GISdevelopment.net ---> AARS ---> ACRS 1989 ---> Forestry

Forest change detection A study of the flooded area in king amphoe phipunProvince of Nakhon SI Thammarat, Thailand

Ake Rosenqvist,Shunji Murai
Institute of Industrial Science University of Tokyo
7-22 Roppongi, Minato-Ku, Tokyo 106, Japan

Suvit Vibulsresth
National Research Council
Bangkok 10900, Thailand


Abstract
In November last year (188) the area of king Amphoe Phipun, in the province of Nalhon Si Thammarat in Southern Thailand, suffered seriously from flood and landslides (earthslides) after three consecutive days of heavy rain. One cause f the landslides is believed to be the change of the landuse in the region, were large areas of the natural forest, also on steep mountain sides, have been cut and changed into orchards and rubber plantation's

multitemporal satellite data was used in order to monitor the forest change for the period of the last 4 years. Principal component analysis on images recorded in April 1984 (MSS) and March 1988(TM) as applied to detect areas that had changed from vegetation to non-vegetation and vice versa. The mages were also classified using maximum data available and an on-site investigation not yet had been performed the classified images were used as preliminary reference.

To an effort to find out the relationships the landslides, the landuse ad the slope inclination, a SPOT image recorded in February 1989 (2.5 months after the flood0 was used as well as a DTM of the area.

Background
Due to the catastrophy, the area around King Amphoe Phipun has received very much attention lately ad a lot of research is now being carried out to find the causes of the disaster. If this can be done it may then be possible to predict where future disasters may strike and also, in a long term to prevent this from happening again.

Although forest change probably not is the single reason for what happened it is most likely that it played an important role. Concerning this area, the change is to a great extent natural forest that has been converted into rubber plantations. Problems occur when this is made not only on flat land but also on mountainsides. Since the roots of rubber tress are very shallow compared to those of natural forest, it make rubber tress less capable to hold the soil, for instance in case of heavy rain. This is also one theory of what actually happened in Phipun last November. Smaller landslides occurred as rubber dams of mud and logs. As the rainfall continued these occurred and propogated as fallen tree trunks, water and mud tore up fore tress downhill. Many people were killed as masses of logs, debris and mud reached the village on the plain downstream.


Fig.1 Landslides on the mountains near Ban Kathun Village

The Change Detection
  1. Preprocessing

    1. Geometric correction

      A Landsat MSS scene registered on the 12th of April 1984 and a Landsat TM scene registered 4 years later, on the 30th of March 19088, were compared to detect the change of the vegetation during that period. The MSS image was geocorrected using a third degree polynomial transformation and resampled from 80m to 50m-pixel size. The TM image was geocorrected applying the same kind of transformation and resample from 30m to 50m-pixel size. The images differ from each other with in average less than 0.5 pixels in both X and Y.

      An area of 24*24 km, covering the region worst struck by the flood, the river plain and the water catchments area of the two river channels where most of the mud and debris came down, was chosen as study area.

    2. Sensor characteristics

      Because of differences in the spectral characteristics of the sensors of the multispectral scanner and the Thematic Mapper. Only the bands corresponding in both satellites could be used in the principal components analysis. MSS band 1 and 2 (visible green and visible red) correspond well to TM band 2 and 3 respectively, but when it comes to the invisible reflected infrared band, which is very important for determining biomass content, the sensors differ some. However, the band 4 covers most of MSS band 4 and the two were therefore chosen as the third pair of bands to be compared in the principal component analysis.

    3. ls excluded from the computations

      When it comes to remote sensing data, and especially of this latitude it is a well known fact that we often have problems with cloud cover. Although the MSS image in principle was cloud free, we still had the consider this problem since the TM image contained some scatters clouds the mountains in the northern part of the study area.

