For climatological studies, it is very important to address the issue of threshold selection. For extreme rainfall/temperature time series, apart from block Maxima series, POT (peak over threshold) series also provides vital informed. The question arises about selection of threshold value for analysis. Some literatures state that the threshold value shall be computed corresponding to a base period which is consistent with recent World Meteorological Organisation (WMO) operational climatology base period, i.e., at present 1961-1990 or 1971-2000 is the current base period adopted by WMO. For this base period, threshold values corresponding to 95th or 99th percentile values of a variable may be derived by fitting suitable probability distribution function.
The question is whether this method is robust for such estimation. If my data length of concern is 1901-2013, then the threshold values determined for the 'base period say 1961-1990' would be applicable/suitable to analyse the change during other periods, say for the years before 1961 and after 1990?
Please suggest or advice in this regard. Thanks in advance.
Priyank, I would say that while you should consider using conventions such as a 1971-2000 base period, there is no need to be a slave to them. Since you have data spanning a much longer time, (1901-2013), it probably makes sense to treat that full 113 years as your base period to start with - unless there is a clear trend over that time. If there is such a clear trend, then the question of what section of the period to take as a "base" becomes harder - but still mostly arbitrary. You could de-trend your data (in fact you should probably do this anyway), and then look at extremes (2 or 3 standard deviations, or the equivalent on a gamma distribution of precipitation) relative to the de-trended series. Once your data is de-trended, any 30-year base period should be equivalent to any other in any case.
However, running statistical time-series analyses tends to be fairly fast, so it should not be too hard to experiment with different base-periods, and different base-period lengths. You should be guided by the data themselves, and do whatever makes most sense to extract the signals in them from the noise. E.g., some phenomena are "oscillating", and can be fruitfully analyzed in frequency space (via Fourier transform) - think of energy spectra - while other phenomena are more "episodic" (e.g., extreme rain events), and are better analyzed with "composites" rather than frequency spectra. Choose your statistical too to suit the characteristics of your data.
Dear Priyank, in the case of more than 100 years for rainfall data, I believe the multi-annual trend can be clear enough and also you should estimate the rainfall thresholds by using the whole data series and results may be much better than thresholds estimated based on a certain period. In the case of annual TRS values, you may decide the highest / the second highest value of each year to subset the annual maximums (for rainfall data).
Tanja Porja Thank you very much for your response. I too feel that consideration of entire period is more plausible than a certain sub-period.
- Badges
- Science topic
- Similar topics
- Geoscience
- Atmospheric Sciences
- Climate Science
- Climate Change