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

Remote Sensing for forestry applications In the tropical rain forest of Peninsular Malaysia

Yoshio Awaya, Itsuhito Ohnuki,
Harrrrruo Sawada, Osamu Nakakita
Forestry and forest products research institute of Japan
P.O. Box 16, Tsukuba Norin Kenkyudanchi-nai, Ibaraki, 305 Japan,

Khali Aziz b.Hamzah
Forest Research Institute Malaysia
Karung Berkunci 201, Kepong, 52109 Kuala Lumpur, Malaysia


Abstract
Remote sensing is an advantageous technology for forest resource management. Remote sensors send periodical image data, which can be used for various types of forest resource evaluation such as harvest monitoring and forest type classification. However clouds with tropical rain forests in the image data obstruct remote sensing. In the case of utilizing the visible and infrared radiometer, cloud free data might never be obtained from these areas and finding effective usages are now primary research topics.

This study concerns cutovers' detection and effective display techniques for displaying both original Landsat Thematic Mapper (TM) data and processed images, which show changes. As digital image processing has limited ability in accurately classifying forests both in the sun and shade and in accurately distinguishing cutovers from clouds, enhancement and display techniques for image interpretation are primary research themes in the tropics. Histogram normalization and Wall is enhancement was consecutively executed to provide contrast stretching in forests. TM channels 4,5 and 1 (TM 4,5,1) (red, green and blue in color respectively: RGB) made a very good combination for enhancing forest types and about 7 different types could be interpreted. Then principal component analysis (PCA) was executed to enhance images of cutovers using a combination of SPOT High Resolution Visible (HRV) and TH images, and normalized vegetation index (NVI) images of the HRV and TH respectively. Both results were adversely affected by clouds and their shadows, though the obtained with NVI had greatly enhanced cutover with reduced topographic effects. Compartment superimposed composites made it possible to precisely locate each forest cutover and they can effectively supplement the forest inventories on the ground.

Introduction
Forestry is one of the important industries of Malaysia and sustained timber harvesting is a major problem there. Forests are usually managed by the selective harvesting system. Forests are harvested every 30 or 55 years (MPI 1988) and their regeneration is a primary concern. "Harvesting has spread widely and it seems difficult to check exact harvested locations, extent of disturbance and growth. Due to poor accessibility and inadequate forest information, satellite remote sensing is a desired technology to be used in place of field checks (Wan Yusoff 1988)". However, since the appearance of clouds in combined NOAA Advanced Very High Resolution Radiometer (AVHRR) images using three weeks' data was reported by Malingreau (1986), it has been found that clouds obstruct practical remote sensing applications in the tropics. Since forest monitoring by digital image processing leads to numerous errors in the tropics producing interpretative photo products is a simple and practical way to utilize remote sensing data for forestry applications (Awaya et al. 1988).

As the topography of Peninsular Malaysia is very sleep, shadows may lead to misinterpretation, and also the varying sun elevations and azimuth angles of each image make complex mosaics of shadows in combined images. The objective of this paper is to determine how to reduce shadow effects and to enhance images of forest types and selectively harvested fporests in Malaysia by image enhancement techniques, i.e. histogram normalization, Wallis enhancement, NVI and PCA. Forest compartment boundaries are superimposed on the enhanced images to make it easy to interpret locations for forestry usage.

Study Area
The Ulu Selangor to Gombak area, which is adjacent to the north of Kuala Lumpur and covers an area of 25km by 25 km, was selected as a study area (Figure 1,2). The topography of the area various greatly. The western part is hilly with widespread oil palm and rubber plantations, and there are secondary forests in the north. The eastern part, which is from Batang Kali (Batang Kali, Serendah forest reserve and so on) to the Gentling Highlands (Ulu Gombak protoected forest, Ulu Gombak forest reserve), is mostly a steep mountainous area with elevations up to 1770 m above sea level and it is covered with forest.

