Title: New Modeling and Data Analysis Methods for Satellite Based Forest Inventory
1New Modeling and Data Analysis Methods for
Satellite Based Forest Inventory
2Consortium members
- Rolf Nevanlinna Institute, University of Helsinki
- National Forest Inventory of Finland, Finnish
Forest Research Institute - Laboratory of Space Technology, Helsinki
University of Technology
3Participating researchers
RNI Lasse Holmström1) Petri
Koistinen Jukka Sarvas Pasi Ylä-Oijala Lisa Zurk
NFI Erkki Tomppo2) Juha Heikkinen Matti
Katila Kai Mäkisara Ilkka Taskinen
LST Martti Hallikainen3) Marcus
Engdahl Laila Hokkanen Jaan Praks Jouni
Pulliainen Teemu Tares
1) Consortium coordinator, coordinator at RNI 2)
Coordinator at NFI 3) Coordinator at LST
4The general goals
5The general goals
- Develop new statistical data-analysis tools for
optical and microwave sensor based forest
inventory.
6The general goals
- Develop new statistical data-analysis tools for
optical and microwave sensor based forest
inventory. - Develop a new microwave scattering model for
forest imaging using mathematical modeling and
efficient computational implementation.
7The general goals
- Develop new statistical data-analysis tools for
optical and microwave sensor based forest
inventory. - Develop a new microwave scattering model for
forest imaging using mathematical modeling and
efficient computational implementation. - Apply the scattering model and improved
data-analysis tools to fully utilize the new type
of data being made available by the emerging
sensor technologies.
8The EMFORSIM Simulator
- Tree modeled as a collection of cylinders on a
reflecting half-plane. - Fully polarimetric calculations, various
scattering mechanisms. - A fast implementation in MATLAB that allows
models or realistic complexity.
9A realistic tree model LIGNUM
- A tree growth model developed at the Finnish
Forest Research Institute (METLA) - Thousands of cylinders (14 824 here)
10Scattering Calculation
Incident wave at Qi, fi
Direct scattered wave (infinite cylinder approx.)
Qi
Multiply scattered wave
Scattering types direct ground-cylinder cylind
er-ground ground-cylinder-ground cylinder-cylind
er
Qs
Ground scattered wave (from image theory)
Qs
- Rigorous treatment of first order (dominant)
scattering - Phase relationships (i.e., path lengths) are
included - Fully polarimetric scattering calculation
- Future extensions can include adding ground
roughness, snow layers, leaves, etc.
11Results
12Results
- Structural differences detected in simple trees.
13- each tree has the same biomass
- C-band scattering (5.3 GHz)
- average of 30 forests
- each forest consists of 10 random realization of
a single tree type
14Results
- Structural differences detected in simple trees.
15Results
- Structural differences detected in simple trees.
- Random trees and random transformations of a
single tree give similar results
perhaps no need for complex forest models.
16Results
- Structural differences detected in simple trees.
- Random trees and random transformations of a
single tree give similar results
perhaps no need for complex forest models. - Statistical scattering properties of forests of
complex trees similar to real data (LST).
17Results
- Structural differences detected in simple trees.
- Random trees and random transformations of a
single tree give similar results
perhaps no need for complex forest models. - Statistical scattering properties of forests of
complex trees similar to real data (LST). - Ground model (mirror surface) needs to be
improved (LST).
18Forest parameter estimation developing a new
statistical approach
- A typical forest parameter average trunk volume
in some area. - Available data ground truth (field plots) in
various locations, satellite measurements from
the area of interest. - Current approach given a satellite measurement,
find field plots with similar characteristics and
compute a weighted average of their forest
parameter values (k-NN regression).
19A new approach
- Use a more general regression method local
linear ridge regression (LLRR). - Determine the level of smoothing in local
regression by area-based cross-validation - takes into account spatial correlation
- directed towards the actual goal estimation of
regional averages
20The idea of area-based cross-validation
- The area for which a forest parameter needs to be
estimated has size A
A
- Training data (field plots) are divided into
subareas of about size A. - Cross-validation is used to select the required
smoothing parameters.
21The idea of area-based cross-validation
- The area for which a forest parameter needs to be
estimated has size A
A
- Training data (field plots) are divided into
subareas of about size A. - Cross-validation is used to select the required
smoothing parameters.
22The idea of area-based cross-validation
- The area for which a forest parameter needs to be
estimated has size A
A
- Training data (field plots) are divided into
subareas of about size A. - Cross-validation is used to select the required
smoothing parameters.
23Finding the best level of smoothing using
area-based cross-validation
24Comparison of the new approach and the current
approach estimation of total trunk volume