GISdevelopment.net ---> AARS ---> ACRS 1989 ---> Digital Image Processing 2

Texture analysis using differnce statistics for land cover classification

Michiyusu Akasaka, Katsunoii Furuya and Ryutaro Tateishi
Remote Sensing and Image Research Center Chiba university
1-33 Yayoi-Cho Chiba City Chiba 250 Japan


Abstract
To improve land cover classification accuracy spatial information should be considered this study deals with difference statistics method statistics which is one of texture analysis the optimum parameters of difference statistics were investigated and the way to apply difference statistics to land cover classification was proposed among four known features from to provide statistics inaugural second moment and entropy are found to provide better results for land cover classification concerning multi spectral band for difference statistics near infrared band 3 land sat is found to be best.

Introduction
Satellite imagery data include both the spectral information and the spatial information how ever most of approaches to land cover classification has been used only spectral information systematic and suitable land cover classification method using spatial information is not yet established in this study difference statistics were picked up to extract the textural information of land sat Tm image SPOT HRV image authors investigated the use of textual features for classification of land cover.

Texture contain information about the spatial distribution of tonal variations with a band texture analysis is separated in to statistical texture analysis and structural analysis four standard approaches to statistical texture analysis make use of features based on co-occurrence matrix on difference statistics on run length matrix and on the fourier power spectrum respectively co- occurrence matrix and difference statistics have a capability of discriminating textual feature than the others co occurrence matrix method is based on the second order joint probability densities of parts of gray levels while difference statistics method is based on first order probability density function therefore distribution on difference statistics is more stable than co-occurrence matrix method fore this season difference statistics method is applied for texture analysis in this study.

Difference statistics
Procedure to obtain density function to calculate difference statistics is given in fig 1.

ex.) displacement d ={ r ,q )
          = { 4, 0- 360° )
r : inter sample spacing distance
q : angle

Pd (k) = f (K) / N K : difference of image data at two points with the displacement d (r, q ) ( k=0-225, integer )
P d (k) : density function (Probability of difference k)
f(k) : frequency of difference k
N: number of surrounding pixels
( N=24 in this ex. )


Fig.1 Procedure to obtain density function.

Four texture are defined from each of those density functions pd they are follows.
  1. Angular Second Moment ( A.S.M )


  2. Contrast


  3. Mean


  4. Entropy

Data Description
Land sat tm data and spot HRV data were used in this study the test site contains urban residential areas forests paddy golf fields course and marches tm data and HRV data were resample to 30 m and 20 m respectively test data in this study is as follows.

  Land sat SPOT HRV data
Band : 1-7 : 1-3 ( multi spectral )
Path Row : 107-35 : 331-279
Date : 24 July 1987 : 7, April 1986.
Processing level : Bulk : 1 B ( bulk)

Determination of optimum parameters for land cover classsificastion
Optimum parameters of differences statistics was determined by the following procedures.
  1. Collection of ground truth data human interpretation of the image with help of existing land use map and topographic maps.

  2. Comparison parameters of difference statistics.
    The following three parameters were investigated.

    Inter sample spacing distance : from 1 to 10 pixels
    band : Land sat -TM -1-7 ( except bands 6), SPOT HRV =1-3
    feature : A.S.M., Contrast, Mean, Entropy

  3. Selection of optimum parameters.
Land cover classification
The study area was classified by maximum likelihood method using multi spectral data and textual data which is derived by the determined optimum parameters of difference statistics land cover categories were forest paddy field urban open space water and golf course we compared classification accuracy in the cas3e of using multi spectral data and textual data with the case of only multi spectral data .


Fig.2 Flow of processing.

Results of comparison difference statistics
  1. Intersample Spacing Distance
    Fig. 3 shows density function of each inter sample spacing distance of forest by SPOT-HRV band 2.

    The distribution is almost the same with inter sample distance exeeds 4 pixels our formar study in mountain area used land sat tm data same result.

  2. Band
    Fig. 4 shows mean and plus or minus standard deviation of entropy of four categories for each spot hrv band in the case of thar inter sample spacing distance is 5 pixels.

