Adjustment of Temperature Trends in Land stations After Homogenization (ATTILAH)
Final Report Abstract
Temperature observations from climate stations are affected by inhomogeneities caused by relocations of stations or changes in the measuring techniques. Homogenization algorithms applied to eliminate them consist of two major modules, the detection and the correction part. Additionally, relative homogenization needs to build differences between neighboring stations to reduce the dominating climate signal. For each of these three categories we chose two commonly applied techniques. By combining them, eight prototype homogenization algorithms are constructed and their performance is tested by simulated data. This consists of two superimposed signals, the break signal modeled as a step function with random breakpoints plus noise always remaining even for neighboring stations. In a first paper, we investigated the detection part and showed that the signal-to-noise ratio (SNR), defined as the standard deviation of the breaks divided by that of the noise, is a key parameter for the skill. The fundamental problem is that random segmentations are able to explain already one half of the break variance. For low SNRs, segmentations optimized for the noise explains a maximum of variance, while the additionally explained break variance leads to a spurious significance of the solution. The two alternative methods of detection gave comparable results, although the multiple breakpoint approach is slightly better than the much simpler hierarchical splitting. A second paper is dedicated to the correction part. For the ANOVA method, we found that the network-mean trend error after homogenization is not improved when the quotient k/SNR is larger than 6, where k denotes the number breaks in each station series. However, these are random errors, which could be reduced when many networks are available. Severer are trend biases arising when jumps are predominantly positive or negative. These could be entirely corrected only if the break positions are perfectly known. In realistic cases with detection errors, the trend error is only partly corrected. The alternative correction method investigated is the stepwise jump correction, which gave comparable results. The third category, i. e. the method how to build difference time series, turned out to be the most crucial. In the Composite Reference (CR) approach the average time series of all stations of the network is built and subtracted from the candidate station considered. The Pairwise Differences (PD) approach builds the difference of all possible station pairs separately. However, an attribution step is necessary to assign any break found to one of the two involved stations. These two alternative approaches are tested for their ability to correct a mean trend bias. While the detection and correction methods are held constant, two scenarios are applied. One representing a high station density as found in the U. S. or Germany, the other for low station densities as prevailing in most regions of the globe. The CR approach seems to have general problems with the mean trend bias. In the high density scenario, only 35% of trend bias is corrected, decreasing to 11% for low data density. This means that a potentially found global trend correction has to be multiplied by a large factor in the order of 9. Such a large post-correction is not acceptable. The PD approach works satisfactorily only in the dense data scenario, where it explains 99% of the bias. However, for low station densities, the performance is again weak (34%). Thus, none of the tested prototypes seems provide satisfactory results on global scale. However, it is shown to be sufficient to improve the attribution step of the PD approach. For simulated data, we can easily subtract the climate signal making the difference building obsolete. In this case, 97% of the trend bias is explained even for low data density. Different research groups assume different statistical properties of the break signal. During the project it became clear how important it is, whether the levels between two jumps are considered as independent variable or the jumps themselves. In the first case a Random Deviation (RD) signal results, in the second case a Brownian motion (BM) signal. Two further papers discuss the importance of the break-type assumed. One paper presents a method to derive number and strength of both RD and BM breaks from real data, while the second focus on the very different effect the break-types have on the resulting trend error.
Publications
- 2018: On the reduction of trend errors by the ANOVA joint correction scheme used in homogenization of climate station records, Int. J. Climotol., 38, (14), 5255 – 5271
Lindau, R. and V. Venema
(See online at https://doi.org/10.1002/joc.5728) - 2018: The joint influence of break and noise variance on the break detection capability in time series homogenization, Adv. Stat. Clim. Meteorol. Oceanogr., 4, 1 – 18.
Lindau, R. and V.K.C. Venema
(See online at https://doi.org/10.5194/ascmo-4-1-2018) - 2019: A new method to study inhomogeneities in climate records: Brownian Motion or Random Deviations? Int. J. Climatol., 39, 4769 – 4783
Lindau, R. and V. Venema
(See online at https://doi.org/10.1002/joc.6105) - 2020: Random trend errors in climate station data due to inhomogeneities, Int. J. Climatol., 40, 2393 – 2402
Lindau, R. and V. Venema
(See online at https://doi.org/10.1002/joc.6340)
