0916-B1
Zhao Xianwen, Li Chonggui, Si Lin, Tian Yonglin, Yan Kaixian 1
As a new system, its establishment is related to technical methods, technical standards, work circuit (operation process) and benefit assessment. This paper is just an introduction to these aspects.
Forest is the main body of land ecosystem, and is absolutely necessary for human survival. So undoubtedly, Change of forest resources is highly concerned by the society. Traditional forest resources inventory is a hard mission, which needs much time, work and money. Since the end of 1970's, remote sensing technology has been introduced into China. A lot of foresters use this new technology in the activities of forest resource inventory, especially on estimating forest stock volume. This research shows the remote sensing related application fields as following:
In recent years based on the former study and by using modern statistic measurement and non-parameter estimation, new explore and deeper studies were done. These have solved some key problems of applications of remote sensing in forestry resource inventory, such as: registration between sampling plots on the ground and remote sensing image, decreasing field work, increasing accuracy of stock volume on compartment, fast selecting optimal function, and avoiding of all factors that are hard to be determined by remote sensing data (such as age group, crown canopy, etc.). In both theory and practical application has confirmed that remote sensing information play a leading role for estimating forest stock volume. These studies have made a good base for application of remote sensing in forest inventory, and given strong support for establishing a new system of forest inventory, which use remote sensing technology as primary measurement.
At the end of this paper, a frame of new system of forest resource inventory is put forward. The suggestion on key problem in the new system, work circuit (operation process) and benefit assessment are also discussed.
The ridge trace method of modern regression analysis can overcome the disadvantageous influence without optimum solution caused by ill-condition from multi-collinearity existing among the factors influencing stock estimation, which can not be resolved by LS estimation. Ridge trace method can effectively determine optimum factors influencing the estimation of dependent variable [1].
The concrete way of doing is to select effective factors for stock volume estimation by means of ridge trace analysis. The result indicates that the quantitative factors obtained from remote sensing data are the base; the qualitative factors such as canopy density, aspect, elevation, land type also play a certain part. Among the qualitative factors, canopy density's function is the largest. Because canopy density cannot be obtained directly from remote sensing data, it needs to be estimated. In the estimation equation of canopy density, the quantitative factors from remote sensing data are still the base, among qualitative factors; land type (Here land type is only divided into forest land and non-forest land, It is very easy) plays an important part. So, there are conclusions as follows:
The principle of selecting regression independents must be few and perfect. When discarding some variables, the LS estimation and predicted variances of the remained variables will always reduced. But, the estimation and prediction in this case is generally biased. If the influences of the variables discarded are really small, the regression coefficients and predicted mean square errors of the remained variables will surely reduce [2]. The principles for selecting variables are as follows:
There are many kinds of principle for variable selection to LS estimation [3], thinking of the effectiveness and convenient for programming, here the principle of residual mean squares is adopted. According to the principle of regression tree, the main arguments subset can be screened from all available variable subsets. When the number of variables is
, the residual mean squares (RMSq) can be expressed as:
where numerator 
is residual mean squares of stock estimation, meanwhile, 
, meaning the residual mean squares for stock estimation will be gradually reducing with the increasing of the variable number . The minimum value of residual mean squares will be reached when all variables influencing stock estimation are included in regression equation. When selecting variables on the basis of the above expression, with the increasing of variable number , 
will increase. But at beginning, numerator reduced more, leading to 
reducing gradually. When variables increase to a certain degree, the important of them almost been selected, at this time if add other variables, numerator will reduce very slowly, consequently the increase of can not be counterbalanced, at last begin to increase after reaches a minimum, referencing Fig. 1. According to the principle of , the variable subset corresponding to the minimum value of it is the main subset influencing stock estimation. Based on this, the software for selecting optimum equation has been developed. During 4 hours, 2 millions equation can be screened, which ensure the selected equation to be optimum; it can reflect the situation of forest resource in monitoring region.
Fig. 1 The changing curve of RMSq with q
Using remote sensing data to estimate forest resource, especially to estimate forest stock, the main purpose is to improve inventory quality and reduce amount of work and outlay. So, how many sample plots can be reduced from the amount of sample plot needed for traditional inventory under the requirement of accuracy is concerned by forest scientist and enterprises of forestry [4].
The research is carried out with a full TM remote sensing image of Simao district in Yunnan province in 1992 and the sample plots inventory data of the forest district corresponding to it as an example. Adopting the method of systematic sampling, 30% and 50% sample plots are respectively sampled to establish estimation equations. The equations of canopy density and stock are
Where 
is the estimation of canopy density; 
is the estimation of stock; 
is forest land; 
is elevation; 
is shady slope; 
is canopy density.
Assume the surveyed canopy density is 
, the relative error of the prediction deviation of total canopy density and the standard error of prediction deviation can be calculated in the light of below expressions.