    Band 3 SPOT-HRV data is compartively better than at the other bands for discrimniation of four categories bt our formarstudy lqand sat band 4 is ther best foer classfication wave length of land sat tm band 4 is 760 to 900 nm and spot hrv band 3 is 790- 890 nm both bands have almost the samewave length region larger tahn the value of urban at band 3 however this relation is reverse at band 1 nad 2 that means textures are different in differnce band image this effedct tells us the use of multi spectral. texture analysis.

  3. Features
    Four Features of SPPOT-HRV band 3 with difference inter sample spacing distance are given in Fig.5 (a) A.S.M., (b) Contrast, (c) Mean, (d)Entropy, respectively.

    These figures statistics this standard deviations of contrast and mean are so big that difference statistics by their features have less useful ness to extract the textural information entropy and A.S.M are found to provide better results for land cover classification .

    The distribution of entropy of land sat band TM 4 given in fig 6.

    Distribution of paddy field nearly equal to that of urban.

Fig.3 Density function of each distance.
(SPOT_HRV band2, golf course)


Fig.4 Comparison of bands.
(SPOT-HRV, distance=5, Entropy)


Fig.5 Four features.
(SPOT-HRV band3)
a)A.S.M., b)Contrast, c)Mean, d)Entropy


Fig.6 Entropy of Landsat-TM band4.


Results of classification
classification accuracy by maximum like hood method used spot data is given table used only spectral information used both spectral information and textual information parameters of difference statistics to extract textual information are distance 5 band 3 and features entropy.

In spite of textural information was added classification accuracy was almost the same or was deteriorated on almost on categories mis classified pixels were found on border of some categories this tendency is shown also in the case of tm data.

Table 1 Classification accuracy.

a) only spectral information
  FOREST PADDY URBAN WATER OPEN GOLF
FOREST 12220 1845 109 0 28 28
PADDY 9 6341 544 3 1 0
URBAN 191 1439 3653 205 36 15
WATER 0 0 4 6261 0 0
OPEN 0 1133 198 0 735 9
GOLF 0 115 9 0 20 956
UNDEFINED 0 0 86 14 0 0
TOTAL 1420 19873 6603 6483 820 1008
PERCENT 85.92 58.32 85.61 96.58 89.63 94.84

b) spectral and textual information.
  FOREST PADDY URBAN WATER OPEN GOLF
FOREST 1183 2036 101 18 44 40
PADDY 11 5875 511 7 0 0
URBAN 224 1803 5699 721 42 2
WATER 0 0 4 5723 0 0
OPEN 0 1001 200 0 727 7
GOLF 0 158 6 0 7 959
UNDEFINED 0 0 86 14 0 0
TOTAL 1420 10873 6603 6483 820 1008
PERCENT 83.31 54.03 86.31 88.28 88.66 95.14

Conclusions
As the results to apply difference statistics to spot HRV data and land sat data the following things are found out. optimum parameters of difference statistics for land cover classification are.
  1. Inter sample spacing distance : 1-4 ( pixel )

  2. band pixel spot HRV band land sat TM band 4

  3. Feature entropy or Angular Second Moment (A.S.M)
Textural information by differences statistics of some categories have enough capability for land cover classification how ever simple introduction of textural information by difference statistics in to spectral information can not raise up classification accuracy for improvement of land cover classification analysis we have consider the followings.
  1. Texture information by differences statistics of some categories have enough capability for land be applied only in the area where spectral analysis can not provide good classification accuracy.

  2. Multi spectral texture analysis is effective when the texture of difference bands is different.

  3. Interpretation spacing distance of difference statistics should be small when a size of area of uniform category is small in order to reduce misclassification on borders.
Reference
  • J.S. Weszka C.R.Dyer and A Rosenfeld : "A compartive study of texture measureses for terrian classification", IEEE Trans., vol. SMC-6 no.4, pp. 269-285 (1976)
  • R.W Conners and C. H Harlowa : "A theroretical comparison of texture algorithms IEEE trans. Vol. PAMI-2, n0-3 pp 204-222 (1980 )
  • M. Akasaka, K. Furya and R. Tateishi: "Land cober classification using difference statistics", Proc. conference of JSPRS, Tokyo pp. 107 112 (1989)