      Apart from the fact that clouds actually conceal information in the images, they also have a strong effect on the principal components analysis. Clouds will be interpreted as areas of very strong change and this results in that the effect of areas actual change will become very faint. Pixels where clouds or cloud shadows could be found were therefore set to zero in both images and excluded from the computations.

      Since the investigation concentrated on the change in the mountainous areas. It was also possible to improve the accuracy by excluding pixels on the flat land beneath the mountains. Points below the altitude of 100 m were therefore considered to be of less interest although some may have had occurred.


  2. Principal Component Analysis (PCA)

    1. Some theory

      The basic principal of the detection is that areas that have not changed in the period in question will have approximately the same not changed in the period I both images and therefore be highly correlated, whereas areas of significant change consequently will have low correlations.

      The property of a principal component transformation is that it reduces redundant information in the input data (in this case 6 channels; 3 from MSS and 3 from TM) and creates a new set of coordinate axis (here 6) where the data in uncorrelated. Areas of significant change will then occur a high or owl values in the higher principal components.

    2. The analysis

      From Table 1 it can be seen that the first principal component (PC-1) contains most of the variance in the data, PC-2 contains the next major part and so on. PC-5 and PC-6 have very low variances and consist mainly of noise.

      Table.1 Information contens of the principal components
      PC-1 PC-2 PC-3 PC-4 PC-5 PC-6
      Variance 3.436 1.567 0.57 0.349 0.072 0.018
      Info. Content (%) 56.3 26.1 9.3 5.8 1.2 0.3
      Cumulative (%) 57.3 83.4 92.7 98.5 99.7 100.0

      Areas of significant change could be found in PC-3 as especially high and low values. The high values represent change from vegetation 1984 to non-vegetation 988 and the low values indicate change n the other direction. The class non-vegetation includes bare soil as well as very sparse vegetation where the ground still can be seen from the air. When the thresholds for the areas of change had been found, PC-3 was classified. Some noise was reduced by using a threshold in PC-1 as well.

    3. PCA Results

      The analysis show that a total area of 1510 hectares had been changed into non-vegetation during these 4 yeas and that only 210 hectares had recovered and changed in the other direction. This totals in a net decrease I the vegetation cover with 1300 hectares. The figures are of course approximate but they give an indication of the magnitude of the change.

      Although the image quality of Fig 2 is bad, the detected change from vegetation to non-vegetation can be seen as white spots, especially on the mountains sides closest to the flat land in the valleys.


      Fig.2 King Amphoe Phipun. hill shading from DTM (24 km * 24km). White spots: vegetation 1984= > non-vegetation 1988

    4. Land cover classification

      In order to check the correspondence with the PCA, as well s to find out the present land use (i.e March 1988) in the area, both images were classified using Maximum likelihood method. However, since no ground truth information of the area was available and an on-site investigation not yet had been performed, only three classes were used, i.e. bare soil, parse forest and dense forest. Baresoil had the same meaning as in the PCA: bare soil and very sparse vegetation where the soil can be seen from the air. Spare forest includes young rubber and orchards while the dense forest class includes natural forest and most probably also dense rubber plantations.

      As for the PCA, pixels below 100m as well s cloud pixels in either of the images were excluded from the classifications.

    5. Results

      The results of the two classifications can be found in table 2 it can be seen that approximately 3500 hectares of dense forest in the test area had been converted into some other kind of land use during these 4 yes, a decrease with 14%. The areas of sparse forest and bare soil increased at the same time with 18% and 20% respectively.

      Table 2: Maximum likelihood classification
      Land type MSS 1984
      hectares      		 TM 1988
      hectares      		 Change
      hectares       		Ratio
      1988/1984
      Bare soil 560 1850 1290 3.30
      Sparse forest 12480 14710 2230 1.18
      Dense forest 26010 22490 -3520 0.86
      Total 39050 39050    

      The 1290 hectare figure for bare soil can be confirmed by comparing it with the PCA, which showed a net of decrease of 1300 hectares in the vegetation. The correspondence is almost surprising while the figure itself if depressing.
The cause of the landslides - an interim report
  1. Present research

    It is now of course of great importance to try to find out the reasons why the landslides actually occurred and what we can do to prevent this from happening again. An investigation is at present being carried out but is yet not finished. Some interim results will however be presented below, as well a short discussion about the methodology and influencing factors.

