Detection of forest change using multi-spectral scanner data

Detection of forest change using multi-spectral scanner data

Xu Dingcheng, You Xianxiang, Han Xichun
Beijing Forestry University, Beijing, China

Abstract
In the experiment, MSS satellite data tapes of two periods (May, 1976-October, 1985) were using and the following methods which were studied include image D-value, D-value of ratio vegetation index, Normalized D-value vegetation index, Multi-temporal KL analysis and Monitoring of classified comparison. These methods have been used in the forest dynamic monitoring.

Research on region information acquisition and methods
The experimental region is located in Pingquan County, Hebei Province (at 41°2'19" - 41°14'10" N and 118°31'43" - 118°47'36" E). The area of experiment is about 600 square kilometers, Two MSS data tapes of two periods were used in the experiment (see the following table.)

satellite number	2	5
date	Many 16,1976	October 7, 1985
index number	13/31	13/31
sun angle	56°	48°
position angle the sun	123°	135°

Besides, the following aerial photos were also used in the experiment:
Black-white aerial photos 1:50,000, taken in 1979; infrared color aerial photos (1:,30,000, 1:70,000 and 1:130, 000)' and a number of forest distribution maps in different periods and topographic maps, etc.

In the experiment, binary cubic multinomial and bilinear interpolation was adopted to take samples and the temporal image MSS of 1985 was corrected. During the correction 19 control points were evenly used. After correction, the image of 1985 was taken as the standard, while the image of 1976 was registered with that of 1985. The mean square root errors in X and Y after recombination are 0.331 pixel, and 0.274 pixel, respectively.

Dynamic monitoring method of forest area
Dynamic monitoring of forest area was conducted by mans of the difference between the two-temporal images. The two-temporal changes can be divided into two kinds The first kind includes atmospheric condition, soil moisture, the difference in satellite detection process, etc. which influence most or all pixels. The influence can be eliminated or decreased through operations or rotary data space. the second involves only part of pixels, such as forest felling, afforestation regeneration seasonal difference etc.

To collect the information of forest dynamic changes, the following methods were used in the experiment:

Method of Image D-value
The growth and decline of forest will induce the changes of images of red-light waveband as well as near infrared waveband. The method of image D-value is used to substract lumin-ance value of the first temporal image from the matched original image of the second temporal. Theoretially, positive and negative values indicate the changed pixels and zero indicates the pixels without changes.

Since the luminance value of image is between 0 and 255, a constant is usually added in the method of D-value to eliminate the negative value.

The formula is as follows:

DX ( ijk ) = X( 2 )( ijk ) - X ( 1 )( ijk ) + C

where
DX indicates changed image
X (1), X (2) indicate the first and second temporal images
C indicates constant; i indicates line;j indicates row; k indicates waveband.
The histogram of the D-value image MSS7 and MSS5 produced by the method of image D-value isdistributed like a bell in shape (see Figure 1).

Fig. 1. Diagrammatic sketvh of D-value image threshold detection

Although the two wavebands are highly sensitive to vegetation changes, the combination of the changes of pixel luminance caused by aspection difference, atmosphere, location of satellites and difference of soil moisture, and the changes of vegetation coverage makes the two kinds of changes unable to be separated. As a result, it is difficult to collect all of information, While the influences of different factors on different wavebands are not the same, the colour combination of the D-value of 3 wavebands synthesizes the dynamic information of different wavebands. Therefore, the information of vegetation changes can be conspicuous.

If D-value images MSS7, MSS5 and MSS4 are respectively matched with the colour synthetic images of red, green and blue, the increased area of vegetation will be red; the damaged area of vegetation is dark drown; the greater part of cyan represents type of land without any change. Simple D-value synthetic image produces excellent visual interpretive effects.

D-value Method of Ratio Vegetation Index
Since the ratio of IR/RED is closely related to plant biomass, this vegetation index of comparative two-temporal can effectively monitor the forest vegetation changes. At th3 same time, the ratio can also eliminate the influence of the atmospheric condiction, soil moisture and sun angle on the image, and reduce the difference of changes not caused by the type of land, and , consequently, the changes of land type become conspicuous.

