4.2 Examples. 
For example : model(leaf area index, leaf angle) -> 
multispectral reflection. Inversion of the model : multis- 
pectral data -> green vegetation index -> leaf_area 
index and multispectral data -> sum of photons all 
bands -> intensity -> leaf angle. leaf area index & leaf 
angle -> vegetation subclass. 
Reflected and emitted photons: Landsat TM data with 6 
reflective pass bands and one infrared band. As there 
are two completely different processes at work it would 
be senseless to just combine them into one picture. 
(Temperature is a state variable, the observed emitted 
radiance is a function of the heat balance over a long 
time interval. The use and interpretation of the data 
requires therefore a model for heat flow with many 
parameters which are not observable by RS techni- 
ques). So the reflected radiance is used to measure the 
amount of radiative energy absorbed at the moment of 
observation. This is extrapolated over the previous 
period. Other components of the heat transfer model 
can be derived through surface class membership. If 
the emittance of the objects is known then the 
temperature can be estimated from the IR data. In 
model based image analysis, the predicted temperature 
is compared with the estimated temperature, and an 
optimisation subroutine varies the remaining variable 
parameters until a minimum cost of estimation / 
classification is found. 
Temperature can only be determined if at least the 
(directional) emittance of the object under observation 
is known. 
Feature extraction depends also on class hypothesis: 
for waterbodies and the problem of thermal polution the 
temperature distribution (flow model) is the feature, for 
crop monitoring relative temperatures indicate the 
degree to which plants cope with heat stress and for 
soils the temperature is related to soil moisture because 
water has a high heat capacity. 
In applications of meteo sat and NOAA VHRS data, the 
temperature estimation is relevant for the determination 
of height of clouds, which in turn feeds into a rainfall 
likelihood model. External data are in this case provide 
by rainfall gauges and predicted pressure and windfield 
distribution plus vertical profiles of humidity, pressure 
and temperature. 
900 
Spatial features are defined through geometric (solid; 
modelling, parameters such as object position, size anc 
orientation are estimated from the RS images. Progress 
is made in the geometric modelling of buildings 
[Schutte,1992], trees and homogeneous vegetatior 
canopies . 
The parameters of the microwave reflection model are 
mostly geometric, regular like reflecting plane surfaces 
or irregular like area roughness and scattering veget- 
ation canopies. Woodhouse,1990 has build a simulatior 
model and has demonstrated how a sequence o 
parameter adjustments reduces the difference betweer 
predicted and actual radar image to a noise picture. The 
next important parameter is the dielectic constant ir 
combination with (surface) conductivity. Last, anisotropic 
scatterers / reflectors change the polarisation and phase 
of the incoming e.m. wave. This leads to an estimater o! 
e.g. tree branch directional distribution or directiona 
distribution of fissures and cracks in rocks and ice. Foi 
water applications the parameters describing sea state 
are important. Further research into the model basec 
analysis of SAR images takes into account the spectra 
classification at an earlier data plus a Markov estimatoi 
for the change with time. 
4.3 The method. 
As feature extraction depends so much on class 
definition and the physical model describing the interac- 
tion between radiation and matter, the hypothesis driven 
reasoning of knowledge based systems is selected. 
The following strategy is used : 
- problem analysis leads to the definition of querries in 
terms of classes and subclasses of objects and state 
parameters of the object. - the present state of the 
model representing the previous state of the world is 
used to predict a priory probabilities for class members- 
hip and parameter values. - for each [class , parameter, 
source) combination, the appropriate features are ext- 
racted. - for each object in the scene the class 
likelihoods and the process / state parameters are 
updated. - the GIS used for modelling (specific views of) 
the world stores the class likelihoods and relevant state 
variables / parameters together with a time tag. 
MY ey — 
-— ff^. "1 OY A (0 =i ty 
- — p —— )* en o Du ow "M veu ea 
FN pe VU OU OM ]