Assume you have lots of data measured by weather stations, but unfortunately their temporal coverage is not sufficient to compute a full 30 years climatology (e.g. over the 1991-2020 period). However, you still want to compute some reference for the station which will allow one to get an idea on how warmer/colder or drier/wetter a certain period (week, month, season, year...) was.
I came up with a cool trick, that I believe someone else in the literature used, although I could not find evidence anywhere. Using reanalysis (ERA5-Land) data on a period where I have coverage from both model and station data I can attempt to build a relationship that links the two: can be something simple from a linear regression to something more complicated like find a SVM model that fits the closest grid points values from the reanalysis to the station data.
Once this "model" is found, I can use it on the reanalysis data of the period 1991-2020 to get, as output, a "fictitious" climatology for the station. This works pretty well for temperature, as it has a clear seasonal cycle and no distinct day-to-day variability, but fails completely to capture the precipitation sudden changes, maybe also due to the fact that reanalysis are hardly capturing local precipitation features.
Does someone have any literature suggestion that could make me improve the model?
Dear Guido,
Many thanks for your question. Try to check the ERA5-Land (and other reanalysis datasets such as NASA POWER) with only years when station data are available. Then you can assume that the reanalysis is good to fill in years when station data are not covered.
Best Regards
Samy
Dear Guido,
Reanalysis can be one of the options on the table that one can rely on to fill the gaps in the station observations (option 1) as Samy Ashraf Anwar suggested. But of course, this entails that the reanalysis data reasonably compare to the observations to implement. You may want to undertake sufficient statistical comparison and in some cases take some remedial actions such as correcting the reanalysis for systematic errors before using the reanalysis dataset.
As you mentioned in your question, you have many station measurements, which also means that you can just consider filling the gaps from neighboring stations (option 2). In this case, you do not need the reanalysis data.
If you are interested only in the climatology of the stations, then you can experiment on the effects of varying record length (arising from the data gaps) on the calculated climatology of the stations. We did this for temperature observations in Ethiopia and found that the mean uncertainty in the calculated climatology arising from the data gaps is actually small (see Fig.2c in the attached paper). This gives another option to try: computing the climatology from the station observations although they have gaps and then adding the uncertainty to the calculated climatology. Article Gridded daily 2-m air temperature dataset for Ethiopia deriv...
Dear Guido Cioni,
Recently, I successfully applied a PCA-based approach to fill in missing daily minimum and maximum temperature data (DOI: 10.3390/atmos13081186). The main advantage is in avoiding the reanalysis bias. I think this approach can be cleverly used for precipitation data if, firstly, one recovers monthly precipitation and then disaggregates it to daily precipitation following the regime of the closest neighboring stations.
Best regards,
Krastina
Hi Guido Cioni,
just some questions to clarify your situation:
Which region exactly are we talking about?
Which meteorological parameters are to be analysed?
What is the station availability or station density in the region?
Depending on the station availability, the best model probably will be a model using neighbouring stations as predictors. Maybe you have a look at https://doi.org/10.1127/metz/2018/0908, or the method applied to enhance time series e.g. for precipitation time series: https://publikationen.sachsen.de/bdb/artikel/40314(German language only)
Yes, the use of reanalyses can be as consistent as the use of observable data for comparative research purposes, as long as the quality and accuracy of the reanalyses are evaluated and verified in comparison with observable data. Reanalyses are generated using climatic models and assimilated atmospheric and oceanic observations, and the uncertainties associated with these processes should be taken into consideration. It is important to critically analyze the quality and limitations of reanalysis data before using them. During my studies in the Climate and Environment Master's program at the Federal Institute of Santa Catarina - IFSC - Brazil, we identified that precipitation values in the reanalyses were not the same as in observable data. However, by analyzing the variation of precipitation values using moving averages and percentiles, we found that the variations were similar in both reanalyses and observable data. These results show that it is possible to use reanalyses for comparative analysis, as long as caution is taken in the data analysis. Unfortunately, our studies have not yet been published, but I will verify if it is possible to make them available to enrich the discussion in your research proposal.
Hi Guido,
have a look at our papers, with this kind of tools we have reconstructed the whole climatology of the Umbria region.
Best.
Paolina
Reanalysis data is a useful tool for reconstructing the climatology of a station. However, the accuracy of the reconstructed climatology depends on the quality of the reanalysis data and the availability of station data. The accuracy of reanalysis data has improved over time with the development of new techniques and the availability of more observations. A study published in Nature evaluated the spatial-temporal variation performance of ERA5 reanalysis data and found that ERA5 has high accuracy in reproducing the spatial-temporal variation of temperature and precipitation.
- Badges
- Science topic
- Similar topics
- Phytochemistry
- Extracts