The relative error of the prediction deviation of total canopy density is
The standard error of prediction deviation is
Where nn is total number of predicted sample plots.
The prediction accuracy of canopy density and stock can be calculated respectively on the basis of above expressions. The relative error of the prediction deviation of total canopy density and the standard error of prediction deviation are separately 3.405% and 0.131. The errors of stock prediction are respectively 7.788% and 4.731m3. The following conclusions can be obtained through study.
1 With the basis of the remote sensing and GIS information of ground sample plots , the canopy density and total stock of forest can be predicted effectively through establishing the equation based on the unit of pixel.
2 According to the practical example, it is demonstrated that only sampling about 30% to 40% sample plots needed for traditional method, the canopy density and stock can be monitored effectively.
This is a key problem. If registration cannot be solved precisely, the thought of remote sensing data as the base for forest resource estimation will be shaken. During the research by means of the introduction of GPS and GIS technology, this problem has been resolved. The main technology contents are as follows:
The registration error is less than one pixel.
This is another problem concerned in resource survey. The research result indicates that by adopting neural network, according to more that six sample plots the accuracy of estimation region can meet the needs [1].
1 To the region that will not be exploited or managed in near three to five years, this method can guarantee the accuracy of locating to compartments. Because when the area of compartment is 200hm2, there will include six systematic arranged sample plots. (according to current regulations, in the area of a forest bureau, the systematic sample plot is 1km×1km.)
2 To the region that will be opened up soon, in order to estimate the stock volume of sub-compartments, the original method is still need to use. Namely, the stock volume of sub-compartments is estimated by angle gauge assisting eye survey, sample plots entirely control and correction according to weight.
In special segments, such as the accuracy of sub-compartments should meet special needs and the area is relatively large, in this case, some random sample plots should be added to guarantee the total number of sample plots is not less than six so as to ensure the accuracy of stock volume estimation in this segments.
Area is an important composition in forest resource. The inaccuracy of area will affect the estimation accuracy of total stock volume, so the estimation of area is always a key problem concerned by remote sensing scientist. But during recent 20 years, this has not gotten large breakthrough.
1 According to the theory of fractal, introducing the fractal dimension value into comprehensive texture sorting, the efficiency of sorting is increased comparison with the sort based on original gray value of remote sensing data.
2 Using intermediate-resolution satellite image (TM) to calibrate low-resolution satellite image (NOAA), the method of fractal is also introduced, it is a strong support for the classification estimation on large region or the whole world.
However, low cost and low precision, high precision and high cost are not selected for us. So efficiency is defined it as follow:
Efficiency (K)=precision/cost
The efficiency of estimating stock volume using traditional method is Kt=32.09% and remote sensing method is Kr =41.81%.
The result shows that satellites remote sensing data in forest resource inventory is economical and high efficient.
The forest investigation using satellites remote sensing data will save 12,000,000 (RMB) in one year in China
Satellites remote sensing data (TM) was mainly used and ground and other information are assistant.
Class A and Class B are used to investigate middle and long terms developing forest regions.
4.3.1 Land type division
According to north of china, northeast, northwest and the other regions to divide. To the former, pure woods should be classified into different tree species. .
4.3.2 Method of land type division
Combination satellites remote sensing data (TM) and other information
4.3.3 Estimating forest stock volume
The method of multi-variable regressive, which using quantitative and qualitative factors of satellite data as independent variable with a small quantity of volume of ground sample plots (dependent), is used to estimate forest volume.
4.3.4 Stand measurement factors
The canopy density is the only factor need measuring.
4.3.5 Ground sample plots
Angle gauge cruising is used to measure. For re-checking, central stake can be left and positioned by GPS.
4.3.6 Field map
Satellites remote sensing information only are applied for field maps (hard copy)
4.3.7 Making map
All contents was drawn on field maps should be transfer to base map
4.3.8 Calculate area
Supported by GIS, inputting map into computer in terms of different layers and calculating the area of different land type, controlling and adjusting error.
4.3.9 Forest stock volume fluctuation
Forest stock volume fluctuation is weighed by the margin of two yeas volume.
4.3.10 Precision
Area precision: more than 95% are required; volume precision: the regions of standing volume surpass 500,000,000m3 more than 95% are required, and the other regions is 90%. In the future developing regions, precision can be cut down 5%.
One map will be harvested and information of forest resources will be offered per year.
Fig. 2 diagram of forest management inventory by satellite data
[1] Zhao Xianwen, Li Chonggui. quantitative estimation of forest resource based on "3S"-theory, method, application and software. China Science and Technology Press, 2001,5.
[2] Cheu Xiru, Wang Songgu. Modern regression analysis, An Hui education Press, 1987.
[3] Tang Shouzheng. Multi-variant statistics analysis, China Forestry Press, 1989.
[4] Zhao Xianwen. Forest remote sensing quantitative estimation, China Forestry Press, 1997.
1 The Chinese academy of Forestry, Beijing 100091