Classification algorithms
Three methods have been tested in estimating
interesting forest characteristics
- a variation of k-nearest neighbour classification method
- regression analysis
- inversion method
K-nearest neighbour
The basic K-nearest neighbour classification algorithm is same
that is used in [National Forest Inventory]
Linear Regression
Linear regression estimation has been applied with
NFI sample plots and segmentation based average smoothing of
backscattering coefficients. The models have been
derived by tree species.
Inversion
The most promising results so far has given the inversion method
developed by Pulliainen.
It utilizes semi-empirical forest backscattering coefficient model,
multi-temporal ERS-1 SAR images and estimated soil and canopy
moisture of images. The semi-empirical model needs field
measurements of forest parameters (growing stock volume) on
a subarea, e.g. on sample plots. The model has been derived by
means of helicopter borne scatterometer, HUTSCAT.
It the first stage, forest bacscattering model is fitted to
sample data with growing stock volume and ERS-1 SAR backscattering
coefficients. The output is the soil and canopy moisture estimates for
each ERS-1 SAR image applied in the analysis.
The model fitting gives the avarage behaviour of backscattering coefficient
as a function of tree stem volume.
In the second stage, the stem volume estimates for all sub-areas
are obtained by fitting the backscattering model into a multi-temporal
set of ERS-1-based backscattering coefficients (average
backscatering values for each sub-area).
The values of soil and vegetation moisture are those
determined in the first stage.
Contact information
- Finnish Forest Research Institute METLA
- Unioninkatu 40A, FIN-00170 Helsinki, Finland
- tel. +358-0-857 051, fax +358-0-625 308
- Internet: Erkki.Tomppo@metla.fi, Petri.Mikkela@metla.fi
[METLA] [National Forest Inventory]
[ERS1-SAR Project]
PM, August 30, 1995.