Detection of the windfall damage to forests caused by the Typhoons 9117 and 9119

Detection of the windfall damage to forests caused by the Typhoons 9117 and 9119

Gen Takao
Forestry and Forestry Products Research Institute of Japan
P.O. Box 16, Tsukuba Norin Kenkyu Danchi, Ibaraki 305, Japan

Abstract
The typhoons 9117 and 9119 attacked successively northern Kyushu with exceptionally strong wind in September of the last year. it left severe damages to the agriculture and forestry in those prefectures.

The author carried out the windfall-damage area detection on the forest in northern Kyushu using MOS MESSR images. The damage appeared as the inorement of the CCT value of the Band 2 on the images. Thus, the damage areas are detected using the two dimensional histograms of Band 2s of before and after the typhoons. The atmospheric effects to the images are considered locally.

Using the result of the detection, the incidence of the damage in relation to topographical factors were analysed with DTMs.

Introduction
In September of 1991, the typhoons 9117 and 9119 successively struck northern Kyushu ( the most west of Japan’s main islands ) with extremely strong wind. Among them, the typhoon 9119 was especially strong that it broke records of the maximum instantaneous wind speed at many observatories; e.g. 60. 9 m.s at Mt. Aso, 52.9 m/s at Kumamoto city or 44.4 m/s at Hita city ( Meteorological Agency, 1991 ( Figure 1)

Fiture 1. The locus of the typhoons (yamamoto, 1992)
From those typhoons, forest productions in northern Kyushu suffered considerable damages ( Figure 2) . For examples, in Ooita pref., where suffered the most damaged in the prefectures of this area, the damage amounted about 430 million dollars and the suffered area reached 220 km2, which is about one-tenth of whole forest area of the prefecture ( Forest Agency, 1992). The most part of the damage was caused by windfalls on coniferous plantations. Windfall surveys were carried out in every towns and villages by the prefectures using aerial photographs and / or ground observations.

Figure 2. The windful damage in a coniferous stand (Forest Agency, 1992)
To understand the distribution and tendency of damages to the topography in this area, the overall survey with a criteriaon is necessary instead of one as the sum of small scale surveys among which the cirterion may vary. In this paper, a detection of wide – speeded windfall damages of forests over northern kyushu is attempted using satellite imagery.

Data and Preprocessing
Two sets of MOS MESSR data taken before and after the typhoon are used. Each set has contiguous two scenes on an orbit. One set was observed in Apr. 15, 1991 ( hereafter referred to as the “April image” ) and the other was in Nov. 5, 1991 ( hereafter referred to as the “November image”) These scenes cover the most severely damaged area, i.e. whole Ooita pref., western Fukuoka pref. and northern Kumamoto pref., No clouds can be seen on the April image but some clouds on mountains on the November image.

a The DEM and the municipality map of the Digital National Land Information are prepared to classify the slope direction of damages and to round off the damage for each municipality. The unit of the data is the ¼ mesh’ of the Digital National Land Information, which has 45 sec. width in longitude and 30 sec. width in latitude.

Two scenes of each set of images are merged into an image and the consequent images are projected on the longitude – latitude coordinate. The size of a pixel is equivalent to one-fifth in either direction of a ¼ mesh, and is about 58 meters along longitude and 46 meters along latitude around Hita city, Ooita pref. ( N320 20).

Forest areas are extracted from the April image by clustering method. Decidous forests are excluded because the leaves had not opened yet at the time observed. The other forest types, such as coniferous plantations, pine forests, evergreen or mixed broad – leaved forests are not separated from each other in the forest classes. The optical characteristics of forested areas are rather stable except for deciduous trees in winter ( Kaufman, 1989). Then, the change detection is considered to be possible within the forest classes. Clouds and their shadows on the November image are extracted and the corresponding part of the forest areas are reduced. Then, the edge of the forest areas are removed from the further analyses to avoid detecting errors on overlaying . Finally, a forest mask on the images is prepared for following analyses.

For the most part of processing and analyses, the International Imaging System’ the System 600 on a Sun-3 and a Sun Sparc Station 2 are used, and some detailed analyses are carried out with programs coded in Sun Fortran,

Change Detection

The difference between damaged and non – damaged forest on the images
Pixel values of the damaged and non – damaged forests are compared to clarify the difference between those forests on the images.

As sample data, forests in the Kusu and Kokonoe region are used. This region is one of the most severely damaged areas in Ooita pref. Sample pixels are selected, referring aerial photographs taken immediately after the typhoons, and finally 48 pixels of completely fallen stands and 48 pixels of non-damaged stands are picked up.

Scattergrams of those pixels, value before and after the typhoons in every bands and the NDVI are shown in Figure 3. From these scattergrams, some characters of the damaged pixels can be pointed out: no clear differences between damaged and non-damaged pixels can be identified on the infrared bands ( Band 3 and Band 4 ) and the NDVI, on the other hand the most clear differences appear on the red band ( Band 2) This distributions of damaged and non-damaged pixels seem to be nearly separated on the scattergram of Band 2, and non – damaged pixels distribute in a narrow range for each Band 2 value in April ( See Table ).

Figure 3. The scattergrams of each band before and after the typhoons
As the method to detect cut-over areas, multibands analyses such as the principal component analysis ( Onuki et al, 1982 ) or the change vector analysis ( Awaya et al, 1986) had been applied . The windfall damages, though, appear in only the red band on the images. Allen and Lynham ( 1981) employed the level slicing of the difference of a visible band to map cutovers. This is a simple method but does not consider the local effect of atmosphere as mentioned below.

