J. For. Res. 6 : 1 3 9 - 1 4 6 (2001)

Vegetation Inventory of a Temperate Biosphere Reserve in China by Image Fusion of Landsat TM and SPOT HRV Qi-Jing Liu I and Nobuo Takeuchi Center for Environmental Remote Sensing, Chiba University, Chiba 263-8522, Japan. Vegetation cover types on Changbai Mountain, a nalural biosphere reserve (2,000 kn]:) in northeast China, were derived by using multisensor satellite imagery fused with Landsat TM and SPOT HRV-XS. DEM data were used for finproving classification accuracy. Cover lypes were classified into 20 groups. Bands 4 and 5 of Landsat TM image acquired on July 18, 1997, and band 1 of SPOT HRV-XS image acquired on Oct. 19, 1992, were fused Io a false color image, and maximum likelihood supervised classification was perfi~rmed. Data fusion showed high accuracy of idenfilication, compared to individual images. The overall accuracy of classification of individual images by SPOT HRV-XS reached 56%, and TM 66%, while the fused data set provided accuracy of about 78%, which was raised to 81 ~7~after recoding by using DEM. There were five vegetation zones on the mountain, from the base to the peak: hardwood forest zone, mixed forest zone, conifer forest zone, birch {brest zone, and tundra zone. Spruce-fir dominated conifer tbrest was the most prevalent (nearly 50~}) vegetation type, tollowed by Korean pine and mixed forest (17%) and larch forest (5%). HRV image taken in leaf-off season is useful for discriminating forest from non-forest, and evergree~ forest from hardwood forest, while the summer image (TM) provides detailed inlormation on the difference in similar vegetation types, like hardwood forest with different compositions. Key words: Changbai Mountain, Landsat TM image, multisensor, SPOT HRV-XS, supervised classification

Changbai Mountain is one of the world's very important biosphere reserves, a still-pristine natural ecosystem high biodiversity. Intensive studies on the vegetation by traditional field investigation have been carried out since the 1980s. Research on the vegetation distribution and dynamics by remote sensing, however, amounts to very little. As basic data for the study and management of the biosphere, an inventory of natural resources by satellite remote sensing is very time- and cost-effective, especially for regional scales and topographically inaccessible areas. With high quality image data, vegetation can be precisely detected. Accuracy of classification by imagery can be affected by many factors, e.g., stand age (Mickelson et al., 1998; Abuelgasim et al., 1999), understory structure (Spanner et al., 1990; Curran et al., 1992: Joffre and Lacaze, 1993: Nemani et al., 1993), tree height, live basal area, leaf area index (LAD, tree size (Nilson and Olsson, 1995; Jakubauskas, 1996), and density or timber volume (Gemmell, 1995; Trotter et al., 1997), etc. On the other hand, classification accuracy can be significantly improved by using multi-season imagery (Jeon and Landgrebe, 1999: Kurosu et al., 1999) and multi-sensor data (Gong et al., 1994; Dai and Khorram, 1998; Toutin, 1998; Haack and Bechdol, 1999: Zhukov et al., 1999, Michelson et al., 2000), because data from different sources and seasons have different separability on cover types, and the complementation can improve the accuracy of identification. Assistant data like DEM (digital elevation model) are also useful, because vegetation distribution, especially natural communities, is closely related to elevation and the relationship can be used for verifying classification results. A previous report (Shao et al., 1996) about the vegetation mapping of the target area of this study was based on ISOI Corresponding author (E-mail: [email protected]).

DATA unsupervised method with single-date imagery of Landsat TM. But the reliability of the results could be improved with more robust data, because the accuracy by single-date imagery is poorer than multi-temporal data, and unsupervised classification is less realistic than supervised classification. For example, the evergreen conifer forest was grouped in one category, when in fact there are at least two types, pine dominated and spruce dominated stands. Theretore, additional data like DEM and phenological information from different seasons are expected to prove efl'ective for improving discrimination. This study is the first attempt to use fused data from different sensors of TM and HRV, with DEM data, for cover types detection for the Changbai Mountain Natural Reserve.

