Open AccessA Novel Approach for the Detection of Inhomogeneities Affecting Climate Time Series
Author(s) -
Andrea Toreti,
Franz G. Kuglitsch,
Elena Xoplaki,
Jürg Luterbacher
Publication year - 2012
Publication title -
journal of applied meteorology and climatology
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.079
H-Index - 134
eISSN - 1558-8432
pISSN - 1558-8424
DOI - 10.1175/jamc-d-10-05033.1
Subject(s) - computer science , homogenization (climate) , series (stratigraphy) , homogeneity (statistics) , metadata , identification (biology) , data mining , markov chain , variance (accounting) , context (archaeology) , segmentation , time series , econometrics , climatology , machine learning , artificial intelligence , mathematics , geography , geology , biodiversity , paleontology , botany , accounting , archaeology , business , biology , operating system , ecology
Sudden changes caused by nonclimatic factors (inhomogeneities) usually affect instrumental time series of climate variables. To perform robust climate analyses based on observations, a proper identification of such changes is necessary. Here, an approach (named the “GAHMDI” method, after its components and purpose) that is based on a genetic algorithm and hidden Markov models is proposed for detection of inhomogeneities caused by changes in the mean and variance. Simulated series and a case study (winter precipitation from a weather station located in Milan, Italy) are set up to compare GAHMDI with existing methodologies and to highlight its features. For the identification of a single changepoint, GAHMDI performs similarly to other methods (e.g., standard normal homogeneity test). However, for the identification of multiple inhomogeneities and changes in variance, GAHMDI returns better results than three widespread methods by avoiding overdetection. For future applications and research in the homogenization of climate datasets (temperature and precipitation) the use of GAHMDI is encouraged, preferably in combination with another detection procedure (e.g., the method of Caussinus and Mestre) when metadata are not available. Since GAHMDI is developed in the generic context of time series segmentation, it can be applied to series of generic variables—for instance, those related to economics, biology, and informatics.