K-Gaps: a novel technique for clustering incomplete climatological time series

Leopoldo Carro-Calvo,Fernando Jaume-Santero,Ricardo García-Herrera,Sancho Salcedo-Sanz

doi:10.1007/s00704-020-03396-w

Abstract

In this paper, we show a new clustering technique (k-gaps) aiming to generate a robust regionalization using sparse climate datasets with incomplete information in space and time. Hence, this method provides a new approach to cluster time series of different temporal lengths, using most of the information contained in heterogeneous sets of climate records that, otherwise, would be eliminated during data homogenization procedures. The robustness of the method has been validated with different synthetic datasets, demonstrating that k-gaps performs well with sample-starved datasets and missing climate information for at least 55% of the study period. We show that the algorithm is able to generate a climatically consistent regionalization based on temperature observations similar to those obtained with complete time series, outperforming other clustering methodologies developed to work with fragmentary information. k-Gaps clusters can therefore provide a useful framework for the study of long-term climate trends and the detection of past extreme events at regional scales.

Highlights

Marked variations in regional climate patterns arise as a response to persistent changes of the climate system
The performance of k-gaps has been compared with that of two other clustering techniques: the k-POD algorithm (Chi et al 2016) and the k-means algorithm
We have presented a novel clustering technique for incomplete datasets known as k-gaps

Summary

Introduction

Marked variations in regional climate patterns arise as a response to persistent changes of the climate system. Within this framework, classical clustering techniques, such as the k-means algorithm (Hartigan and Wong 1979; Phillips 2002), have become widespread in the past few years as dimensionality reduction methods are able to extract relevant information from extensive climate databases (Bernard et al 2013; Bador et al 2015; Zhang et al 2016). Classical clustering techniques, such as the k-means algorithm (Hartigan and Wong 1979; Phillips 2002), have become widespread in the past few years as dimensionality reduction methods are able to extract relevant information from extensive climate databases (Bernard et al 2013; Bador et al 2015; Zhang et al 2016) These methodologies can arrange data according to their internal structure by defining spatial regions for datasets with geolocated climate information (Rao and Srinivas 2006). Pattern recognition of climate trends has been explored by cluster analysis of individual samples (DeGaetano 2001), which provided valuable information about the evolution of regionally located climate variables (Scherrer et al 2016)

Objectives

Results

Discussion

Conclusion