Study Area and Methods 1 Study area The study area was the Changbai Mountain Natural Reserve which is located on the border between China and North Korea, and the coordinates are 127°42 ' E/41 °41' N- 128 ° 17' E/ 42o26 ' N (Fig. 1). The area of the reserve is about 2,000 km 2. The ranges of the reserve are 60 km south-north and 40 km east-west. The highest elevation is 2,734 m above sea level (a.s.1.). The vegetation is vertically divided into five zones, and there are three forest zones inside the reserve, i.e. ( 1) broadleaved-conifer mixed forest zone ( 6 0 0 - I , 100 m a.s.1.), dominated by Korean pine (Pinus koraiensis), bass (Tilia amurensis), maple ( A c e r m o n o ) , elm ( U h n u s p r o p i n q u a ) etc'., (2) conifer forest zone ( 1,100-1,700 m a.s.l.), dominated by spruce (Picea jezoensis vat. komarovii, P. koraiensis) and fir (Abies nephrolepis), and (3) mountain birch forest zone (1,700-2,000 m a.s.l.), dominated by mountain birch (Betula ermanii), except for the east slope which is occupied by larch (ixwix olgensis). The evergreen component increases towards

140

J. For. Res. 6 (3) 2001:

Fig. 1 TM image (TM4 of July 18, 1997) of the Changbai Mountain Natural Reserve. Elevation of thc peak is 2,740 m a.s.I. The mountain is a dormant volcano, with a crater in which the water is 347 m deep. The line in the image shows the boundary of the natural reserve of Changbai Mountain.

the upper limit ( 1,700 In a.s.I.) of the conifer forest zone, and disappears in the mountain birch forest zone (Liu et al., 1998). In lower elevations, secondary vegetation of poplar birch forest, dominated by poplar (Populus davidiana) and white birch (Betula plaOT~hylla), and oak (Quercus mongolica) is very common. The hardwood forest is mainly mixed hardwood forest and birch-poplar forest. Oak forest is also frequently encountered in the lower areas, which is a secondary vegetation alter repeated cutting.

Changbai pine (Pinus syh,estrifornlis) (an endemic species of the reserve) forest has a small area, but shows a wide range of distribution, from 700 to 1,400 m a.s.1, on the northern slope.

2 Data A Landsat TM image (path 116 row 31) acquired on July 18, 1997, and a SPOT HRV on Oct. 19, 1992, which were provided by NASDA (National Space Development Agency of Japan) were used for cover type identification. For the SPOT image, the path was 302, and row was 265-266 (two scenes). The parameters of TM and HRV sensors are shown

Liu and Takeuchi

141

Table 1 Band parameters of Landsat TM and SPOT HRV sensors. Band

Color

TM B 1 TM B2 TM B3 TM B4 TM B5 TM B6 TM B7 HRV B 1 HRV B2 HRV B3

Blue Green Red Near infrared Short wave infrared Far infrared Short wave infrared Green Red Near infrared

Wave length (,um) 0.45-0.52 0.52-0.60 0.63-0.69 0.75-0.90 1.55-1.75 10.40-12.50 2.08-2.35 0.50-0.59 0.61-0.68 0.79-0.89

in Table 1. The sun elevation and azimuth for HRV path 302 row 265 were 37.8 ° and 168.8% respectively, and that for path 302 row 266 were 37.3 ° and 167.6% respectively. The pointing angle of HRV was 1.96 °. For TM image, the sun elevation was 59.67% and azimuth was 123.91 °. The two scenes of HRV were mosaiced into one image to cover the whole study area. The data processing status for both sensors was level 2. In this paper, bands 1-7 of Landsat TM are abbreviated as TM1, TM2, ..., TM7, respectively. Similarly, bands 1-3 of SPOT HRV are called HRV1, HRV2, and HRV3, respectively. Spectral radiance, in the form of digital value, was picked up from 20 sample points, one point for each vegetation type. The radiance of each sample point was expressed by the average value of 9 pixels (3 by 3 pixels) for the TM image, and (4 by 4 pixels) or (5 by 5 pixels) for the HRV image, consistent with the TM image in the ground area. A 1 : 100,000 scale contour map was scanned with a drum scanner for geometric correction. The source map, which was published by the former Soviet Union in 1982, is deposited in Gifu Prefectural Library World Distribution Map Center (http://www.smile.pref.gifu.jp/map/index.html). The contour map was also digitized with a kind of software called ATLAS GIS (Q + E Software, Inc.). The digital map was then converted into ASCII format, and a grid file was created by the Minimum Curvature method with SURFER (Golden Software Inc.), and the grid size was set to correspond with the satellite imagery. The generated data set (DEM file) was imposed onto the classified image for verifying cover types. This was expected to be meaningful for distinguishing the different types of vegetation, especially conifer forest, which are similar in spectral radiance. Image processing software programs for the analysis were PCI (PCI Ltd.) and ER Mapper (Earth Resource Mapping Pty Ltd.). Geometric correction was performed by the fourth-order polynomial method, and the root mean square (RMS) error was controlled within one pixel. More than 100 ground control points (GCP) were picked from each image of the two sensors. Radiometric correction was not performed. The dimensions of the TM data were 1,995 pixels by 2,842 lines, and those of the same area for the HRV image were 2,993 pixels by 4,263 lines (resolution 20 m by 20 m). Data fusion was pertbrmed on the two sensors. HRV

