GISdevelopment.net ---> AARS ---> ACRS 1990 ---> Poster Session

A high accuracy landcover classification method of multi spectral images using dempster- shafer model

Sun-Pyo Hong , Haruhisa Shimoda
Kiyonari Fukue, Toshibumi Sakata
Tokai University Research and Information Center
2-28-4 Tomigaya, Shibuya-ku, Tokyo 151, Japan


Abstract
Three new classification methods for multi spectral image are proposed. They are named as a like hood addition method a like hood majority method and a dumpster Shaffer rule method basic striates data and to combine obtained like hood for final classification these three methods use different combining algorithms.

From the classification experiments following results were obtained the method based on Demister Shafer's rule of combination showed about conventional method. This method needed about 16% more processing method showed 1% to 5% increase of classification accuracies how ever processing times of these of these two methods are almost the same with of a conventional method.

Introduction
With the launch of second-generation high resolution sensors like thematic mapper and HRV many kinds of researches have been done to certificate the capability of these sensors studies have shown that classification. Most of the results of these sensors have shown that classification accuracies using these sensors are not so high as expected when applying conventional supervised maximum like hood classifier using only spectral information. These results have made many researches to study spatial features like texture or more sophisticated classifier like expert systems or fuzzy classifiers.

One of promising methods, which can be through to increase classification accuracies, is to utilize multi spectral data. The most popular method of combining multi spectral temporal data is to just increase the dimensions of classification feature space. In other words multi temporal data are considered to be set of multi not necessarily shows improvements of classification accuracies because variances of each training data are usually increased this conventional method is called as a simple combination method in this image.

Proposed Methods
A pixel wise maximum like hood classifier based on spectral features is used as a basic classifier .Let be like hood of class c derived from multi temporal data set .in conventional simple combination (SC) method is calculated, Lc is calculated as


Where
c: Class, n : number of spectral bands m: number of temporal data.
t: transposed matrix determinant
Sc variance covariance matrix of class c
Mc mean vector of class -c x:pixel vector shown as
X = {x1(t1)x2(t1).............xn,(t1).x1(t2),x2(t3).........x1(tj),x2(tj),.................,xn(tj),.........,x1(tm),x2(tm),...........xn(tm)
t: ID of observation date,..........xi:pixel value of spactral band-1

Then a decision class is determined to the class showing the maximum like hood as follows

DC = C-max, if Lc-max=maxc[Lc]

In this method temporal features are treated as the same feature with spectral features .That is the dimension of feature space is equal to the product of the number of spectral bands and the number of nulti temporal usually increases decreases because the variance of each class usually increase compared to that of single temporal case. Consequently the SC method does not always show improvements of classification accuracies.

Three new methods of utilizing multi temporal data have been tested in this research .the first and second on are named like hood addition method and like hood majority method respectively the last one is based on Dempster sharfer rule of combination and is named as a DR method in these proposed methods the like hood of each class is calculated from each temporal image that is the like hood of class obtained from temporal image t is calculated for a pixel vector x ={x1(t),x2(t),........xn(t)} last the like hood claculated from temporal data.
  1. like hood adding method

    A score of class-c, S(c), is calculated in the LA method by the following equation.


    where. k is the number of classes. A decision class(DC) is determined to the class "c- max" if S(c-max) shows the mazimum score. which can be written as follows;


  2. Like hood majority (LM) method

    In the LM method scoring of like hood and decision of class are calculated using eq. (4) and (5) respectively


    note that the function f coverts the value of Lc(t) to binary data.

  3. Dempster rule DR method

    Dempster rule of combination is expressed by the following formula/1


    for a¹f, where K is the normalization coefficient and are expressed as


    Eq.(6) Shows that the degree of evidence from the first source which focuses on set B and the degree of evidence mc t2 from the second source which focuses on set c are combined by taking the product mb t1 mc t2 which focuses on the inter section of B and C this is exactly the same way in which the joint probability distribution is calculated from two independent marginal distributions.

    In this research is treated as a score of each classes B and C are defined as a subset of 1st 2nd and 3rd candidates of decision classes for each temporal data in order to aboid computational explosion 1st 2nd and 3rd candidates of decision classes correspond to classes having the largest 2nd largest and 3rd and when it is assumed that C1 C2 are the 1st 2nd and 3rd candidate class respectively the sub set B and C are expressed by.

