GISdevelopment.net ---> AARS ---> ACRS 1990 ---> Land Cover/Land Use

The study of land use survey in the tropics using Multi Season and Multi Sensor Remote Sensing data

Yasushi Shimoyama, Izumi Kamiya,
Masanori Koide, Tokio Mizuno
Photogram metric research and development office Geographical survey institute
Kitazato -1 Tsukuba,305 Japan


Abstract
In the tropics it is important to employ multi- season and multi sensor data for continual environmental monitoring and for more accurate land use survey which copes with the seasonal land cover changes the accuracy of land use classification using multi season data is controlled by accurate registration among the image data. Conventionally each image data set is rectified independently to a standard map coordinate system for registration purpose .How ever this method causes small registration method using correlation coefficient must be developed in order to lessen the small errors.

This study concentrates on the image registration method for LANDSAT MSS & TM SPOT HRV (XS) and MOS-1 MESSR it includes following four steps.
  1. Image overlaying on a display for rough geometric correction

  2. optimization of sub-pixel differences at a set clipped patch area by the coefficient between two images.

  3. calculation of coefficient of affine transformation at local area whose vertexes consist of the patch area.

  4. execution of geometric correction at each local area.
This method improved the registration accuracy land use classification from these image data sets also showed significantly higher accuracy than those for a single image this results indicated the significance of employing multi season and multi sensor data for accurate land use survey.

A Registration Method
  1. Necessity of image registration

    Use of multi season and multi sensor data is desirable in order to monitor temporal environmental at a wide area and the achieve accurate land use multi season and multi sensor data the accuracy of land use classification depends considerably on that of image registration consequently accurate image registration is required.

  2. image registration by data correlation

    the image of the highest resolution was rectified to a map coordinate system to create a standard image then the rest of the input image were registered to the standard image

    A set of patch area is manually clipped from points the in order to simplify the search for their conjugate points the image correlation method was employed to utilize computers and the realize accurate results with low resolution images in which edges are not well defined the patch areas were shifted bit by bit to find a point where the correlation coefficient is largest the interpolation of the standard image to the grid of registered image.

  3. Geometric correction

    The geometric correction was executed by the affine transformation by the unit of each local area the coefficients at each local area were calculated using the residents of four area are vertexes which consist of conjugate points surroundings of local area are also corrected by the affine transformation of the nearest local area
Case Study
The test site in the this study is phuket island of Thailand . the specification of the image data for the case study are as follows :

1
2
3
4		 SPOT
MOS-1
LANDSAT
LANDSAT		 HRV-XS
MESSR
TM
MSS		 1,2,3
1,2,3,4
1,2,4,7
4,5,6,7		 bands
bands
bands
bands		 12.13.1988
02.06.1989
02.04.1989
04.20.1987

The standard image data was SPOT HRV data and the other images were registered to the spot image.
  1. The correlation coefficients at each patch areas.

    The optimized correlation coefficeients at each patch area are shown at table 1 the correlation coefficients of highest calculated except for band 1 of each combination because patch 2,5 and 8 are chose at land where topographical features are inferior and the other patch areas are all set at the seashore the correlation of 2,5 and 8 were low.

    After the correlation coefficients were computed patch areas are moved vertically and horizontally by the interval of 0.2 pixel and optimized reasonable are set where the coefficient of correlation is largest .

    At the surroundings of the optimized position of patch area this coefficient of correlation changes in table 3 this proves that the coefficient is dominant around the optimized position and that residuals can be measured reasonably by the unit of sub pixel.

  2. the coefficient of correlation of test site.

    The affine transformation was employed to geometrically correct the image data of the test site to investigations yo accuracy of geometric correction the changes the correlation coefficients were analyzed in an urban area of 5*k km at the southeast of test site in each combination the coefficients were so much improved the effect of the image matching method.
Acquisition of training samples and island uses classification.
  1. Acquisition of training samples

    Training areas were collected as polygon data with a digitizer by comparing the image displayed on CRT with the field survey data on this maps training samples were acquired from the image pixels surrounded by the training areas data.
    Table 1 optimized correlation coefficient at each patch area
      Optimized correlation coefficient
    Patch area No. Between HRV & TM Between HRV & MESSR Between HRV & MSS
    1 0.90 0.87 0.90
    2 0.76 0.79 0.53
    3 0.83 0.94 0.91
    4 0.85 0.92 0.92
    5 0.87 0.91 0.70
    6 0.93 0.94 0.90
    7 0.83 0.93 0.95
    8 0.75 0.88 0.87


    Table 2 correlation coefficient before the registration.
      HRV
      1 2 3
    T
    M    		 1 0.78 0.68 -0.49
    3 0.71 0.75 -0.28
    4 -0.66 -0.34 0.91
    7 -0.12 0.17 0.56
    M
    E
    S
    S
    R    		 1 0.72 0.53 -0.63
    2 0.35 0.50 -0.03
    3 -0.60 -0.28 0.86
    4 -0.63 -0.33 0.73
    M
    S
    S    		 1 0.76 0.49 -0.74
    2 0.75 0.68 -0.53
    3 -0.42 -0.08 0.72
    4 -0.52 -0.23 0.73


