Determination of optimum sample number and location
According to the requirements of the project, the accuracy and reliability should be reached 90% and 95% respectively.

The optimum number of samples is able to be calculated according to following formula:

n0=square [SNh*Sh]/V= 471 (sampling units)               (1)
Where: n0, primary sampling size;

Nh, the figures of sampling units in class No. h ;
Sh, the square roots of the real variance of class No. h;
V, the figures of pre-assuming variance.
Because fpc=471/2541=18.5%>5%(fpc : finite population correction), finite population correction is needed. After FPC, the number of samples should be :

n=n0/[1+SNh*Sh*Sh/V]=238.                        (2) As we mentioned before, 9 special units have to be taken into account. Therefore, the optimum total number of sample units will be is 238+9=247.

Based on the optimum allocation method, nh= [(Nh * Sh) /S(Nh*Sh)] * n , the 238 counties can be allocated into each class. The result is shown in Table 1.

Table 1 The allocation of 238 samples in each class

Class No.	1	2	3	4	5	6	sum
Nh (Total units )	115	223	454	1187	419	143	2541
nh (Sampling units)	97	23	17	16	17	68	238

Experiments and Discussions

Experimental results
The experiment was conducted using Landsat TM images. The sampling units of 202 counties under remote sensing investigation are chosen from Sichuan, Jiangsu, Heilongjiang, Guangdong, Gansu, Jiangxi, Shanxi Provinces and Xinjiang Autonomous Region. The total land area and total arable land area in selected region cover 370,000 square kilometers and 190 million mu respectively. In addition, 78 county level units (mainly distributed in class 1 and class 6) were selected for test. Therefore, total 280 units were used in our experiment. The estimation formula are shown below :

y=(1/N)*(SNh*yh)                                         (3)
Where: y, estimated value of population average;
N, total unites of population;
yh, average value of stratum No. h.

V(y)=[1/(N*N)]*[SNh*(Nh-nh)*Sh*Sh/nh]               (4)
Where: V(y), the variance value of population average;
Sh, variance value of stratum h;
nh, real sampling unites in stratum h.

Table 2 shows the estimated results calculated from eq. (3) and (4).

Table 2. The estimated results

class	Nh	Real nh	Sh	Average for sample y	Y population	Standard Error of Y
1	115	97	17941.97	-2812.74	-7358127.80	708410.92
2	223	26	4578.77
3	454	25	3534.51
4	1187	39	2446.92
5	419	19	3474.61
6	143	74	18679.89
sum	2541	280

Discussions
The errors occurred in this project came from two sources. One was the sampling error, which is S1=70.8 (sampling standard error); the other was the error in visual interpretation of satellite imagery (TM) . The visual interpretation work had been carried out for 3 years. The error in visual interpretation was about 5%. Hence the misinterpreted error is S2 = 7415000* 5% ˜371000 mu. Therefore, the general error (S) can be calculated as: S = square root (37.1*37.1+70.8*70.8) = 800,000 mu According to the relation between accuracy and reliability(Y±S*t), one can easy calculate accuracy according to the reliability requirement.

• Assuming the reliability requirement is 90% (t = 1.64),
  then, the limiting error is t*S = 1.64*800,000= 1.31 million mu;
  therefore, the accuracy of estimation is (741.4 - 131) / 741.5 = 82.5%.
• Assuming the reliable interval is 85% (t = 1.46),
  then, the limiting error is t*s = 1.46*80= 1.17 million mu;
  So, the accuracy of this estimation is (741.4 - 117) / 741.5 = 84.2%.

Similarly, when the reliability requirement is 68%, the estimation accuracy will be about 89% .

It should be mentioned that the comparison of estimated results from remotely sensed imagery with the statistical data may not really obtain the accurate results. The major reason was that the original statistical data could not reflect the real situation. Taking Dongguan City of Guangdong Province as an example, the city lost several dozens thousands mu of arable land annually in recent years, but the decrease information was not reflected in its statistical datum. Some of the counties in Gansu Province reported their statistical datum which were of wide difference compared with the datum from remote sensing investigation. Moreover, the absolute value was same but the symbol was just opposite. Such problems can be only solved gradually with the deepening understanding of the population in the work and continuing improvement of class program.

Conclusions
This paper described a practical model for estimating the nationwide arable land change with some sampling data. The results of experiment shows that the professional technology method with integration of remote sensing technology and area sampling technology, is one of the most effective and economic methods for implementing the monitoring of the nationwide arable land resources changes.

The method and technology proposed in this model may be easily applied for monitoring or estimating other utilization status of the land, such as the planting area of main crops. In these cases, the sampling method may have to be modified.

In the future, we are going to test the model in other countries or regions. The possibility of applying proposed model in as macro investigation and management may also be investigated and discussed.

References

• William G. Cochran: Sampling Techniques, Third Edition, John Wiley & Sons,USA.1977
• Wigton W., Area frame sampling, documents for training course, USDA. 1979.
• Gallego,F.J., Delince,J., Area estimation by segment sampling. In Euro-Courses: Remote Sensing applied to Agricultural Statistics. JRC-ISPRA,ITALY,1993
• Meyor-Roux J., The ten years research and development plan for application of remote Sensing in Agricultural Statistics. Joint Research Center, EEC. publication No: JRC SP 1.87.pp39,24.

Page 2 of 2
| Previous |