The vertical distribution of the natural forests of Malaysia is shown in table 1 (Wyatt-Smith et al. 1963). Natural or primary forests including the Upper Dipterocarp Forest and Montane Forests are founded in the Genting Highlands and Ulu Gomak protected forest. Forests have been classified into eleven broad forest types by 19811-82' second global forest inventories organized by FAO (Wan Yusoff, 1988). According to the classification, forests are divided into two groups, based on elevations above and below 1000 m, i.e. unproductive and productive forests. They are further divided by stand volumes and species. Productive virgin forests were classified at the foot of the Genting Highlands by prescribed inventories. There are disturbed forests in the southeast of the study area, which have been harvested since 1988 and logged-over forests are common, though harvest years vary from the mid-70's up to present. Forests harvested prior to 1980situated in the south east and recently harvested forests in the east of the study area. Some forests have been converted forests in the east of the study area. Some forests have been converted into agriculture lands recently (Table 2, figure 2,5,8). Average timber volumes harvested in Batand Kali are about 15 tons (wet weight) par hectare.

Materials and Equipments
SPOT HRV (6 July 1987, path 269, row 343, @CNES 1987) and Landsat. TM (29 February 1988, path 127, row 58) images were used in this study. Because of rough ground resolution, TM 6 was not utilized. About 20% of both images were covered with clouds and the HRV image did not cover about 220% of the eastern study area. The sun elevation and azimuth angles of the TM and HRV images were 51o 35 and 6o , 46', and 60o 08' and 46o 54 respectively. Differences in the sun's location was a factor leading to errors in change detection.

Topographic maps (scale 1:63,500) with forest compartment boundaries were used for selecting ground control points, making a digital elevation model and digitizing compartment boundaries.

A Fujitsu M340 s computer was used for basic data processing, and as an International Imaging System's (1Z S ) system 600 image processing system on a Toshiba AS3160HM computer (OEM of Sun 3-160) was used for advanced data processing. A drum scanner produced by the Sekkej inc. was used for digitizing maps

Method
  1. Preprocessing


    2. Geometric correction was executed on Landsat TM data with pixel size of 2225 m using 26-ground control points on the longitude and latitude coordinate. The HRV data were registered to the TM data using 19 control points. Second-order polynomials with nearest neighbor resembling was used for both corrections.

    Elevations were digitized with pixel size of 200 m and registered to the images using bi-linear interpolation. Then 1000 m contours were delineated. Forests compartment boundaries were digitized by the drum scanner, then they were thinned and overlaid on the images.

  2. Image Enhancement


    3. Desirable channel combinations were checked for making forest-interpretative color composites of the TM data by transformed Divergence (TFD) (Swain et. al. in press). Mean values of the TFD were used for selecting the best combination of three channels (Table 3). Categories used in the analysis were wamp forest, clear-out forest, four different stages of rubber, Acacia mangium, three different stages of oil palm, pine, logged-over forest, secondary forest, secondary to primary forest and primary forest. Then histogram normalization referring to vegetation training areas and Wall is enhancement were consecutively applied to several channel combinations in accordance with the former results. Tested combinations were TM 4,5,1, TM 4,5,2,Th 4,5,3 and TM 4,5,7 (RGB) (Figure 4).

  3. Change Detection


    4. The NVIs of TM and HRV, along with a normalized difference ratio (TM5-Tm4)/(TM5+TM4) (NDR54) image were produced after subtracting the minimum value of each channel for adjusting the origin of two dimensional space (Crippen 1988). The NDR54, after histogram normalization was processed, was used supplementary to make a color composite.

    Two image processing were executed to enhance forest changes, especially cutovers, using the TM and HRV data. The first was, PCA using TH 2,3,4 and HRV 1,2,3 (pair-image PCA). The second was PCA using TM and HRV's NVI (NVI-PCA). Then several color composites were examined and selected combinations were the 4th, 3rd and 2nd components (RGB) for the pair-image-PCA, and the 2nd component, Ist component and NDR54 (RGB) for the NVI-PCA. Wall is enhanced TM 3,4 and HRV 2 (ORG, in RGB) was used as a reference image. Cutover means a logged-over forest between observed dates of the two images in this paper.