    Table 3 the changes of correlations coefficients at the surroundings of the optimized position of patch area
      Sift to the column direction
    -0.4 -0.2 0.0 0.2 0.4
    Shift to
    the row
    direction    		 -0.6 0.44 0.62 0.79 0.82 0.78
    -0.8 0.46 0.66 0.85 0.86 0.80
    -1.0 0.40 0.67 0.90 0.87 0.77
    -1.2 0.05 0.50 0.81 0.73 0.60
    -1.4 -0.48 0.19 0.64 0.57 0.45


    Table 4 Correlation coefficients after the registration
      HRV
      1 2 3
    T
    M    		 1 0.81 0.71 -0.51
    3 0.75 0.79 -0.31
    4 -0.64 -0.31 0.94
    7 -0.12 -0.22 0.56
    M
    E
    S
    S
    R    		 1 0.86 0.69 -0.62
    2 0.45 0.66 0.04
    3 -0.31 -0.31 0.92
    4 -0.68 -0.37 0.92
    M
    S
    S    		 1 0.77 0.51 -0.71
    2 0.65 0.63 -0.41
    3 -0.50 -0.12 0.80
    4 -0.62 -0.28 0.85



    Figure 3 Acquisition of training area

  2. Land use classification

    the maximum likelihood method was employed for the land use classification of training data .Classified result is shown in table image data was superior to the other image data then a single image data was classified independently .How ever when two image data were simultaneously one reason is that the combination of low resolution data with high resolution data enables us to perform the accurate classification considering the surrounding information of the pixels of high resolution data another reason which may be dominant is that the data of MSS data acquisition is different from that of the other data. Table 6, 7 and 8 show the error matrix of classified results by MESSR , MSS and the combination of MESSR and MSS respectively In the classification by MESSR urban areas and water areas are accurate and on the other hand in the classified by MSS land cover of the vegetation is accurate so in their combination all of land cover class becomes more accurate consequently this accuracy increase is one of the advantages of using the multi season and multi sensor remote sensing data.
Conclusion
A couple of conclusions of this study are summarized blow.
  1. The development o image registration method.

    The image registration method developed in this study significantly reduced the registration errors among input images also simplified the data processing of multi season and multi sensor images

  2. Improvement of the accuracy of the training data classification

    The combination of three images acquired in different seasons improved the accuracy of training data classification .

    We are now improving our method to employ geological data soil data and DTM of the test site these data are expected to improve the land use classification of this study.
Acknowledgements
We thank TDD Ministry of agriculture and cooperatives Thailand especially mr. Manu OMAKUPT Ms Promchit TRAKULDIST and Mr Anusorn Chantanaoj for assisting our field survey.

Table 5 Acuracy of training data classificaion(%)
One sensor Two sensor Three sensors Four sensors
H 87.9 H+T 96.0 H+T+ME 98.9 H+T+Me+Ms
99.7
T 90.1 H+Me 96.3 H+T+WS 99.2
ME 89.9 H+MS 97.4 H+ME+Ms 98.1
MS 87.9 ME+MS 96.8 H+Me+Ms 99.4
  T+MS 98.1 H........HRV       T........TM
Me+Ms 98.2 ME.....MESSR       T.......M


Table 6 The error matrix of classified results by MESSR
  Classified land use
Urban Paddy Rubber Coconut Forest Mangrove Mine Water
Original
Land
Use		 Urban area 560 11 0 11 0 0 21 0
Paddy 5 556 34 24 0 0 9 0
Rubber 0 642 14 40 40 0 0 0
Coconut 4 11 39 91 0 0 0 0
Forest 0 4 6 42 361 22 0 15
Mangrove 5 13 0 0 16 485 0 2
Mine 5 0 0 0 0 0 282 0
Water area 20 0 0 0 20 0 2 534


Table 7 The error matrix of classified results by MSS
  Classified land area
Urban Paddy Rubber Coconut Forest Mangrove Mine Water
Original
Land Use		 Urban area 425 134 0 28 0 0 12 0
Paddy 65 499 0 2 0 16 0 46
Rubber 0 5 666 0 25 0 0 0
Coconut 0 8 0 137 0 0 0 0
Forest 0 0 9 20 418 3 0 0
Mangrove 4 8 0 0 0 506 0 0
Mine 8 0 0 0 0 0 279 0
Water area 6 33 0 6 0 0 28 503


Table 8 The error matrix of classifed results by MESSR and MSS
  Classified land uses
Urban Paddy Rubber Coconut Forest Mangrove Mine Water
Original
Land Use		 Urban area 580 10 0 1 0 0 12 0
Paddy 9 615 0 4 0 0 0 0
Rubber 0 3 689 02 0 0 2 0
Coconut 2 4 0 139 0 0 0 0
Forest 0 0 0 1 449 0 0 0
Mangrove 4 5 0 0 0 509 0 3
Mine 1 0 0 0 0 0 286 0
Water area 1 0 0 0 0 0 5 570

accuracy = 98.2%