    B= {C1,C2,C3,C1 U C3 C2 U C3 c1 U C3)     (7)

    C={C1,C2,C3,C1 U C3 C2 U C3 c1 U C3)      (8)

    Then mB (t1) and mc t2 are calculated by the formula.


    And a decision class is determined as follows.


    If j of "tj" is greater than 2 such as t = {t1,t2,t3,t4,..........}, S(c) is calculated for all tj by applying eq. (10) iteratively.
Experiments
In order to evaluate purposed methods described in chapter following four seasonal land sat TM data were classified by using new three methods anda conventional SC method. [Object area}

Sagami river basin ( in Japan) which has area of 12.8 km X12.0km [Observation area]

Nov 4 91984) jan 23 (1985) aug 6 (1986) and may 21 (1987) {image size}

512x480 pixels size =25m x 25m [Used channels}

TM Ch 1,2,3,4,5, and 7 figure 1 shows test images used in the experiments these images were registered in the identical UTM coordinate system.

Items in left hand side in table 1 shows 15 classification categories used in the experiments howevrt the total number of classification classes were fifty nine since each the same training area for each seasonal image thus training data set are consisted of four Training data cores spending to four seasonal images.

At the first stage like hoods of each class ie were calculations in all test images by using each training data. this calculation done independently for each test image that is Lc (t1) Lc (t2} is calculated by using training data corresponding to the test image of t1 t2 t3 respectively.

At the second stage three proposed methods ( the LA LM and DR methods ) were applied to four seasonal like hood data obtained in the conventual's methods land cover classification using the first stage on the other hand performed according to eq 1 and 2 in order to compare with a case of single temporal classification conventional maximum like hood of classification wee conducted for each set images the same training data set 2 shows classification results.

Processing times of MLC for a single image was about 15 minutes those of the SC LA and LM methods are about 60 minutes because of process for foue seasonal images. the DR methods needed processing times or about 70 minutes experiments were done by using HP9000 /835 mini computer system

Finally classification accuracies were estimated quantitatively by using digital test site data as shown in fig 3 test site data contain about 50 land cover use categories .As the categories used in test site data differ from these in land cover classification conducted in this experiments fifteen classification categories were merged to five major categories were merged to five major categories as shown in table 1 accuracy evaluations were performed based on these five major categories.

The MLC for each test image shows average accuracies from 62% to 65% the SC method indicates an average accuracy of 65% which is almost the same value to the largest none of the MLC result LA LM and DR method showed about 1% 5% and 12% larger average accuracies respectively combined to the results of the MLC and the SC method.

Conculsion
Four classification methods for multi spectral data were evaluated by experiments using four seasonal land sat TM data the first method is the features conventional simple combination methods which combines spectral features and temporal features in the same manner feature same manner as space and performs a classification with the conventional maximum like hood classification other three methods are newly proposed in this research.

The first purposed method is named as a like li hood additional method in which method scores of each class are calculated by adding like hoods obtained from each temporal data. the second method proposed is named as a like lihood majority method in which method decision class is determined to the majority class in decision candidate classes derived from each temporal data the last one is named as a dumpster saffers rule method which based on rule of combination.

From results of classification experiments following conclusions were obtained.
  1. The SC method and the LA method did not show large improvement classification accuracies.

  2. the LM method can be conducted by relatively simple algorithm however it showed 5% higher accuracies compared to the SC method.

  3. the method which showed the highest classification accuracies is the DR method accuracies were improved about 12% compared to that of the SC method.

  4. As a conclusion the DR method is recommended from the view pint of classification accuracies however the DR method needs about 16% more processing times than cases of the other methods.
References
  • Geroge J Klir and Tina A Folger Fuzzy sets Uncertainly and information Prentice Hall (1988)
Table 1 Classification categories
Classification
categories		 Number of classes major categories
1. Coniferous forest (3) Vegetation
2. broad leaved forest (5)
3. mixed forest (3)
4. Shadow of mountain (2)
5. Paddy (9) Paddy
6. High density urban (4) urban
7. Low density urban (3)
8. Housing area (3)
9. factoring (5)
10. sea (2) water
11. river (2)
12 farm (6) other
13 grass land (5)
14 waste land (6)
15 sands (1)



(a). JAN. 23, '85
(b) MAY 21, '87

(c) AUG. 6, '86
(d)NOV. 4. '84
Figure 1 Test images.

(a) SC Method
(b)LA Method

(c) LM Method
(d) DR Method
Figure 2 Classified results.



Figure 3 Digital test site data.
(overlaied on the image observed in NOV. 4)