Calculation formula:

DRij = (MSS7(2)ij / MSS5(2)ij ) C2 - ( MSS7(1)ij / MSS5(1)ij ) C1 + C

DR(ij) ratio D-value image; MSS(1) first temporal; MSS(2) second temporal;
C, C1, C2 constant; i line; j row.
On the ratio image, pixels of high luminance mean the sharp increased area of ratio vegetation index, while low luminance the sharp decreased area of ratio vegetation index; a greater part of pixel assumes intermediate grey, signifying an area of slight change of vegetation index. In view of the distribution of histogram pixels of the violently changed vegetation index distribute at the two tails of the histogram At the left tail is the distribution of pixels of sharply decreased vegetation index, while at the right tail is the distribution of pixels of sharply increased vegetation index. Vegetation index changes slightly in a greater part of area and distributes in the intermediate position of the histogram. The luminance distribution of the entire vegetation index D-value image is continous. The two-temporal vegetation index changes are as follows:

Changes are caused by the difference of seasons. However, if only it is in the growing season, no great changes will take place with the changes of seasons.
Changes are caused by those of vegetation viability. Nevertheless, owing to the long growing cycle and slow senility of forests, and in the condition of short plastochron, changes of vegetation index will be slight.
Changes of vegetation index are caused by the succession and reform of forest vegetation. Changes caused by this factor, in fact, are different from those of vegetation index caused by vegetation type. However, the former changes, in general, are not great either.
Changes in the growth and decline of forest vegetation are, mainly, caused by felling, fire and afforestation. The sharp changes of vegetation index caused by this factor are even greater that those caused by other factors.

Based on the fact that the growth and decline of forest vegetation exerts a remarkable influence on the changes of vegetation index, the changes of forests can be monitored according to the changes of vegetation index. Dynamic monitoring of forest area, in general, can be divided into three kinds, namely, newly-increased forest land, untouched forest land and damaged forest land. For this reason, vegetation index D-value image must be divided according to a certain threshold to detect the position and size of the changed area. In order to determine the division accuracy of different threshold, 225 sample points are set up at the entire experimental window. According to the standard of dynamic results interpreted by two-temporal aerial pictures, accuracy of different thresholds is checked and taken. Consequently, optimum threshold is determined. During the interpretation of aerial photos, in consideration of spatial resolution of image MSS and classification method of forest investigation of our country, changes of forest land are determined as follows:

newly-increased forest land:

forestless land > forest land, shrub forest land
thin stockedl land > forest land, shrub foreest land
damaged forest land:

forest land > forestless land, thin stocked land; shrub forest land > forestless land
In order to estimate the accuracy of dynamic monitoring, during calculating, monitoring accuracy, average accuracy and total accuracy should first be calculated respectively, and the average value of both is taken as the criterion to compare different thresholds (see Table1) .

Table 1 shows that monitoring accuracy detected with 1.25 thime s of standard difference is the highest; average accuracy amounts to 78.5% and comprehensive accuracy, 76.85%.

Table 1 Image threshold detection table of ratio vegetation D-value

standard difference time K	average value x=89. 177 standard difference STD=15.288
	accuracy of correct classification unit %
	average accuracy	total accuracy	comprehensive accuracy
0.75	68.53	69.78	69.20
1.00	70.32	72.40	71.36
1.25	78.50	75.20	76.85
1.50	73.35	74.35	73.79

Calculation methods of different accuracy in Table 1 are as follows (the same with others)

average accuracy = (( monitoring accuracy of correct changes + accuracy of correct non - changes) / 2) 100%

total accuracy = (correct total / total number of samples ) x 100%

comprehensive accuracy = (( average accuracy + total accuracy ) / 2 ) X 100%

Normalized D-value Vegetation Index Method
Like ratio vegetation index, normalized D-value vegetation index reflects vegetation. In the area of sparse vegetation and great interference of soil background , the application of normalized D value vegetation index is superior to ratio vegetation index.

calculation formlua:

DNDij = ND(2)ij - ND(1)ij + C

NDij(k) = ( MSS7(k)ij - MSS5(k)ij ) / ( MSS7(k)ij + MSS5(k)ij ) ) * CK
where:
DND stands for D-value image; normalized vegetation index; MSS (7) the 7th waveband; MSS 5 the 5th waveband; k temporal; i line, j row and constant. The result of this method is shown as Table 2, The average acuracy is 75.6% total accuracy, 73.7% and comprehensive accuracy, 74.7%.

Method of Multi-temporal KL Analysis
Two-temporal wavebands of MSS are taken as eight-channel data. The extended data, by KL analysis, separate the vegetation information changes taken as a type of noise from the high order KL producing vegetation dynamic information, in the process of multi temporal structure protation Landsat image is transformed by means of KL, The first KL is luminace, the second green, greater part of the third and fourth, noise.

Two conditions are needed for carrying out dynamic monitoring by means of multi-temporal KL, e.i., two-temporal images possess dimensionality of two-dimension, namely, luminance and green, Land coverage and changing extent of vegetation exceed a certain limit. With the two conditions, and through exact registration, these multi-temporal multi-dimension data in numeral space rotation produce spectral reflection changes caused by different dynamic changes which are separated each as one-dimension component. Of the new KLs, the four KLs, stable luminance and stable green, changing luminance and changing green are of significance.

