reflectance product of relative or quantitative value (errors of 20% or more), to a product that can be used 
quantitatively in the global modeling applications where the requirement is for errors to be limited to around 
5-10%. The results demonstrate optimum sampling and averaging strategies for creating hemispherical 
reflectance maps for photosynthetic and climate studies (Kimes et al. 1993). 
The range of view angles that is required in order to obtain accurate inferences of hemispherical 
reflectance using the above techniques was also studied as applicable to instruments such as AS AS, MISR, 
and POLDER that look forward and aft Full and half strings out to 60°, 45°, 30°, and 15° were tested on 
ground data of a wide range of cover types and sun angles. Both a visible and near infrared bands were 
tested using VEG. Using string techniques that calculate a single coefficient from an internal spectral data 
base, the results showed small errors for the full string +60° (less than 4 percent error), for a string +45° 
(less than 7 percent error). Much larger errors occur in those cases where long extrapolations must be calcu 
lated (e.g. strings to only 15° and 30°). Systems that view fore and aft out to 60° would be desirable to 
minimize errors assuming atmospheric corrections could be made at these angles. Sensor systems capable of 
only viewing in the fore or aft directions (half strings) as opposed to both directions (full strings) greatly 
increase the error (Kimes and Deering 1992). 
3.2. Ground Cover and Plant Height 
The Learning System was tested in classifying various classes of percent ground cover and plant height. A 
description of the algorithm is discussed briefly in Section 2.4 and in detail by Kimes et al. (1992). Table 1 
shows a few examples of tests of the system for different classes, view angle combinations, wavelengths, and 
solar zenith angles. Each run produced class solutions (description of best relationships for class discrimina 
tion) which were then used to classify all appropriate samples (all cover types from the data base that fit the 
solar zenith angle and wavelength intervals). The percent of correctly classified samples was then used to 
summarize the degree of classification accuracy achieved by the learning technique. The entire historical data 
base was used in these tests. 
Table 1. Various runs of the system showing the identifying run number, the user defined classes, the view 
angles available (zenith, azimuth) where the azimuth is relative to the sun (0° is forward scatter and 180° is 
backscatter), wavelength in pm, solar zenith angle used for learning and classification, and finally the result 
ing classification score of all the appropriate samples. This score is the proportion of samples correctly 
classified using the class solutions produced by the Learning System. 
Run # 
User Defined 
Classes 
View Angle Available 
(Zenith Relative-Azimuth) 
Wavelength 
(pm) 
Solar Zenith 
Angle 
Score 
#1 
Ground Cover 
(0-30%) and (31-100%) 
String Data 
(60 35) (45 315) (30 315) 
(15 315) (0 0) (15 135) (30 135) 
0.91 
35° 
1.00 
#2 
Ground Cover 
(0-30%) (31-60%) 
and (61-100%) 
String Data 
(60 35) (45 315) (30 315) 
(15 315) (0 0) (15 135) (30 135) 
0.91 
35° 
1.00 
#3 
Plant Height 
(>10 m) and (<10 m) 
String Data 
(60 35) (45 315) (30 315) 
(15 315) (0 0) (15 135) (30 135) 
0.91 
45° 
0.67 
#4 
Ground Cover 
(0-30%) and (31-100%) 
Poorly Dispersed Data 
(0 0) (10 0) (15 0) 
0.68 
40° 
0.65 
Run #1 used directional data similar to what forward and aft looking sensors (e.g., MISR) could collect 
for an unknown target as input to the system. A "string" of off-nadir data occurring in an azimuthal plane of 
1357315° was used as the directions available to learn class descriptions. A near infrared band centered at 
0.91 pm and a solar zenith angle of 35° were used in the learning technique. The directional data used were 
(60 315) (45 315) (30 315) (15 315) (0 0) (15 135) (30 135). Two classes based on percent ground cover 
were defined by the user- (0-30%) and (31-100%). 
The classification score of Run #1 was 1.00 indicating that all samples were classified correctly. The