Olsson, Häkan 
  
independent view of large forest holdings for relatively low cost per hectare. In theory, a new image over the same area 
during the vegetation season every 1-3 years could be used to detect unexpected changes, as well as identify departures 
from expected forest growth predicted by the models. Such information, at the very least, could be used to direct field 
inventory activities in a more efficient way. Ideally, satellite image sequences could be fully integrated into a spatially 
and temporally explicit estimation scheme to improve estimates of the current state of the forest. 
Yet despite considerable research effort in satellite remote sensing for forestry, it has not been widely used 
operationally. We could speculate on the historical reasons for this as (1) a lack of facilities for handling spatially 
referenced digital data at the end-user level, (2) lack of precise position information for sample plots, (3) the cost and 
complexity of ordering data, and (4) a lack of suitable methods for handling images together with sample plots to 
extract useful information in the forest inventory context. At least points 1-3 above have changed recently with 
widespread use of GIS and relational databases for storing inventory information, GPS for positioning plots in the field, 
and efficient web-based tools for ordering low-cost data such as from Landsat-7 ETM+. This project addresses the 4^ 
point, and specifically methods for extracting information about growth from temporal image sequences. 
1.2 Spectral Development of Forests Over Time 
The spectral reflectance of tree crowns and forest canopies have been measured and modeled to try to understand the 
relationship between spectral signatures and forest parameters. Through normal growth and compositional change, the 
spectral signature will change over time in response to changes in the conditions on the ground. This relationship is 
complex and difficult to model accurately, and unfortunately the traditionally most important inventory parameters 
(stem number, height, volume, basal area, etc) may not be the most important factors for determining the temporal 
course of spectral reflectance (Nilson and Peterson, 1994). During the early stages of stand growth, the satellite signal 
is dominated by the reflectance characteristics of the field-layer vegetation and exposed soil or rocks. As trees grow 
and the canopies close over, the background vegetation becomes less important and the species composition and total 
leaf area of the canopy becomes dominant. Height growth comes into effect mostly through the amount of internal 
shadowing, especially in the shortwave infrared wavelengths because of good atmospheric penetration. In a mature 
forest with a fully closed canopy and stable leaf area, increases in basal area, and thus volume, have minimal effect on 
spectral reflectance. The spectral development of a forest stands, as a function of age, are well approximated by a 
decaying exponential function of time within each spectral band. Of course disturbances will introduce discontinuities 
into this otherwise smooth profile. 
It has been proposed that the spectral behavior of forests over time, or its spectral-temporal trajectory, could be used to 
monitor forest development. Háme (1991) refers to the "spectral life cycle" of a forest stand as its spectral trajectory 
over a full rotation. He constructed descriptive models for life cycles that included discontinuities caused by periodic 
stand thinning. Jupp and Walker (1996), outline the potential in this area and suggest using geometric-optical models to 
construct expected profiles, to which observed data can be compared. Nilson and Peterson (1994) propose that a set of 
tables or curves could be produced that represent expected spectral development for a number of site conditions, as a 
direct analogy to growth curves used widely in forestry. Here the geometric-optical canopy reflectance model provides 
a link between the forest inventory data and remotely sensed data. These developments with canopy reflectance 
modeling are certainly encouraging, but to be used effectively, image data must be calibrated to physical units of 
surface reflectance factors to be compared to model outputs. The alternative is to use methods based on statistical 
relationships rather than physical considerations. This is the approach taken in this study- we are interested in picking 
out the general spectral trends over time and comparing these profiles in a relative manner. 
1.3 Objectives 
In this study we explore the possibility to compare forest growth rates on sample plots from a sequence of Landsat TM 
imagery. The emphasis is using a realistic, rather than ideal, image dataset and using a random sample of forest plots 
from the Swedish NFI that reflects the full variability of forest conditions in the area. The main question is whether a 
spectral-temporal profile derived from a normalized image data sequence can explain differences in forest productivity. 
We focus on the parameter site index, since it is an important predictor of forest growth, and it is widely used in growth 
modeling. 
2 STUDY AREA AND TEST DATA 
  
1082 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000.