imagery has a spatial resolution of 20 m by 20 m. TM imagery was resampled from the original 30 m by 30 to 20 m by 20 m in grid size by the nearest neighbor method in order to be overlaid with the SPOT HRV image. NDVI (Normal Difference of Vegetation Index) was calculated with the equation (TM4 -- TM3)/(TM4 + TM3) for the TM image, and (HRV3 -- HRV2)/(HRV3 + HRV2) for the HRV data. To avoid real numbers, NDVI was scaled to a range of 0 - 2 0 0 . As to the utilization of DEM for classification, the generated data set (DEM file) was imposed onto the classified image for recoding cover types. This was expected to be meaningful for distinguishing the different types of vegetation that have a close relation with elevation. Based on our knowledge about the vegetation distribution, the classified image was recoded by the DEM. The elevational thresholds of cover types used in the algorithm, however, did not strictly adhere to the actual ones because of the existence of ecotones. For example, the upper limit of Korean pine forest (class 8) is determined at 1,100 m, but patches dominated by this species at higher elevations can be often found. Such patches were rare above 1,200 m. Theretbre pixels with elevation higher than 1,200 m (not 1,100 m), which were classified as Korean pine forest were included in the class of conifer forest (class 2). The algorithm for recoding the classification is shown in Table 2 and Fig. 2. 3 Ground data Ground data were obtained by field investigation. More than 100 plots were surveyed, including some permanent ones. The detailed inlbrmation about plant communities was

Table 2 Regimen for recoding cover types in the classified image with DEM. Initial 1 2* 3* 4* 5* 6 7* 7* 8* 9 10 11 12' 13 14 15 16" 17" 18 19" 19' 20

Threshold (m) >2200 < 1100 1200-2000 < 1100 < 1100 > 20O0 1200-1400 1400-2000 1200-2000 >2000 > 2000 > 2000 < 1400 No recoding No recoding < 1500 1200-2000 1400-2000 > 2000 800-1100 1100-2000 >2000

Recode 15 8 5 3 3 15 17 12 2 15 15 15 17 13 14 14 13 12 15 8 2 15

Initial, class code by initial classification: Recode, class code after recoding: * Recode to class 15 (tundra) if elevation > 2,000 m. See Table 3 for class codes.

142

J. For. Res. 6 (3) 2001 Ilnitial

code7 I ~ 1 ~

160

~

140 120 100

so o

I,

SPOT-HRV1,,2/10/1,

i o Water - ~ Conifer forest ""~. /~ - ¢ - - Mixed forest "11.~ s~ - - X - - Har~ood forest "~;i • - +- - LarGE forest "~/~& • ~ Mountain birch forest tjjl~'ll - - e - Rock and semi desert 1;7 --II-- T~dra '7 /

60

.......

.--~_.~" ,/ / / / /

-,--

40 20

No

0

Fig. 2 Recoding of classified image by using DEM. The case of code 7 (hardwood forest) is shown• Code 12. mountain .birch forest; Code 15, tundra; Code 17, poplar birch forest. recorded for three layers, i.e., tree layer, shrub layer, and herb layer. Tree diameter at breast height (DBH) was recorded with plot sizes ranging from 10 m by 50 to 100 m by 100 m. For shrub and herb layers, height, density and coverage were measured with quadrat sizes of 2 m by 2 and 1 m by 1 m, respectively. 4 Classification procedure According to the results of ISODATA unsupervised classifications, which were performed with several data combinations of the two sensors, the maximum likelihood supervised classification was applied for image analysis. Image separability was evaluated by radiance distribution curves, as bands that showed larger differences between cover types are considered meaningful for classification. The final vegetation map output was based on the overall accuracy of classification, i.e. the one which had higher accuracy. Based on the vegetation composition in and around the reserve, it was grouped into 20 cover types with supervised classification. Typical areas of each category were selected, mainly on the China side, as training areas for the classification. Similarly, testing areas were sampled for accuracy assessment. The accuracy of category i was represented by pi : r t i / N i , and the overall accuracy was po ~ani/Ni, where ni is the number of pixels correctly identified, and Ni is the number of pixels of the category in training area. i = 1,2, ..., C, where C is the number of categories (C = 20 in this study). After recoding with DEM, the classified image was smoothed with the median filter of 9-by-9 window to remove noise or tiny points. =

Results By comparing the difference in spectral radiance of bands, as well as the results of ISODATA unsupervised classification of different band-combinations, suitable band sets were selected for image composites for the supervised classification. In this section, for understanding the spectral characteristics of the vegetation, radiance of different cover types observed by the two sensors is presented, in addition to the classification results. 1 Landsat TM imagery The TM image was taken in midsummer, when the vege-

I

I

I

I

1

2

3

NDVI

200

I

150

if\,.

