Using an expert deviation carrying the knowledge of climate data in usual clustering algorithms.

摘要: In order to help physicists expand their knowledge of the climate in Lesser Antilles, we aim identify spatio-temporal configurations using clustering analysis on wind speed and cumulative rainfall datasets. But show that L2 norm conventional methods as K-Means (KMS) Hierarchical Agglomerative Clustering (HAC) can induce undesirable effects. So, propose replace Euclidean distance (L2) by a dissimilarity measure named Expert Deviation (ED). Based symmetrized Kullback-Leibler divergence, ED integrates properties observed physical parameters knowledge. This helps comparing histograms four patches, corresponding geographical zones, are influenced atmospheric structures. The combined evaluation internal homogeneity separation clusters obtained was performed. results, which compared silhouette index, five with high indexes. For two available datasets one see that, unlike KMS-L2, KMS-ED discriminates daily situations favorably, giving more meaning discovered algorithm. effect patches is spatial representative elements for KMS-ED. able produce different makes usual structures clearly identifiable. Atmospheric interpret locations impact each cluster specific zone according KMS-L2 does not lead such an interpretability, because represented spatially quite smooth. climatological study illustrates advantage new approach.