  4. Superimpossion


    5. Forest compartment boundaries, compartment numbers and 1000 m elevation contours were superimposed with the color composites and their effectiveness was examined.
Discussion
  1. Image enhancement


    2. Several TM channel combination studies have been reported (Awaya et al.1985 etc.) and the common results have been that combinations including TM 4 and 5 were highly ranked for effective channel selection. the TFD analysis showed almost the same results, but TM 7 seemed to be more effective than other studies' results (Table 3). Miwa also mentioned TM 4, 5and one other channel in coloring red, green and blue as an adequate color combination. The color composite usually shows vegetation differences very well.

    Histogram normalization was effective in enhancing less varied channels in forests and various color tones were recognized. Then Wallis enhancement contrasted borders of different forests (Figure 3)in all channels and improved color balance of the composites very well. As Wallis enhancement is a space variant function which enhances pixel by means and variances of small local windows, clouds badly effect its application for original TM data. Thus pre-enhancement was necessary to reduce the effects.

    The effectiveness of the third channel was also checked (Figure 4). The Ulu Gomak protected forest was strongly enhanced by TM 1 (dark at bottom center) and it contrasted very well with the forests on step slopes in the Genting Highlands (medium brightness at upper right) or secondary forests (slightly dark at middle left). On the other hand, the brightness differences were reduced in other channels as wave lengths were longer.

    The color composites were compared with each other (Table 6) using the compartment superimposed products, which made it possible to distinguish forest types in each compartment (Figure 5, Table 2). The most colorful and interpretative composite was TM 4,5,1. Basically colors were identified in the forests in the TM 4,5,1 image, but they became less visible as the wavelength become longer in the blue colored channel. A ground truth study must be executed to make sure of the relationships between the colors and forest types. The compartments made it possible to identify forest types at exact locations which is useful information for forest management.

  2. Change Detection


    3. Change detection using two images under very different solar angle is extremely difficult task because of various shadows effects in mountainous area. Detailed spectral adjustment between two corresponding channels such as using linear regression (Fogelman 1988, Ohnuki et al. 1988) was nearly impossible without topographic correction.

    PCA is a very popular transformation for reducing data dimensions and enhancing images, but it is a data dependent transformation and produces various results case by case. Thus the training area which is desired to be enhanced should be appropriately selected for PCA. However PCA has the following advantages. 1) As the first component has the greatest variance of data space, the lower components might be less effected by topography. 2) Although a linear regression cannot express the longer axis of data distribution in a low correlated two dimensional space, the first component is robust enough to express it and the longer axis usually shows the most common relationship between the samples. 3) PCA using a correlation matrix reduces the difference of sensor sensitivity.

    Vogelmann (1988) pointed out that "A drawback of PCA is the difficulty of relating the results to reflections change for individual bands for the areas undergoing change." As he indicates, slight changes might be undetectable but reflection changes distinctly after logging roads, logging yards and badly damaged forests are detectable by PCA.

    The 1st component of the pair-image PCA showed the total brightness of the images, but it was effected by baze and topography, and this was why the component wasn't used in the color composite. However the topography became invisible with the lower component. The 2nd component enhanced general differences between the two images and the 3rd component displayed vegetation. Bare soil and cutovers were also enhanced in the 3rd and 4th components respectively (Table 4, Figure 6).

    Rationing is a well known technique for removing shadows caused by topography (Crippen 1988), and NVI was performed instead of direct rationing to enhance non-vegetated areas, and topographic effects were invisible in the two NVI of TM and HRV. Cutovers appeared dark in the 1st component and bright in the 2nd component of NVI-PCA, but clouds is the TM data were also enhanced in the same way. It seems that change detection would be impossible, if clouds and their shadows should not be removed at data processing or be superimposed on the final results. Rare soil such as logging roads was contrasted with cutovers in the 2nd component with dark brightness (Table 5, Figure 7).