,'i

=

~1oo r~ 50

1

2

3

4

5

7

Bands

Fig. 3 Spectral radiance of the main vegetation types on Changbai Mountain in the images of Landsat TM and SPOT HRV. The TM image was acquired on July 18, 1997, and the SPOT HRV image on Oct. 19, 1992. Radiance is represented by digital values• tation was growing vigorously. As a whole, the spectral radiance showed low in TM1, TM3, and TM7, and high in TM4 and TM5 (Fig. 3). TM4 and TM5 showed meaningful for cover types interpreting, because the radiance was significantly diversified among cover types. In TM4, the difference of radiance between maximum and minimum was over 180, and in TM5 it was about 150. Contrary to this, other bands showed less than 50. This pattern reveals the relationship between vegetation type and spectral radiance that, for visible bands (0.4-0.76/.tm), the absorption by pigments of tree leaves is significant, and the radiance was therefore very low. Contrary to this, the NIR band (TM4) showed great transmission and reflection, which is consistent with previous reports (Japan Society of Remote Sensing, 1997). The meadow, consisting of high grass, showed a high level of radiance, and water revealed the opposite• Tundra was similar to meadow in physiognomy, and the radiance was slightly lower than for the meadow. Radiance in the forest vegetation was strongly affected by species composition. The pure Changbai pine forest, sparsely distributed in the reserve, showed the lowest among all forest types; the mountain birch forest presented the highest, and the mixed forest intermediate. The spruce-fir forest in higher elevations was mixed with larch and birch (Liu et al., 1998), and the radiance was consequently higher than that of the pure conifer forest stand.

143

Liu and Takeuchi

Table 3 Relative area and classification accuracy of cover types on Changbai Mountain derived by Landsat TM and SPOT HRV. Type Water Conifer forest Mixed forest Spruce larch forest Spruce birch forest Cut area Hard wood forest Korean pine forest Larch forest Thin conifer forest Windfall and shrubs Mountain birch forest Rock and semi desert Meadow Tundra Urban area Poplar birch forest Young forest Changbai pine forest Thin forest Overall

Code l 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 --

TM

HRV

THl

TH2

TM

HRV

TH1

TH2

100 83.95 23.82 69.68 87.04 44.58 73.73 35.99 69.38 77.73 87.47 86.28 62.00 79.88 64.33 7.94 59.80 69.36 96.14 59.65 66.35

99.95 53.10 31.90 76.59 67.60 4.86 81.27 44.38 12.23 81.54 83.22 90.11 35.64 66.48 78.93 24.78 32.95 42.79 47.97 42.78 55.73

100 85.72 72.76 87.40 77.33 80.22 74.93 61.88 63.72 95.32 91.67 91.16 71.63 82.71 67.99 90.00 68.75 89.30 92.92 43.40 78.10

100 85.54 70.57 87.23 73.59 92.81 82.25 74.02 67.35 95.45 91.18 91.75 72.51 82.63 62.20 79.78 71.52 90.55 92.66 42.60 77.65

100 84.37 27.66 69.68 87.72 44.58 73.73 42.91 69.38 77.73 87.47 86.57 77.98 83.01 64.33 7.94 60.40 69.36 96.14 59.65 68.16

99.95 60.29 41.10 76.59 70.03 4.86 81.27 47.85 12.23 81.54 83.22 90.11 59.38 66.74 78.93 24.78 32.95 42.79 47.97 42.78 59.52

100 89.11 77.60 87.40 80.12 80.22 74.93 69.29 63.72 95.32 91.67 91.16 93.08 87.01 67.99 90.00 68.75 89.30 92.92 43.40 81.02

100 90.07 75.23 87.23 78.35 92.81 82.25 86.52 67.35 95.45 91.18 91.75 93.04 90.27 62.20 79.78 71.52 90.55 92.66 42.60 81.34

Area % 0.2 I 12.19 11.08 6.86 27.18 0.02 3.77 5.96 5.40 2.00 5.31 3.32 1.67 3.94 3.19 0.00 2.19 0.30 0.91 4.50 --

TM~TM2 + TM4 + TM5; HRV, three bands of HRV; TH I, TM4 + TM5 + HRV l ; TH2, TM4 + TM5 + HRV2. Area % is the relative, after DEM recoded, area of cover type in the naturalreserve, which was from TM4 + TM5 + HRV 1. The improvementof accuracy by recoding with DEM is presented.

