Page 267 - Special Topic Session (STS) - Volume 4

STS583 Yakob M. S.
Area Sampling Frame consisting of dividing the territory in PSUs with physical
boundaries through visual interpretation of satellite imagery, stratification of
PSUs into agricultural classes intensity by photointerpretation and subdividing
the sampled PSUs in to segments.
Another example of comparison can be drawn from the area frame
employed by China’s National Bureau of Statistics in ten provinces (covering 1
652 083 km2). It used a stratified two-stage sampling design with PPS selection
in the first stage (the size of each PSU ranging between 1 km2 and 5 km2) and
random selection in the second stage (each SSU having a size between 2 ha
and 5 ha, with a total sampling fraction in the order of 0.2 percent). For Anhui
province (139 400 km2), a sample size of 6 000 segments leads to a CV of 1.3
percent for wheat (2 200 000 ha), of 0.9 percent for middle rice (1 900 000 ha)
and of 3 percent for corn (1 000 000 ha). The good level of precision obtained
results from the stratification and from the PPS sampling, based on the
classification of GF1 and ZY3 Chinese satellite imagery (with a resolution of 2
m). The associated costs amount to US$75 000 per province, and therefore
approximately US$0.5/km2.

3.2 Improved estimators
At estimation level, merging data from the ground survey and from
satellites is usually achieved through regression or calibration estimators
(Global Strategy, 2015a). Gallego et al. (2014) present detailed results for a
region (78 500 km2) in northern Ukraine, containing 2.45 million ha of
cropland. Ninety, 4 km x 4 km, square segments were field-surveyed in 2010
(with a sampling fraction of 1.8 percent and a field size up to 250 ha). Later,
the entire region was covered with MODIS, Landsat5, AWiFS, LISSIII and
Rapideye imagery. Image classification was trained on data collected along
the road, independently of the area frame segments. For the major crops
(wheat, barley, maize and soybean), the respective mean efficiencies
amounted respectively to 1.59, 1.54, 1.48, 1.50 and 1.50; therefore, the sensors’
performance was approximately equal.
Comparing the cost of the field survey and the cost of imagery (today, the
cost of image classification is so low that it can be set aside), the situation
changes drastically, because only the two free-of-charge sensors (MODIS and
Landsat TM) remain cost-effective, as the purchase price of the other three
sensors make them inefficient (AWiFS, LISIIIand Rapideye).
A general context to the study include: First, field size in Ukraine tends to
be large, allowing for coarse resolution sensors to compare with finer-
resolution ones in terms of classification accuracy. This would not hold in most
African or Asian countries, where fields tend to be rather small. Second, the
study relied only on the Maximum Likelihood Classification (MLC) method -
today, the USDA relies on the decision-tree classification method (See5).
Finally, the availability of freely accessible imagery is increasing. MODIS (250 -

256 | I S I W S C 2 0 1 9