Wenche Aas, Hilde Fagerli, Svetlana Tsyro, Sverre Solberg Comparing trends from the benchmark data set (S,N,O3 and PM) with EMEP model results Part I. Benchmark dataset Wenche Aas, Hilde Fagerli, Svetlana Tsyro, Sverre Solberg

Outline Calculating trends Benchmark dataset the use of Mann Kendall statistics How long time serie is needed to detect a trend? Benchmark dataset Some results on seasonal differences

Statistical analyses, Mann Kendall Basic questions to ask are To what extent will the method detect trends for the given time periods? How reliable are the estimated Sen slopes? What is the effect of missing data? Constructed artificial time series reflecting two different processes leading to variability in atmospheric observational data: A linear long-term trend due to changes in emissions and hemispheric baseline levels; and A “natural variability” due to variations in meteorology from year to year A note by Sverre Solberg available on wiki page)

Results from Mann Kendall statistics on artificial time series annual trend of -1, -2, and -3 % /year calculations of 1000 artificial time series with 11 and 23 years natural variability corresponding to a st.dev of 0.05, 0.10 and 0.15 of the mean the probability that the Mann Kendall detects a significant trend (p=0.05) and the average Sen’s slope

The length of the time series is critical for the performance of the Mann Kendall statistics. The smaller the trend is relative to the inter-annual meteorological variability, the longer time series is needed in order to identify a significant trend. With only 11 years of data, it is difficult to detect trends, except for the cases with a small annual variability and a marked trend (3 %/y or stronger in absolute level, risk of overestimated trends Use less stringent sign. citeria for data with high variability? With 23 years of data, the situation is very different, and the method will most certainly detect the trend, and the estimated slopes will be close to the real trend

First set of the benchmark dataset Same sites as used in Tørseth et al (2012) though added the extra years Uploaded in wiki.met.no in Dec 2014

Feedback and follow up Missing datasets (NL, DE44, RU, (SK)) No comments on the actual data used Decided to include more sites: not only sites with co-located air and precip measurements of sulphur For the shorter periods (1990-2001 and 2002-2012) more sites than 1990-2012. Uploaded extended dataset and seasonal trends on wiki (in April)

Overview results (S,N,PM)

Seasonal differences in trends

Some remarks before conclusions Time resolution (annual, season average) ok? Higher time resolution (monthly, weekly, daily) may help detecting trends For filtering based on sectors etc, original resolution needed Stricter criteria for site selection based on quality? Revisit “strange” data with national experts Change in method may influence trends, Wrong units Change in surroundings Use outlier test to exclude years or time series?