The difference of vegetation types in spectral radiance is considered to be largely influenced by the stand structure and green leaf biomass or stand volume, according to previous reports (Rey-Benayas and Pope, 1995; Kilpelainen and Tokola, 1999). Broadleaved species usually have higher reflectance than conifer species (Tong, 1990). In this study, for forest vegetation, the radiance (TM4) of mountain birch forest showed the highest, while that of the conifer forest was the lowest. The leaf biomass (dry weight) of the conifer forest was 19 ton/ha, and that of other forest types were less than 6 ton/ha (Li and Li, 1981). It is worth indicating that the vegetation on the higher areas of alpine, tundra and meadow showed high values in the near infrared band, and consequently led to the high NDVI. The conifer forest showed a lower value of NDVI, compared to that of other types of forest. This is consistent with the physiological characteristics of conifer species that their photosynthesis that is generally slower than that of broadleafed species (Crawley, 1997).

2 SPOT HRV imagery This image was taken in late autumn, when the summer green vegetation was leaf off, and there was snow on the alpine zone. Since the phenology was similar to winter, some cover types, which were significantly different in summer, showed the same in spectral radiance. For example, the non-forest area, like tundra, meadow and water (ice), were covered by snow, and the radiance in such areas was therefore very high and similar in appearance. The conifer forest was dominated by evergreen species, which showed a dark green color in the image (HRV3 = red, HRV2 = green, HRV1 = blue), and presented the lowest

value of radiance. In brief, the difference in life-form level between cover types was enhanced in this season. It was clear that the non-forest area showed higher radiance among the selected categories while the high dominance of evergreen species resulted in low radiance, i.e. the mixed forest stood in the intermediate. The radiance in mountain birch forest, however, was even higher than in tundra. This is considered to be caused by the existence of snow. The radiance distribution curves (Fig. 3) showed that, for distinguishing vegetation types, HRV 1 and HRV2 were very similar in separability, and were slightly better than HRV3. Therefore, these two bands were respectively fused with the TM image for classification. NDVI presented a reverse trend that the conifer forest was the highest, while other types of pure deciduous communities showed semblable to each other. The mixed forest presented slightly higher values than the summer green areas. This was apparently due to the content of chlorophyll in leaves that remained unchanged in the evergreen conifer species, while other classes were free of chlorophyll.

3 TM-HRV fusion and vegetation map by supervised classification Landsat TM and SPOT HRV images were used to make a false color image. As mentioned earlier, TM4 and TM5 presented significantly visible differences between objects, and the difference among the spectral bands of HRV image was not so large. Several combinations of the two sensors were attempted for classification. In this paper, the results for TM4 + TM5 + HRV 1, TM4 + TM5 ÷ HRV2 and individual images are described. Classifications on individual images showed that the veg-

144

J. For. Res. 6 (3) 2001:

Table 4 Classification accuracy and misidentificationof cover types on Changbai Mountain after recoding with DEM. Class

