92nd American Meteorological Society Annual Meeting (January 22-26, 2012)

Wednesday, 25 January 2012: 12:00 AM
Finding Time-Periods of Abrupt Climate Change: A Data Mining Approach
Room 242 (New Orleans Convention Center )
Xun Zhou, Univ. of Minnesota, Minneapolis, MN; and S. Shekhar and S. Liess

Finding Time-Periods of Abrupt Climate Change: A Data Mining Approach

Abrupt climate change is defined as an unusually large shift of precipitation, temperature, etc, that occurs during a relatively short time period. Identifying such a period is a crucial step for understanding and attributing climate change. In this work, we analyze historical precipitation data series using a novel, automated data mining approach to identify time-periods of persistent, abrupt precipitation decrease (increase). We propose a statistical model called sameness degree to quantitatively evaluate the change abruptness and persistence in a time period. Being a statistical function of all the piecewise change rates in a time series, the sameness degree gives a value between 0 (least persistent or abrupt change) and 1 (most persistent abrupt change). We then design an algorithm to exhaustively examine all the time-periods. Given a user specified threshold, the algorithm finds all the time periods that have sameness degree greater than the threshold, and are not subsets of others (to remove small fluctuations).

We evaluate the proposed method with the Climate Research Unit (CRU) precipitation data, whereby we focus on the Sahel rainfall index. Results show that this method can find periods of persistent and abrupt precipitation changes with different temporal scales.

We also further optimize the algorithm using a smart searching strategy, which always evaluates longer time-periods before its subsets. By doing this, we reduce the computational cost to only one third of that of the original algorithm for the above test case. More significantly, the optimized algorithm is also proven to scale up well with data volume and patterns numbers. Particularly, it achieves better performance when dealing with longer patterns.

Supplementary URL: