RADIOMETRIC CORRECTION OF MULTITEMPORAL LANDSAT TM DATA FOR DETECTING RAPID CHANGES
IN MINERAL SOIL FOREST LAND
Jari Varjo, Researcher, National Forest Inventory, Finnish Forest Research Institute, Unioninkatu 40A, 00170 Helsinki, Finland
ISPRS Commission VII, Working Group 1
KEY WORDS: forest change detection, multitemporal difference images, radiometric calibration
ABSTRACT
Possibilities to produce generic training data for forest change detection were studied applying linear radiometric calibration
combined with studentization. The standwise approach with Landsat TM image pairs was applied and the spectral features used,
were the difference of stand mean and standard deviation. The timing of several forest treatments within a three-year image
interpretation interval did not affect to the spectral separability of the treatments. After the regression calibration and
studentization all the treatment classes were significally separable at least on one difference channel. The possibilities of
composing generic training data where training observations need not come be from the image pair analysed was demonstrated.
After regression calibration and studentization the spectral responses were quite similar on all image pairs except on TM channel
four.
1. INTRODUCTION
Detecting changes in forest canopy is important from
strategic to operational management planning of nature
resources. For operational purposes, forest change detection is
necessary for controlling the quality of the. information used
in forest management because of possible updating errors in
continuously updated data bases, and for detecting forest
damages large enough to affect management decisions. For
smaller scale, updating is needed for detecting trends such as
defoliation, for controlling forest activities over large areas
such as the European Union and for detecting natural disasters
in remote areas. Satellite remote sensing provides an
economical and repetitive source of information for change
detection purposes if specific changes can be detected based
on their spectral response.
Before detecting changes different pre-processing methods
for producing multitemporal satellite images have been
proposed. Normally the satellite acquisitions are first
registered together and rectified to some map coordinate
system. After these different radiometric calibrations have
been proposed to make image acquisitions radiometrically
comparable (e.g. Singh 1989). However, only few methods
have been accurate enough for detecting changes such as
silvicultural activities or forest damages in the Boreal Forest
conditions. When absolute and relative calibration methods
have been compared the absolute calibration has not been
accurate enough, and relative methods have been proposed
(Olsson 1994). Selection of the actual radiometric calibration
method depends on the ground truth available. If there is no
ground truth available, methods such as histogram matching
have been proposed, but if for example forest/nonforest
delineation or forest stand delineation is available, good
results have been obtained by linear regression calibration
(Olsson 1993, 1994 Varjo 1996). In the case of a single image
pair the need for radiometric calibration has not been very
clear. Häme (1991) has shown better forest change detection
728
results without calibration but Varjo and Folving (1996) have
found that calibration improves change clustering results
notably. In the case of several image pairs, such as in
operational updating or control of forest management
information, radiometric calibration seems necessary. If
supervised methods are used it cannot be expected that
training data for change detection is collected separately for
every image pair (Varjo 1996). For producing generic training
data good results have been obtained by combining linear
regression calibration and studentization (Olsson 1994).
In this work, relative radiometric calibration methods are
studied for producing multitemporal data for detecting rapid
changes in forest canopy at a scale of 1:20 000. Forest stands
originating from base line field inventory are used as
observation units (Varjo 1996). The combination of relative
radiometric calibration and studentization methods (Olsson
1994) for detecting changes as small as possible from
Landsat a TM difference images is proposed for supervised
change detection methods. The use of relative calibration is
studied for producing generic radiometrically comparable
training data for several image pairs covering the same
geographic location and for image pairs covering different
geographic locations of a similar forest ecosystem type. The
interval within Landsat TM image pairs vary from one to
three years. There are two test sites available for the boreal
forest zone in Finland. The spectral response due to different
forest management treatments is analysed.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996
The
Easte
fores
conif
opera
inter]
these
mine
obser
the c
496 s
Tabk
Nam:
Untre
Uncc
ning
Com
(C.th
Prep:
cut)
Rege
(Reg
Cleai
Soil
prep.
Ther
area.
were
N, 1
are p
Tabl
N=nt
Date
H 21
H 10
H 31
N 08
N 23
*Fro
Robı
radic
pair.
estin
chan
12
selec
band