1

2

3

4

5

6

7

8

9

10

11

12

13

14

15

16

17

18

19

20

I 100.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 2 0.0089.11 0.02 9.58 16.13 0.00 0.00 0.00 3.14 0.00 0.00 0.00 0.00 0.03 0.00 0.00 0.00 0.00 0.00 0.00 3 0.00 0.00 77.60 0.00 0.00 0,54 3.38 28.93 0.00 0.00 0.22 0.00 0.00 0,00 0.00 0.00 7.86 0.00 0.26 0.62 4 0.00 7.05 0.00 87.40 0.50 0.00 0.00 0.00 1.33 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.52 0.00 0.00 0.00 5 0.00 I).30 4.03 0.76 80.12 0.00 0.00 0.00 0.05 0~00 0.00 0.00 0.00 0.03 0.00 0.00 8.48 0.00 0.00 22.19 6 0.00 0.00 0.00 0.00 0.00 80.22 0.00 0.00 0.00 0.00 0~00 0.00 0.00 0.00 0.00 0.00 0.00 3.48 0.00 0.00 7 0.00 0.00 3.57 0.00 0.00 0.18 74.93 0.00 0.00 0.00 0.49 0.00 0.00 0.00 0.00 0.00 9.73 0.00 0.00 2.17 8 0.00 0.04 1.25 0.00 0.00 0.00 0.00 69.29 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 2.14 0.00 0.92 0.00 9 0.00 0.06 0.04 0.83 0.27 0.00 0.00 0.83 63.72 0.07 0.04 1.35 0.00 0.03 0.00 0.00 0.98 0.00 0.39 0.04 10 0 . 0 0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.0095.32 1.49 0.13 0.00 0.(15 16.29 0.44 0.00 0.00 0.00 0.00 II 0.00 0.00 0.04 0.00 0.00 0.72 0.70 0.00 0.00 2.39 91.67 0.04 0.00 1.05 4.13 0.00 0.00 3.98 0.00 8.37 12 0.00 0.02 0.00 0.00 2.89 0.00 0.00 0.00 31.67 0.00 0.00 91.16 0.00 0.58 4.24 0.00 0.00 0.00 0.00 0.18 13 0.00 0.02 0.00 0.00 0.00 10.07 0.00 0.00 0.00 1.09 0.00 0.0093.08 0.00 2.32 9.56 0.00 0.25 4.72 0.00 14 0.0(1 0.00 0.00 0.00 0.00 0.54 0.00 0.00 0.00 0.89 0.71 0.80 0.00 87.01 5.03 0.00 0.00 2,99 0.00 1.33 15 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.24 0,00 6.53 6.92 2.23 67.99 0.00 0.00 0.00 0.00 0.00 16 0 . 0 0 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 90.00 0,00 0.00 0.79 0.00 17 0.00 0.00 8.97 0.63 0.00 0.00 1.27 0,47 0.09 0.0(I (I.22 0,00 0.00 0.22 0.00 0.00 68.75 0.00 0.00 21.52 18 0.00 0.00 0.00 0.00 0.00 7.55 0.00 0.00 0.00 0.00 1.08 0,00 0.00 0.28 0.00 0.00 0.00 89.30 0.00 0.18 19 0.00100.00 0.00 0.80 0,03 0.00 0.00 0.20 0.00 0.00 0.07 0.00 0.00 0.00 0.00 0.00 0.00 0.0092.92 0.00 20 0.00 0.00 4.48 0.00 0.05 0.18 19.72 0,28 0,00 0.00 4.00 0.00 0.00 8.47 0.00 0.00 0.54 0.00 0.00 43.40 Total 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 100 Source data, TM4 + TM5 + HRV 1. Code numbers represent cover types, consistent with those in Table 3.

etation types with less difference in spectral radiance were very difficult to identify, even for those taken in ideal season. For example, the conifer forest, conifer-mixed forest, and even hardwood forest were partly mis-grouped into one category. On the other hand, the subalpine spruce forest, mountain pine forest, and the forest mixed with other species were similar in spectral radiance, although they were significantly different in species composition, physiognomy and spatial structure. To distinguish vegetation types with high accuracy, multi sensor/temporal image data are expected to prove helpful, because different vegetation may behave the same in spectral radiance in one season, and differ significantly in another season. Table 3 shows the classification results for different data combinations. The accuracy for multi-sensor/temporal imagery was higher than for single-date data. For HRV imagery, the initial overall accuracy was about 56%. This low accuracy was considered to be partly caused by the phenology that the vegetation was leafless. As to the TM imagery which was acquired in midsummer, the overall accuracy was 66%, higher than that of the HRV imagery. On the other hand, the accuracy for fused data from different sensors (TM4 + TM5 + HRV1) was 78%, and TM4 + TM5 + HRVI resulted in similar accuracy. By recoding with DEM, the accuracy was improved in all cases. The accuracy of multi-sensor classification was raised from 78% to 81%. Table 4 is a cross table showing the classification accuracy of cover types after recoding by using DEM. Most cover types, like conifer forest, spruce larch forest, spruce birch forest etc., showed an accuracy higher than 80%. There were five cover types with accuracy lower than 60%. The thin forest (code 20) showed the lowest accuracy (43.4%), and a large

number of pixels were misclassified as spruce birch forest (code 5) or poplar birch forest (code 17). Nearly 30% of the Korean pine forest was misclassified as mixed forest (code 3). The larch forest was some times difficult to distinguish from mountain birch forest (code 12). About 16% of tundra (code 15) vegetation was misclassified as thin conifer forest (code 10) and other forest types, although the spectral characteristics were significantly different. This was due to the complex topography on the high altitudes of the alpine; shadows are considered to have been the main factor.

