Spectral Measurements of Solar Irradiance and Radiance in Clear and Cloudy Atmospheres - Short-Wave Solar Radiation in the Earth's Atmosphere - page 90

overshoot is an error where the values of the radiative characteristics at one
or several spectral points are sharply distinct by a magnitude from the neigh-
boring ones. If the relative difference of two neighbor values (following each
other) of the spectral points exceeds the fixed level (e. g. 10%) the consequent
point will be assumed as an overshoot. Note that a detailed logical analysis is
necessary lest a strong absorption band is attributed to an overshoot, either
it is necessary to account for all possible variants of the overshoot positions
in the beginning or end of the spectrum and the nearby overshoots as well.
An overshoot correction consists of the substitution of the point interpolated
over the neighbor sure points to the error point. After the removal of the er-
rors, the procedure is repeated (because the strongest overshoots can mask
the weaker ones) until there is no overshoot at the recurrent iteration. The
breaks at the boundaries of the UV-VD and VD-NIR regions of the spectrum
are caused by the measurements with different photomultipliers at different
spectrum regions (Sect. 3.1). These breaks are likely owing to the deviation
of the dynamical characteristic of the photomultiplier from the linear one.
The removal of the breaks is accomplished by the adding of the corresponding
constant correcting values to the break spectrum region.
The elucidating of the errors using logical analysis is not effective enough.
Usually, the operator easily identifies the errors visually just because he knows
in advance, what the “right” spectrum looks like. Scientifically speaking he
uses the a priori information about the spectrum shape accumulated from
experience. The following stage of the elucidating and correcting of the errors
is based just on that comparison of the spectrumshape with the certain apriori
spectrum . The spectrum under processing and the a priori one are compared
in relative units (they are reduced to the interval from 1 to 2) for excluding
the relationship between the spectrum shape and the signal magnitude. If the
modulus of the comparison result exceeds the standard deviation of the a priori
spectrummultiplied by a certain factor the spectrumwill be identified as an er-
roneous one. The factor is selected during the process of the system tuning. We
have used the factor equal to 4.2 that differs from the traditional magnitude for
the statistical interval equal to three standard deviations. There is an apparent
dependence between the spectrum and atmospheric pressure together with
solar zenith angle, so the distribution of the resulting error is rather different
from Gaussian distribution that explains the deviation of the factor from 3.
Two stages of the system provide the calculation of the standards and of their
standard deviations. At the first stage, the a priori information is absent and
the block of comparing with the standard is turned off. The standard (as an
arithmetic mean over processed spectra) and its standard deviation are calcu-
lated from the results of the first stage (standards are being obtained separately
for upwelling and downwelling irradiances and for different surfaces). At the
second stage, all spectra are processed again with the block of comparing with
the standard turned on. This systemof algorithms, which are accumulating the
a priori information, is a self-educating system as per the theory of the pattern
recognition and selection (Gorelik and Skripkin 1989).
The practice of the data processing demonstrates that the application of
self-educating systems in algorithms of the preliminary analysis of spectropho-