=1,2,...., N) follows
normal distribution as a function of N(ui,o ;?),
the following equation is shown X and N.
X Sarit noi ÉD a: 0181) (8)
ist i=1 int
So that when the response of spectral reflectance
is independent of each other in the classes, the
following equation is conducted on the observed
vector I and N.
M
I5 9A Bí. NCA:// B, BÀ ZB (9)
iz
Aim Ave Arn, SR Ain ) (10)
7 = dias ( 0O::°, oi^... oin? ) (11)
The probability P; is shown equation(12) when I; is
observed in the i band and proportion vector has B.
Then probability P that the observed vector is I
and the proportion vector is B is shown the next
equation(13),
ftn C) M be pj CI. 1615
169
Sup CO) EY dN EN ad