4 Cover types and vegetation map The classified image is shown as Fig. 4. The forest under 1,100 m a.s.1, was dominated by mixed forest and hardwood forest. The dominance of Korean pine showed a big variability, revealing the diverse stages of the community. The subalpine forest differed among slopes. For example, the eastern slope was covered by larch forest (code 9), with some conifer forest (code 2) patches scattered in. The northern slope was dominated by relatively pure stands of conifer forest, and there was also a large area of spruce larch forest (code 4). The western and southern slopes were primarily dominated by spruce birch forest (code 5). The southern slope was somewhat complex, a mosaic texture of pure conifer forest (code 2), birch-mixed conifer forest (code 5), and meadow (code 14). The southwest slope was characterized by the significant presence of mountain birch (code 12), which is very resistant to wind. The frequency of strong wind is high in the high elevations, and the formation of birch forest is considered the result of wind disturbance. The subalpine birch zone is generally found in the area of 1,700-2,000 m a.s.1. (Xu and Lin, 1981; Liu et al., 1998), but the eastern slope was almost bare of birch. On the western slope, there was little continuum of

Liu and Takeuchi

145

Fig. 4 Vegetation map of Changbai Mountain by supervised classification of fused images of Landsat TM (July 18, 1997) and SPOT HRV (Oct. 19, 1992). Numbers represent cover types (see Table 3). Source data: TM4 + TM5 + HRVI. birch forest, and it can only be found in valleys, especially in the higher elevations. The current timberline differed among slopes, about 2,000 m a.s.1, on the northern slope, and about 1,800 m a.s.1, on the western slope. This is related to volcanic eruptions. The preeruption timberline was found to be at about 2,200 m (Liu et aL, 1993). The slow recovery of forest on the west side is considered to be the result of wind oppression. Forest outside the reserve was mainly secondary forest, including poplar birch forest, hardwood forest and artificial forest. Little forest (mixed forest) remained. As Table 3

shows, spruce-dominated forest (code 2, code 4, and code 5) had the largest proportion of area (46%). Following it were mixed forest (code 3) and Korean pine forest (code 8) (17%) and larch forest (code 9) (5.4%). Other vegetation types were each less than 5%. Discussion It is generally true that multisensor imagery can provide satisfactory accuracy of classification. But the season can significantly affect the result due to the plant phenology. Such temporal variability of vegetation in crown color requires

146 combination of band images from different seasons. Since the differentiation of canopy colors is the most visible in autumn, images taken in this period, instead of winter, are expected to prove meaningful for improving classification accuracy. Forest composition is not the only factor affecting radiance. As indicated earlier, stand structure like basal area or stand volume can contribute significantly to the spectral behavior. Stand density is an alternative form of biomass or stand volume. In relatively thin forest, understory is an important element contributing to canopy reflectance (Chen and Cihlar, 1996; Mickelson e t al., 1998), and spectral reflectance may show high fluctuation with season. Contrary to this, dense stands are stable in reflectance. This was demonstrated in this study. As to the accuracy of different classification methods, supervised classification is generally more reliable than unsupervised classification, especially for cover types with similar reflectance, while unsupervised classification is acceptable when the contrast in spectral reflectance is large between them (Thomson e t a l . , 1998). In this study, the supervised classification showed a satisfactory result. The study area is complex in topography, ranges from 600 to 2,700 m a.s.l., and is steep towards the peak. This is considered an important factor affecting the classification accuracy. Although the accuracy was empirically improved by using DEM data, further studies on the relationship between topography and spectral radiance are necessary for making a more accurate vegetation map. We sincerely thank the Changbai Mountain Forest Ecosystem Research Station, The Chinese Academy of Sciences for supportingthe field investigation. Thanks go to Prof. T. Takamura for providing image data.

Literature cited Abuelgasim, A.A., Ross, W.D., Gopal, S., and Woodcock, C.E. (1999) Change detection using adaptive fuzzy neural networks environmental damage assessmentafter the Gulf War. Remote Sens. Environ. 70: 208-223. Chen, M. and Cihlar, J. (1996) Retrieving leaf area index of boreal conifer forests using Landsat TM images. Remote Sens. Environ. 55: 153-162. Crawley, M.I. (ed) (1997) Plant ecology. 2nd ed. 717pp, Blackwell Science, Cambrige,UK. Curran, P.J., Dungan, J.L., and Gholz, H.L. (1992) SeasonalLAI in slash pine estimated with Landsat TM. Remote Sens. Environ. 39: 3-13. Dai, X. and Khorram, S. (1998) Hierarchical methodology framework for multisourcedata fusion in vegetationclassification.Int. J. Remote Sens. 19: 3697-3701. Gemmell, F.M. (1995) Effects of forest cover, terrain, and scale on timber volume estimationwith thematic mapper data in a rocky mountain site. Remote Sens. Environ. 51: 291-305. Gong, P., Miller, J., and Spanner, M. (1994) Forest canopy closure from classification and spectral unmixing of scene components: Multisensor evaluation of an open canopy. IEEE Trans. Geosci. Remote Sens. 32: 1067-1073. Haack, B. and Bechdol, M. (1999) Multisensor remote sensing data for land use/cover mapping. Comput., Environ. Urban Syst. 23: 53-69. Jakubauskas, M.E. (1996) Canonical correlation analysis of coniferous

