Epsilon Open Archive

Abstract

It is well known that remotely sensed data and forest variables are correlated. For inventories
like the Swedish National Forest Inventory (NFI) covering large geographic areas, the design
for the sample plot layout is often either systematic sampling or simple random sampling. It is
reasonable to assume that the design could be more cost efficient if remotely sensed data are
used as auxiliary information. A simulated model has been constructed to evaluate the gain in
precision when stratified sampling, based on remotely sensed data, is used.
A grid map was created to correspond with a landscape in the northern part of Sweden. First,
the map area was divided into forest stands, and a vegetation class was assigned randomly ac­
cording to a probability distribution obtained using NFI data from Norrbotten and Viisterbotten
counties, Sweden. Thereafter, for each grid element of the map both wood volume and spectral
values were simulated. The wood volumes were simulated using field data from the NFI and
the spectral values were simulated with a regression function based on the wood volumes that
correspond to a Landsat TM scene with a 30mx30m resolution.
Based on the spectral values the grid elements were classified into different strata. Stratified
sampling was then performed and compared with simple random sampling without replacement.
The comparisons show that the stratified sampling, based on remotely sensed data, produce
much more precise inventory estimates of volume than simple random sampling.