204
Chapter 11: Stochastic Modeling of Precipitation
11.4.1 Weather Classification Schemes
Weather classification schemes have been the mechanism used by several
authors to summarize large-area meteorological information. The general
concept of weather classification schemes (see, for example, Kalkstein et al.,
1987) is to summarize measures of similar atmospheric conditions. Most of
the externally forced stochastic precipitation models can be classified according to whether the weather classification scheme is subjective or objective,
and whether it is unconditional or conditional on the local conditions (e.g.,
precipitation occurrence).
Subjective classification procedures include the scheme of Baur et al.
(1944), from which a daily sequence of weather classes dating from 1881
to present has been constructed by the German Federal Weather Service
(Bardossy and Caspary, 1990) and the scheme of Lamb (1972), which has
formed the basis for construction of a daily sequence of weather classes for
the British Isles dating to 1861. The subjective schemes are primarily based
on large scale features in the surface pressure distribution, such as the location
of pressure centers, the position and paths of frontal zones, and the existence
of cyclonic and anticyclonic circulation types (Bardossy and Caspary, 1990).
Objective classification procedures utilize statistical methods, such as principal components, cluster analysis, and other multivariate methods to develop
rules for classification of multivariate spatial data. For instance, McCabe
(1990) utilized a combination of principal components analysis and cluster
analysis to form classifications of daily weather at Philadelphia. The statistical model was compared to a subjective, conceptual model, which was found
to give similar results. Briffa (Chapter 7 in this book) describes an objective approximation to the Lamb scherne, and uses it for regional validation
of selected GCM control runs. Wilson et al. (1992) explored classification
methods based on K-means cluster analysis, fuzzy cluster analysis, and principal components for daily classification of weather over a large area of the
Pacific Northwest. All of the above methods are unconditional on local conditions, that is, no attempt is made to classify the days in such a way that
local precipitation, for instance, is well-described by the weather classes.
Hughes et al. (1993) used an alternative approach that selected the weather
classes so as to maximize the discrimination of local precipitation, in terms
of joint precipitation occurrences (presence/absence of precipitation at four
widely separated stations throughout a region of dimensions ab out 1000 km).
The procedure used was CART (Breiman et al. , 1984), or Qlassification and
Regression Trees. The large area information was principal components of
sea level pressure. Figure 11.4 shows the discrimination ofthe daily precipitation distribution at one of the stations modeled, Forks, according to weather
class. As expected, because the classification scheme explicitly attempts to
"separate" the precipitation (albeit occurrence/absence rather than amount)
by the selected classes, the resulting precipitation distributions are more distinguishable than those obtained by Wilson et al. Hughes et al. also simu-
Chapter 11: Stochastic Modeling of Precipitation
11.4.1 Weather Classification Schemes
Weather classification schemes have been the mechanism used by several
authors to summarize large-area meteorological information. The general
concept of weather classification schemes (see, for example, Kalkstein et al.,
1987) is to summarize measures of similar atmospheric conditions. Most of
the externally forced stochastic precipitation models can be classified according to whether the weather classification scheme is subjective or objective,
and whether it is unconditional or conditional on the local conditions (e.g.,
precipitation occurrence).
Subjective classification procedures include the scheme of Baur et al.
(1944), from which a daily sequence of weather classes dating from 1881
to present has been constructed by the German Federal Weather Service
(Bardossy and Caspary, 1990) and the scheme of Lamb (1972), which has
formed the basis for construction of a daily sequence of weather classes for
the British Isles dating to 1861. The subjective schemes are primarily based
on large scale features in the surface pressure distribution, such as the location
of pressure centers, the position and paths of frontal zones, and the existence
of cyclonic and anticyclonic circulation types (Bardossy and Caspary, 1990).
Objective classification procedures utilize statistical methods, such as principal components, cluster analysis, and other multivariate methods to develop
rules for classification of multivariate spatial data. For instance, McCabe
(1990) utilized a combination of principal components analysis and cluster
analysis to form classifications of daily weather at Philadelphia. The statistical model was compared to a subjective, conceptual model, which was found
to give similar results. Briffa (Chapter 7 in this book) describes an objective approximation to the Lamb scherne, and uses it for regional validation
of selected GCM control runs. Wilson et al. (1992) explored classification
methods based on K-means cluster analysis, fuzzy cluster analysis, and principal components for daily classification of weather over a large area of the
Pacific Northwest. All of the above methods are unconditional on local conditions, that is, no attempt is made to classify the days in such a way that
local precipitation, for instance, is well-described by the weather classes.
Hughes et al. (1993) used an alternative approach that selected the weather
classes so as to maximize the discrimination of local precipitation, in terms
of joint precipitation occurrences (presence/absence of precipitation at four
widely separated stations throughout a region of dimensions ab out 1000 km).
The procedure used was CART (Breiman et al. , 1984), or Qlassification and
Regression Trees. The large area information was principal components of
sea level pressure. Figure 11.4 shows the discrimination ofthe daily precipitation distribution at one of the stations modeled, Forks, according to weather
class. As expected, because the classification scheme explicitly attempts to
"separate" the precipitation (albeit occurrence/absence rather than amount)
by the selected classes, the resulting precipitation distributions are more distinguishable than those obtained by Wilson et al. Hughes et al. also simu-