J. For. Res. 6 (3) 2001 forest spectral and biotic relations. Int. J. Remote Sens. 17: 2323-2332. Japan Society of Remote Sensing (1997) Handbook of remote sensing. 28 lpp, National Space Development Agency, Tokyo. (in Japanese) Jeon, B. and Landgrebe, D.A. (1999) Decision fusion approach for multitemporal classification. IEEE Trans. Geosci. Remote Sens. 37: 1227-1233. Joffre, R. and Lacaze, B. (1993) Estimating tree density in oak savannalike 'dehesa' of southern Spain from SPOT data. Int. J. Remote Sens. 14: 685-697. Kilpelainen, P. and Tokola, T. (1999) Gain to be achieved from stand delineation in Landsat TM image-based estimates of stand volume. Forest Ecol. Manage. 124:105-111. Kurosu, T., Uratsuka, S., Maeno, H., and Kozu, T. (1999) Texture statistics for classification of land use with multitemporal JERS- 1 SAR single-lookimagery. IEEE Trans. Geosci. Remote Sens. 37: 227-235. Li, W., Deng, K., and Li, F. (1981) Study on biomass and primary production of main ecosystems in Changbai Mountain. Res. For. Ecosyst. 2: 34-50. (in Chinese) Liu, Q.J., Kondoh,A., and Takeuchi,N. (1998) The forest vegetationand its differentiation under disturbance in a temperate mountain, China. J. For. Res. 3: 111-117. Liu, Q.J., Wang, Zh., and Wang, Sh. (1993) The effects of volcaniceruption on subalpine vegetation of Changbai Mountain. Geol. Sci. Sinica 13: 51-67. (in Chinese) Michelson, D.B., Liljeberg, B.M., and Pilesjo, P. (2000) Comparison of algorithms for classifying Swedish landcover using Landsat TM and ERS-1 SAR Data. Remote Sens. Environ. 71:1-15. Mickelson, J.G., Jr., Civco, D.L., and Silander, J.A., Jr. (1998) Delineating forest canopy species in the northeastern United States using multi-temporal TM imagery. Photogr. Eng. Remote Sens. 64: 891-904. Nemani, R., Pierce, L., and Running, S. (1993) Forest ecosystem processes at the watershed scale: Sensitivity to remotely-sensedLeaf Area Index estimates. Int. J. Remote Sens. 14: 2519-2534. Nilson, T. and Olsson, H. (1995) Effect of thinningcutting on boreal forest reflectance: A comparison of simulations and Landsat TM estimates. Int. J. Remote Sens. 16: 29-63. Rey-Benayas, J.M. and Pope, K.O. (1995) Landscape ecology and diversity patterns in the seasonal tropics from Landsat TM imagery. Ecol. Appl. 5: 386-394. Shao, G., Zhao, G., Zhao, S.H., Shugart, H.H., Wang, S.H., and Schaller, J. (1996) Forest cover types derived from Landsat Thematic Mapper imagery for Changbai Mountain area of China. Can. J. For. Res. 26: 206-216. Spanner, M.A., Pierce, L.L., Peterson, D.L., and Running, S.W. (1990) Remote sensing of temperate coniferous forest leaf area index. The influence of canopy closure, understorey vegetation and background reflectance. Int. J. Remote Sens. 11:95-111. Thomson, A.G., Fuller,R.M., and Eastwood, J.A. (1998) Supervised versus unsupervised methods for classification of coasts and river corridors from airborne remote sensing. Int. J. Remote Sens. 19: 3423-3431. Tong, Q. (1990) Spectra and analysis of typical earth objects of China. 632pp, Science Press, Beijing. (in Chinese) Toutin, T. (1998) SPOT and Landsat stereo fusion for data extraction over mountainous areas. Photogr. Eng. Remote Sens. 64:109-113. Trotter, C.M., Dymond, J.R., and Goulding, C.J. (1997) Estimation of timber volume in a coniferousplantationforest using Landsat TM. Int. J. Remote Sens. 18: 2209-2223. Xu, W. and Lin, Ch. (1981) Preliminary study of relations between vertical distribution of vegetation and thermal climate in Changbai Mountain. Res. For. Ecosyst. 2: 88-95. (in Chinese) Zhukov, B., Oertel, D., Lanzl, F., and Reinhaeckel,G. (1999) Unmixingbased multisensormultiresolutionimage fusion. IEEE Trans. Geosci. Remote Sens. 37: 1212-1226. (Accepted March 27, 2001 )