  2. Parameters of importance and material used.

    It is likely that a combination of many factors led to the disaster. In the present study the land cover and the slope gradient, together with the frequency of the landslides, have been considered but we also have to keep in mind that other factors, such as the surface geology, the drainage pattern ad the amount of rain at the time of the occurred, also most probably played important roles.

    A maximum likelihood classification of the TM image from March 988, with a pixel size do 25m, has been used as the assumed land cover at the time. The same rough choice of the classes, as used in the change detection above, will remain until the ground truth investigation has been performed (December 1989).

    A digital elevation model (DEM) with 25m grid size has been created by scanning topographical map over the area. To reduce discrete areas, the DEM was smoothed by utilizing a combination of a bilinear and cubic convolution after which the slope gradients were computed.

    A SPOT multispectral image recorded on the 9th of February1989, i.e. 2,5 months the disaster, has been used to map the occurrence of the land slides. The image is more or less cloud free and the traces of the land slides can be seen very clearly. To fit the data size of the rest of the material used, the image was geocorrected and resample to 25 m pixel size.

    As for the change detection analysis described above, pixels below 100 m have been excluded. A cloud mask has also been used, slightly different from the one used above since clouds and shadows in the 1984 MSS image not had to be considered and since the pixel sized here was 25 m instead of 50m.

  3. The methodology

    The aim of this investigation is to figure out which factors, or which combination of factors, that played important roles at the time of the disaster. The traces of the landslides were therefore delineated. (1) Manually, to get the exact positions where the landslides occurred, and (2) By thresholding of the SPOt image, to get an approximate figure of the are that was affected.

    The landslide information was overlaid on the landcover classification and slope gradient mentioned above and compared pixel for pixel, in order to get some indication about any relationship between these. The test area was divided into 6 sub-areas since the frequency of the landslides not was homogenous over the whole region. The sub areas were then compared with respect to the frequency, land use and slope gradient.

  4. Some initial results

    From table 3 it can be seen that an area of approximately 3000 hectares was exposed to landslides; a considerable figure. It is also noteworthy to see that as much a 1250 hectares of the dense forest suffered. It is know that landslides also occurred in the natural forest, but the figure here seems a bit too large. This could however be explained by that the dense forest class apart from natural forest also most probably consists of dense rubber plantations, especially in areas near the villages and along the valleys.

    The unclassified land type consists mostly of areas that were under clouds in the TM image.

    Table 3: Landuse March 1988 and landslided areas November 1988
    Land Type Total area (ha) % Land slide area (ha) % Landslides/total
    Bare soil 1930 4 200 6 10
    Sparse forest 15200 33 1360 43 9
    Dense forest 22640 49 1250 40 6
    Unclassified 6500 14 360 11 6
    Total 46270 100 3170 100 7

    Concerning the occurrence of the landslides in the different subzones, table 4 shows that Khao Kathun and Khao Chong Lom Tai landslided area of 11% and 14% respectively. Area 3, Khao Plai Kathun shows a high landslide percentage on are soil areas while the figure approximately is half in the sparse and dense forest classes. The frequency is lower in the west and east sub-zones and reaches the lowest values in the southern part.

    The figures in parantheses show the representation of the different land use types in each sub-zone.