Assume on the value of Band 2 that, within a small part of the Images, the non – damaged pixels take values within narrow range around a central value in November corresponding to the value in April. The central value of the non-damaged pixels’ distributions can be replaced with the central value of whole forests pixels’ distributions if the number of the non-damaged pixels is much more than the non-damaged ones and exceed a threshold value.

It has been reported that the maximum of the damaged forest rations to the private forest in a municipality is no more than 34% ( Forest Agency, 1992). Then, the central values of the non-damaged pixels distributions can be replaced with a robust central value of whole pixels’ distributions.

Local effect of atmosphere to the images
Northern Kyushu has steep mountains in the center and is surrounded by three different seas; the Korean Strait, the Setonaikai ( the Inland Sea ) and the Ariake Bay. The atmospheric condition of a region of the images is, then, rather different from the other even if there is no clouds on the region. Therefore, the haze correction to Band 2 can not be done uniformly all over the images.

Within a small part of the images, it is possible to assume that the atmospheric effect is constant and each pixel’s value of a band contains a same offset of the haze. Then, the distributions of the values are not affected by atmosphere.

The Method
The damaged pixels are detected according to the following procedures.

The images are divided into parts as small as 256 columns by 256 lines ( the divided process ) . For a comparison, the whole images are processed at once by following procedures instead of the parts of the images ( the full process )
From the pixels of the forest areas in each part of the images ( or the whole images ), a two dimensional histogram of Band 2 of the April image vs. that of the November image is created. The third moments of the distributions of values of the November image for every values of the April image are calculated as the center of the non-damaged pixels’ distributions. The third moment is calculated as follows:

Mi = å( nij³* j ) / å nij³

where, i, j: values of Band 2 in April and in November respectively, Mi : the third moment for i, in April and j in November.
A damaged pixel is defined as the pixel whose value of Band 2 is i in April and larger than 1 + Mi in November.
The result image is compiled to the damage intensity image on the ¼ mesh of the Digital National Land Information. It contains the areas of forest, the area of damaged forest and the ratio of the damaged forest within a mesh, which is equivalent to a corresponding set of 25 pixels of the result image.

Results and Discussion
The accuracy of the change detection for the sample points in the Kusu and Kokonoe region is shown in Table 2. Within this region, the damages are detected almost accurately but rather underestimated. One of parts of the result images by the both way are shown in Figure 4. In the mountainous part, the detected areas are nearly equivalent. But, in the edge of the plain, almost the all forest pixels are classified as damaged by the full process while the detected pixels is rather few by the divided process. The damage is scarcely reported in this region.

Figure 4. A example of the difference between divided and full processes Detected areas are presented in white
Figure 5 shows the ratio of the damaged forest on the Digital National Land Information’s system. Overlaying this image onto the DEM and the municipality map, the damage density map and the ratio of the damaged forest by aspect are calculated for each municipality ( Figure 6 ) . Because there is no reliable ground data over the areas, the accuracy of Figure 6 can not be evaluated, but it is obvious that the damage is superior in south slope.

Figure 5. The image of the ratio of the damaged forest. The darker points damaged more severly

Figure 6. The damage density for each municipallty and the ratio of the damaged forest by aspect
As for the single visible band for the change detection, Banner and Lynham ( 1981) used the LANDSAT MSS band 5 to detect the cutovers. They explained the reason why the vegetation index using infrared bands are not available for the detection that the index is sensitive to the growth of vegetation within the cutovers. In the case of the windfall damage, the reason why the infrared band had not reduced may be that the crowns of the fallen trees still remained and alive on the forest at the time observed and evern after a while.

Conclusion
The change of the red band of the MOS MESSR is crucial to detect the windfall damage of forest. No obvious change can be seen in the infrared bands of the damaged forest. The reason may be that the crown of the damaged trees remained on the forest and the leaves are left alive.

Dividing images into small parts and the atmospheric effect which is varied by locations can be assumed to be constant within each part, then the change detection is carried out with the own criterion within each part.

Overlaying the image of the damaged forest onto the DEM and the municipality data, it is confirmed that the damage is superior in south slope.

Acknowledgement
This project is partly supported by Science and Technology Agency of Japan.

References

Awaya, Y. et al, 1986, Detection of cut-over area by LANDSAT data (V) Application of change vector analysis, Transactions of the 97th Annual Meeting of the Japanese Forestry Society ( in Japanese )
Banner, A.V. and Lynham, T., 1981, Multitemporal analysis of LANDSAT data for forest cutover mapping: a trial of two procedures, Proceedings of the 7th Canadian Symposium on Remote Sensing.
Forest Agency 1992, The record of forest disasters caused by the typhoons in 1991, Nihoon Zorin Kyokai ( in Japanese )
Kaufman, Y.J., 1989, The atmospheric effect on remote sensing and its correction, Theory and Applications of Optical Remote Sensing, Wiley Intersciency, pp. 336-428.
Meteorological Agency, 1991, Geophysical Review, No. 1105, pp. 30-36 ( in Japanese )
Onuki, l. et. al, 1982, Detection of cut-over area by LANDSAT data ( II) Application of principal component analysis, Transactions of the 93rd Annual Meeting of the Japanese Forestry Society ( in Japanese )
Yamamoto, H., 1992, Agricultural damages in Kyushu by typhoons 9117 and 9119, Journal of Agricultural Meteorology, 48 (1), pp. 77-83 ( in Japanese )