World Climate Classification and Search: Data Mining Approach Utilizing Dynamic Time Warping Similarity Function
Abstract:We have developed a novel method for classification and search of climate over the global land surface excluding Antarctica. Our method classifies climate on the basis of the outcome of time series segmentation and clustering. We use WorldClim 30 arc sec. (approx. 1 km) resolution grid data which is based on 50 years of climatic observations. Each cell in a grid is assigned a 12 month series consisting of 50-years monthly averages of mean, maximum, and minimum temperatures as well as the total precipitation.
The presented method introduces several innovations with comparison to existing data-driven methods of world climate classifications. First, it uses only climatic rather than bioclimatic data. Second, it employs object-oriented methodology – the grid is first segmented before climatic segments are classified. Third, and most importantly, the similarity between climates in two given cells is performed using the dynamic time warping (DTW) measure instead of the Euclidean distance.
The DTW is known to be superior to Euclidean distance for time series, but has not been utilized before in classification of global climate. To account for computational expense of DTW we use highly efficient GeoPAT software (http://sil.uc.edu/gitlist/) that, in the first step, segments the grid into local regions of uniform climate. In the second step, the segments are classified.
We also introduce a climate search – a GeoWeb-based method for interactive presentation of global climate information in the form of query-and-retrieval. A user selects a geographical location and the system returns a global map indicating level of similarity between local climates and a climate in the selected location. The results of the search for location: "University of Cincinnati, Main Campus" are presented on attached map. The results of the search for location: "University of Cincinnati, Main Campus" are presented on the map.
We have compared the results of our method to Koeppen classification scheme as well as to existing data-driven climate classification methods and concluded that our DWT-based method provides the new, superior analytical framework for studying climate variability on global scale.