自組織映射圖網路於水文屬性均一區劃分之研究:以設計雨型為例

There are two crucial processes for the delineation of homogeneous areas of
hydrological properties. The first process is the cluster analysis on certain
hydrological properties. Obviously, different numbers of clusters lead to
significantly different results. However, determination the proper number of clusters
is a subjective task for conventional cluster analysis methods no matter what
clustering technique, the hierarchical method or the nonhierarchical method, is used.
The second crucial process for the delineation of homogeneous areas is to establish a
method to extrapolate hydrological data from sites, at which records have been
collected, to other sites at which values are required but unavailable, which is the goal
of the next year’s research. In this year’s study, the self-organizing maps (SOM) are
introduced for the cluster analysis of hydrological properties. The SOM, which can
project high dimensional patterns into a lower dimensional space, keep their original
topological structures and capture the intrinsic statistical features of the patterns, is
one of the unsupervised and competitive neural networks models. A clustering
technique, self-organizing maps cluster analysis (SOMCA), is proposed and applied
to the cluster analysis of the simple scaling Gauss-Markov (SSGM) design
hyetographs of northern Taiwan. From the results, the advantage of the SOMCA
over the conventional cluster analysis methods is that the proper number of clusters
2
can be easily and objectively selected by visually inspecting the classification map
derived from the SOMCA. In summary, the achievements of this study significantly
contribute to the clustering techniques of hydrological properties.