    Table.4 Landslided area/total area for each landuse type (landuse type total area/total sub-zone area)
      Bare soil (%) Sparse (%) Dense (%) Total (%)
    Khao Kathun 19 (3) 18 (21) 10 (76) 11 (8)
    Khao chong Lom Tai 14 (7) 15 (50) 12 (43) 14 (17)
    Khao Plai Kathun 1 (2) 7 (26) 7 (72) 5 (6)
    West 8 (7) 5 (28) 2 (65) 4 (15)
    East 7 (3) 6 (28) 2 (68) 4 (38)
    South 2 (5) 2 (66) 4 (19) 2 (15)
    Total 10 (5) 9 (38) 6 (57) 7 (100)



    Fig.3 SPOT XS image recorded on February 9th 1989. Showing the landslides and the 6 frequency zones. The large white area indicates points below 100m.

    In an effort to investigate the influence of the slope inclination together with the land use, the slope gradient and the land use classification were compares simultaneously at landslided points. As expected, dense forest was the land use type that showed the highest angle of inclination (mean value) at landslided points. Followed by sparse forest.

    Table 5: Slope gradient mean and standard deviation (degrees) at positions where landslided where landslided occurred.
      Bare soil (%) Sparse (%) Dense (%) Total (%)
    Khao Kathun 20 (.6) 21 (8.5) 21 (7.6) 21 (7.8)
    Khao chong Lom Tai 20 (7.0) 20 (8.0) 22 (7.4) 21 (7.6)
    Khao Plai Kathun 22 (5.2) 22 (7.1) 2 (7.4) 21 (7.4)
    West 14 (6.0) 17 (9.0) 22 (9.3) 21 (9.3)
    East 19 (7.5) 20 (7.8) 20 (7.9) 20 (7.9)
    South 17 (8.6) 18 (7.3) 21 (8.0) 18 (7.6)
    Total 19 (7.4) 20 (8.0) 21 (7.8) 20 (7.9)

    Bare soil showed the least capability to hold the soil, which also was to be expected. It is a little surprising though, to find such steep angles for bare sol in the Khao Kathun and Khao Plai Kathun areas and such relatively low values for all the classes in the most southern sub-zone.
Discussion
The change detection analysis show that we have had a considerable change in the land use in the region between 1984 and 1988. large areas of dense forest ha been converted into other kinds of landuse, including more than 2100 hectares of bare soil that has occurred.

From the landslide investigation above we can see that the occurrence occurrence of the landslides (as to be expected) actually is related to the land use, i.e bar soil has a higher frequency of landslides than sparse forest which in turn has a higher frequency than dense forest, but it is also clear that it is not the sole reason. It is known for instance from aerial photographs that landslides, although not so frequency also occurred in natural forest areas.

The slope gradient analysis indicate a lot of bare soil and sparse forest areas on steep slopes in for instances Khao Lathun and Khao Plai Kathun, but any significant relationship between the gradient, the landuse and the land slides can not be found. It is obvious that we will have to include more parameters, such as the surface geology, the drainage pattern and rainfall data, if we want to get a clearer picture of the causes.

Concerning the weather at the time of the disaster, we do have some information from the Royal Irrigation Department in Bangkok, that as much as 680mm of rainfall is to be considered extraordinary rare and this could be the explanation why (1) landslides occurred also in natural forest areas and (2) why so many landslides, even on mountains completely unrelated to each other occurred all at the same time.

Many questions still remain and the next step in this research work will be to include a drainage model and some geological information in the investigation, in an effort to straighten some of the question marks out.

Acknowledgements
We would like to thank Mrs. Siripong Absorndua, associate professor at the Chulalongkorn university in Bangkok, for sharing a lot of useful information with us.

A special thanks also to Mr Ernst Ramberg, a legendary character at the Royal Institute of Technology in Stockholm, Sewden, for giving inspiration and good advice.

References
  1. Arbhabhiram, A et al 1987, Thailand Natural Resources Profile, Thailand Development, Research Institute.

  2. Niblack, W. 1986, An Introduction to Digital image Processing prentice - Hall International, ISBN 0-13-480674-3

  3. Richards, John A., 1986 Remote Sensing Digital Analysis Springer - Verlag Berlin Helderberg, ISBN 0-387- 16007-8.