    The color composites were compared (Table 7). NVI-PCA strongly enhanced cutovers and bare land such as logging roads, and topographic effects were almost completely removed. However cutovers and bare land could not be distinguished from clouds and their shadows and it was almost the same in the pair-image PCA composite. Rapid vegetation growth was distinguished, but topographic effects were visible in the pair-image PCA. Colors of cutovers were varied in the ORG composite, but clouds and their shadows were distinguishable.

    As compartment superimposed products showed exact harvesting locations and neighborhood conditions, they could be effectively used for change detection (Figure 8). For example, as legal harvesting was executed by private companies with licenses (MPI 1988), illegal harvesting would be detectable easily by checking each cutover's location and extent.
Conclusion
A combination of histogram normalization and Wallis enhancement was very effective for making color composites for forest type interpretation. TM 4,5,1 was the best composite for distinguishing forest types were distinguishable. Compartment boundaries and 1000 m contours superimposed products were very useful to ascertain existing forests and logged-over forests. They can be used supplement of forest inventories on the ground.

Change detection using cloud-affected data was terribly difficult. Clouds and their shadows had to be superimposed to show their extent on the color composites in an identifiable way. NVI-PCA was effective in enhancing cutovers without topographic effects, but vegetation growth was not distinguishable.

Further studies using multiple satellites' data must be execute for excluding cloud obstacles and image enhancements would have important roles in remote sensing of the tropics.

References
  1. Awaya, Y., Ohnuki , I. and Sawada, H., 1985, Channel Selection of Landsat TM Data for Forest Type Classification, Proceedings of the 5th Annual Conference of the Remote Sensing Society of Japan,pp77-80, in Japanese _______, ________, _________, 1988, Change Detection in Tropical Rain Forests using LANDSAT MSS data, Transactions of the 99th Annual Meeting of the Japanese Forestry Society, 'pp. 105-106, in Japanese


  2. Crippen, R.E., 1988, The Dangers of Underestimating the Importance of Data Adjustments in Rand Rationing, International Journal of Remote Sensing, 9, pp.767-776.


  3. Malingreau, J.P., 1986, Global Vegetation Dynamics: Satellite Observations over Asia, International Journal of Remote Sensing, 7, pp.439-452.


  4. Ministry of Primary Industries Malaysia (MPI), 1988, Forestry in Malaysia, Inventra Print, pp. 67.


  5. Miwa, T., Uchara, S., Ikeda, T. and Asahi, Y., 1988, Investigation on color composite Images of TM Data (II) Effects of Band Combination and Color Sequence on Visual Interpretation, Journal of The Remote Sensing Society of Japan, 8, pp. 113-130, in Japanese.


  6. Nakakita, O., Sawada. H., Awaya, Y. and Khali Aziz, in press, forest Classification of Peninsular Malaysia by Remote Sensing Data, Transactions of the 100th Annual Meeting of the Japanese Forestry Society.


  7. Ohnuki, I., Awaya, Y. and Sawada, H., 1988, Development of a Monitoring System Model for Tropical Forest Management using Satellite Remote sensing, Proceedings of the ISPRS and IUFRO Joint Session, Kyoto, pp.13-22


  8. Swain, P.H., Davis, S.M. et al., 1978, Remote Sensing: The Quantitative Approach, McGraw-Hill Inc., pp396


  9. Sheffield, C., 1985, Selecting Band Combinations from Multispectral Data, Photogrammetric Engineering and Remote Sensing, 51, pp. 681-687


  10. Vogelmann, J.E., 1988, Detection of Forest Change in the Green Mountains of Vermont using Multispectral Scanner Data, International Journal of Remote Sensing, 7, pp. 1187-1200


  11. Wan Yusoff Wan Ahmad, 1988, Use of Satellite Remote Sensing for Updating Forest Resource Information in Peninsular Malaysia, Presented at Asean Seminar: ASEAN Institute of Forest Management and Forest Department Peninsular Malaysia, pp. 17


