عنوان مقاله [English]
Sometimes, the showers of the northern coast of Persian Gulf are very heavy and disastrous and
hazardous. They cause heavy damages to the people and the infrastructures of the region.
Therefore the economic development of the area is highly dependent upon the identification of
the cause and management of these hazardous phenomena. The main factor controlling the
surface climate is the pressure patterns of the atmosphere. Therefore, the main objective of this
study is to identify the synoptic patterns of these showers. Thus, we can predict their occurrence
and mitigate their damages. The successful achievement is dependent on two major factors:
(a) the methodology of pattern recognition and (b) identification of actual patterns. Most of the
models of pattern analysis are linear while the atmospheric processes are non-linear in nature.
Any methodology that neglects the nonlinear nature of atmospheric phenomena would result in
inadequate classification of atmospheric circulation. For this reason, this research has used the
nonlinear models of classification algorithms to identify the pressure patterns of the heavy rains
of the area.
The study was based on the hypothesis that the daily atmospheric circulation can be explained
by the geo-potential height of 500 hPa level, precipitable water, and the velocity of vertical
∗E-mail: firstname.lastname@example.org Tel: +98 9129376680
Physical Geography Research Quarterly, 46 (3), Fall 2014 5
patterns and heavy rainfall, the data have been collected through 15340 days (1966-2007) for
these three variables of 289 grid-points, with a resolution of 2.5 degrees, from NCEP database.
Daily rainfall data for the same period have also been gathered for Bandar-e-Abbas, Bandar-e-
Lengeh, Boushehr, and Abadan stations from Meteorological Organization of Iran. First, the
daily circulations as micro-patterns have been classified using self-organizing map (SOM)
algorithm, a type of unsupervised neural network. This algorithm begins to calculate the
Euclidian Distance between an input vector and all of the weight vectors to find the 'winner' unit
(BMU) with the weight vector closest to the input vector. The calculation continues to update all
the weight vectors, especially those within neighbouring radius .The iterative calculation
proceeds towards the projection of similar data samples in the high dimensional, complex input
data space to an identical unit area in the map. As a result, the neighbouring units in the map are
similar to each other while distant units are dissimilar. Then, the U*-matrix, as a suitable
method for two-dimensional visualization of the trained SOM that enabled us to recognize the
degree of the similarity among adjacent units in the two-dimensional map, was employed to
identify boundaries among clusters and to extract the actual number of meso-patterns. Finally,
K-means method was utilized to cluster these meso- patterns into distinguished macro patterns.
Results and Discussion
The results revealed that SOM, by classifying the micro-patterns into 289 meso-patterns, could
discriminate the days of warm and cold periods with an accuracy of more than 99 percent.
These patterns were classified into 11 macro patterns through the U*-matrix and K-means
models. Through displaying the number of heavy rainfall events in each station on each unit of
SOM, it was specified that four macro-patterns explained up to 83 % of heavy rainfall events of
the region. These patterns are named as follows: Pattern No. 4 as Syria trough becomes deeper,
Siberian high pressure moves towards west, and the moisture of Arabian and Oman Seas move
to PG. Similarly, the identification of pattern 6 is possible by Sudan low, subtropical jet stream
velocity increase, and its base decreases. Pattern 7 is identified by cut-off low system, very low
pressure, and closed low up to upper troposphere. Pattern 9 is specified by two characteristics:
(a): the simultaneous presence of warm and cold season components of atmosphere during the
seasonal change, and (b) dense isobars over PG.
On the basis of the results, we concluded that the combination of SOM classification method,
U*- matrix and K-means clustering methods can be employed as an appropriate instrument to
classify nonlinear atmospheric variables, in one hand, and to resolve the problem of extracting
the actual synoptic patterns, on the other. Of the four synoptic patterns of heavy rainfall, cut-off
low and seasonal transition patterns should be taken into account more seriously because of the
persistence and startling nature of their heavy rainfall as well as the vulnerability of society for
the probable damage.