Table 2 shows characteristic root and characteristic vector of the KL transformation of MSS two-temporal experimental window images. Each of the first and second KL of two-temporal includes more than 98% of the information of the four original waveband images Consequently, the basic dimensionality of the two-temproal's original data is two-dimensional; each of the first KL of two-temporal is positive value, being luminance of image, accounting for about 90% of total information; while the second KL in waveband of visible-light is negative value, in infrared waveband, positive value. This KL reflects the characteristics of vegetation, called "green".

Dynamic changes can be analyzed from the transformation of multi-temporal KL. From Table 3 it can be seen that the characteristic vectors of the first principal component are positive value, reflecting the stable luminance of multi-temporal image, and including 71.8% of all variable information. Multi-temporal first principal component image reveals that where the luminance is high on the image of different wavebands of two-temporal, the luminance on the first principal component image is also high, such as waste land and farmland, etc. In the first four passages of the second principal component, the characteristic vectors of the first temporal at the four wavebands are all negative values. While in the latter four passages, the characteristic vectors of the second temporal in the four wavebands are all positive values. This principal component reflects the changes of two-temporal luminance. The second principal component image shows that the luminance changes of bench and arable land near gullies and valleys, along rivers and streams, are greater. It can be seen from image of the 16 May , 1976 that because it was spring, soil humidity of these regions was greater; and, hence the luminance in different wavebands was low. However, on the image of the 7th October, 1985, because it was autumn, the bare soil was dry, and the luminance of different wavebands was higher, The general trend of characteristic vectors of the third principal component is negative value at the visible wavebands of different temporals, but positive value at infrared waveband. Therefore, this waveband is stable green. In the third principal component image, where there is vegetation cover, the pixel luminance is higher. The characteristic vectors of the fourth principal component are positive value at the first temporal visible-light waveband and the second temporal infrared waveband; first temporal infrared waveband and the second temporal visible-light waveband are negative value. This principal component gives prominence to changes of the two temporal reflection spectra, caused by vegetation. In this principal component image, the luminance of newly increased vegetation region is high, that of damaged vegetation region, low. Compared with the first four KLs, the KLs of higher orders contain very little information and it is difficult to determine its significance. For dynamic monitoring of forest the fourth principal component is the information we want to extract, which reflects the situation of vegetation changes. The monitoring accuracy of dynamic image separated by 1.50 times of standard difference, using standard difference threshold, is the highest. The average accuracy amount s to 78.4%, total accuracy, 80.64%, and comprehensive accuracy, 79.52%.

Table 2 Characteristics root and characteristic vector of MSS
experimental window image of 1976, 1985

	Statistic	Principle Component
	Statistic	1	2	3	4
1976	Characteristic root	109.635	9.729	1.369	0.958
	Contribution Rate	90.1%	8.0%	1.1%	0.8%
	Accumulated contribution rate	90.1%	98.1%	99.2%	100%
1985	Characteristic root	148.783	13.926	1.836	1.438
	Contribution Rate	89.6%	8.4%	1.1%	0.9%
	Accumulated contribution rate	89.6%	98.0%	99.1%	100%

Table 3 Multi-temporal KL characteristic
root and characteristic vector

Channel	Principle Component
Channel	1	2	3	4	5	6	7	8
1	0.413	-0.094	-0.546	0.239	0.481	0.475	-0.093	0.006
2	0.436	-0.109	-0.421	0.163	-0.506	-0.316	0.451	-0.000
3	0.420	-0.369	0.082	-0.147	-0.145	-0.271	-0.745	0.089
4	0.354	-0.531	0.507	-0.237	0.180	0.194	0.449	-0.087
5	0.317	0.459	-0.010	-0.392	0.519	-0.502	0.113	0.035
6	0.343	0.478	0.086	-0.408	-0.431	0.525	-0.045	0.131
7	0.209	0.286	0.300	0.415	-0.038	-0.016	-0.116	-0.754
8	0.218	0.197	0.405	0.589	0.056	-0.040	0.033	0.630
Characteristic root	1828.54	455.141	146.449	53.942	25.619	18.652	13.589	6.197
Contribution rate	71.8%	17.9%	5.8%	2.1%	1.0%	0.7%	0.5%	0.2%
Accumulated contribution rate	71.8%	89.7%	95.5%	97.6%	98.6%	99.3%	99.8%	100%

Monitoring Method of Classified of Classified Comparison
On the basis two-temporal classification image, contingency table is obtained by comparing pixels one by one. This contingency table can explain the changing situation of different type of land. Method of classified comparison is conducted on the basis of two-temporal classification image. Its accuracy is the product of two-temporal classification accuracy. This indicates that this method makes high demands on the classification accuracy of different temporal images, before the results of dynamic monitoring can reach acceptable accuracy. At present, however, owing to the restriction of classified accuracy, accuracy, accuracy of classified comparison method is not satisfactory.