  12. Wyatt-Smith, J., Panton, W.P. and Mitchell, B.A., 1963, Manual of Malayan Silviculture for Inland Forests, Malayan Forest Record 23, Forest Research Institute Malaysia
Table 1 Vertical distribution of Natural Forest at the study area (by Watt-Smith, 1963)
Elevation
(meter) 		Forest Type Common Species at upper story
and tree height
- 300 Lowland Dipterocarp Forest Dipterocarpaceae 30-60m tall
300-750 Hill Dipeterocarp Forest Especially Shorea curtissi
750-1200 Upper Dipterocarp Forest Shorea, Calopyllum spp. 24-30m
1050-1500 Montane Oak forest Fagaceae, Lauraceae spp. 18-24m
1500- Montane Ericaeous Forest Sphagnum spp. 12m


Table 2 Latest Forest Harvesting History (Batang Kali Forest Reserve)
Year Compartment Number
before '60 1, 2, 4*, 5*, 6
'60-69 3*, 12
'70-79 10a, 10b, 13a, 17-24**, 28b**
49a-c, 50c
'80-85 42,43a-b, 44a-b,45,46,48
50a-b, 51, 52, 53, 54, 56,
57b-c, 58a-b, 58c, 58e
'86-88 7, 8, 9, 52b, 57a, 58-d,
Reserve 26-41

* harvesting since '88
** converted into agriculture land Part of each compartment was logged over

Table 3 Results of transformed Divergence(Nakakita et al. in press)
Channels Divergence
1 2 3 4 5 6 7 1893
1 4 7 1887
2 4 5 1867
2 4 7 1855
1 5 7 1855
3 4 5 1855
1 2 5 1850
1 3 5 1847
1 3 4 1840
3 4 7 1839

* Top tens are listed.

Table 4 Results of Principal Component Analysis using HRV 1, 2, 3 and TM 2, 3, 4
Principal Component
  1st 2nd 3rd 4th 5th 6th
Enginvalues 3.2612 1.1828 0.9166 0.3834 0.1481 0.1079
Cum. Proportion 0.5435 0.7404 0.8934 0.9573 0.9820 1.0000
1987 HRV 1 0.4276 0.5221 0.0335 0.0798 -0.7048 0.2010
Eigen-Vector HRV 2 0.3869 0.5634 -0.2390 0.2463 0.6297 -0.1365
HRV 3 0.3275 0.0682 0.7699 -0.4960 0.2060 -0.0831
TM 2 0.4751 -0.3539 -0.2315 -0.0541 -0.2029 -0.7425
TM 3 0.4312 -0.2884 -0.4372 -0.4712 0.1302 0.5484
1988 TM 4 0.3855 -0.4438 0.3230 0.6797 0.0791 0.2863


TABLES 5,6,7 ARE MISSING


Figure 1 Location of the the study Area
Figure 2 Landsat image of the Study Area Montane forests and disturbed forests. Genting H. is at the right edge.

Figure 3 Comparison of Histogram
(a)TM 3 Normalized   (b)TM 3 Wallis
(a)TM 3 Normalized   (b)TM 4 Wallis
Figure 4 Comparison of Blue colored channels
(a)TM 1, (b)TM 2,(c)TM 3, (d)TM4
Dark place at the bottom center is Upper Dipterocarp to Montane For

Figure 5 Compartment Imposed Composite
The 1000m contours are also imposed. Each number is a compartment No.
Dark area seemed still good forests.
Figure 6Results of Pair-Image PCA
(a)1st,(b)2nd(c)3rd and
4th component white spots at the center are culovers.

Figure 7HRV 2, TM 3 and NVI-PCA
(a)HRV 2, (b) TM 3, (c) 1st and (d)2nd component Dark areas are none vegetation in(c)While areas are culovers, clouds and their shadows in d
Figure 8 Compartment Superimposet Products for Change Detection Logged-over forests are in comp. and 6. Cultivation at converted farm lands are comp. 15-24.