The aim is to investigate how well hold Landsat data can be calibrated and
utilized for monitoring past changes in forest cover status. The material c
onsists of three MSS images and a TM image covering a period of 19 years. T
he classes of interest were 1) no change, 2) deforestation, and 3) reforest
ation. Various relative calibration methods are first compared and their ef
fects on the results of interpreting the "change image" are then studied. A
s it is sometimes impossible to locate unchanged areas for calibration, the
use of both unchanged and changed training areas for calibration purposes
is tested here. The change detection method is fixed to parametric supervis
ed change image classification. The longest time difference between image p
airs was 19 years and R-2 with a time span of this order was around 40%. R-
2 for the models increased about 10% when time span was reduced from 11 yea
rs to 3 years. Smoothing methods can predict band-by-band regression very w
ell, and R-2 values are 10-15% higher than with the robust regression techn
ique used. RMSE was slightly larger when multiple band models were used tha
n with smoothing methods. When general calibration models were applied to u
nchanged data, the differences between the calibration errors were not stat
istically significant. Thus it was satisfactory to use an entire target are
a for image calibration and include changed areas in the training data for
calibration models. Models created from multiple occasions did not perform
well in classification. A fair agreement with estimates and the ground trut
h was possible using four images. The best result was obtained when using t
hree image pairs and multiple band robust regression calibration. Classific
ation with calibration was only slightly better than without calibration, t
he difference not being statistically significant. (C)Elsevier Science Inc.
, 1999.