During the classification of images of experimental area, it is difficult to select training area, because different types os land are comparatively broken.

Two-temporal experimental window is classified using unsupervised classification. The combined images MSS 7, 5,4 of different periods are classified. The temporal images of 1976 are sorted into 18 classes and those of 1985, 15 classes. With reference to te collected colour infrared aerial pictures, black-white aerial pictures and two-temporal forest distribution maps, the corresponding relations of classification results and practical classification of surface features are determined . Because the key point of this experiment is forest vegetation, and, at the same time , considering the influence of two-temporal image on the difference of land types, the land types should be combined to the utmost extent. Different-temporal land types are combined into the following 5 classes:

1. water area; 2. Conferous forests; 3. Deciduous forest; 4. Shrub forests; 5. Farmland and wik grassland.

The method of sampling is used in examining the accuracy of classification results. At the entire window are set up 225 specimen points; the results of comparative classification are examined by means of the interpretation results on aerial pictures. The results of examination shows that the temporal classification accuracy of 1976 is 75.4% the temporal classification accuracy of 1985, 80% According to this accuracy, the dynamic matrix obtained on the basis of comparison between pixel and pixel and pixel has only 80.0 x 75.4 = 60.32%. The error amounts to as high as 39.68%. Thus, it can be seen that the dynamic change matrix produced in this way is unreliable. However, the dynamic changes among and within the land types can still be understood there from to some extent.

Discussion
From the comparison of different methods it can be seen that the highest comprehensive accuracy of multi-temporal KL analysis method reaches 79.52%; that of ratio vegetation index, , 76.85% and that of normalized vegetation index, 74.7%. The reason why the accuracy is not very high is that the spatial resolution of MSS image is not high. It is not easy to monitor the small area of forest land. Spectral reflection changes caused by the factor of seasons accounts for the error. If the information sources such as TM, SPOT images area adopted and the same temporal is used as much as possible in conducting dynamic monitoring , higher investigation accuracy can be obtained.

Both ratio vegetation index and normalized vegetation index, going through the operation process of the same temporal image ratio, eliminate the changes of gain in terms of time existing on the image of single waveband. However, ratio will strengthen the random noise on the image.

By using the method of multi-temporal KL analysis, both the data relationship and the influence of atmosphere, the sun angle, soil moisture and noise on the image are eliminated. The process, in which the method of multi-temporal KL analysis is used in monitoring the dynamic changes of vegetation is, in fact, that in the process of multi-temporal image numeral rotation, the changes [the increase of near infrared wavebands (MSS4, MSS5) reflection] caused by vegetation to soil reflection are speparated as high-order KL. When this methods is used, the changing area of vegetation should accounts for only a small proportion of the total area, owing to a small part shared by the high order principal components. In addition, it is required that what the first two of different temporal images KLs reflect the "green". That under what proportion this method could be used remainds to be further studied.

Image D-value method directly compares two-temporal corresponding single wavebands. Owing to the different influences of atmosphere and the sun angle on two-temporal image, if dynamic information is extracted from the D-value image of single waveband, the two-temporal image must be through radiation-correction. Normalization of the sun angle or un-changed land type is taken as sampling point; the registration of two images' luminance is con-ducted by using the method of linear regression. Landsat-MSS 7th and 5th wavebands can be used in monitoring vegetation dynamics.

By using classification comparison method, based on the comparison between pixel and pixel the situation of changes among and within different types of land can be understood. But it is required that different temporal images be first exactly classified, for the broken area of land type is liable to cause the phenomenon of mis-classification, thus restricting the accuracy of classification. Classification comparison method, used in monitoring dynamics in the complete area of land type, can be expected to reach higher accuracy.

Reference

Liu Xinghan and Zhao Huachang, The Application of Remote Sensing Technology in Land Utilization dynamics, Geography Press, 5, 1988.
Xu Guanhua and Xu Jiyan, Research of Renewable Resources, Science Press, 1988
Yankai and Lujian, Treatment of Remote Sensing Image, Surveying and mapping Press, 1988.
John, E.E., Doouglas S; Johan, R.J; Monitoring land Use and Cover Changes, Remote Sensing of Resources Management, 1985.
James, V., Detection of forest Multi-spectral Scanner Data, international Journal of Remote Sensing, 1988.