More specifically, can it be argued that satellite data are more reliable because they are direct measurements of physical quantities, while reanalysis data depend on the processing and merging of different datasets ?
Unfortunately, your question has no answer. Both have errors. Both can be unreliable. Your choice really depends on how you plan on using them.
Satellites instrument contain sources of error, such as bad observations (wildfires in the NIR), mapping problems (especially near the poles), and representation error. Don't underestimate those sources of error- Geer and Bauer found that representation error can result in errors in brightness temperature larger than 20 K. Reanalysis combines both observation errors and model errors. A good reanalysis will reduce the influence of both errors, but it is still ingrained into the product. Satellites still report back observations.
Reanalyses can be unreliable. For example, this paper shows that TRMM and PERSIANN‐CDR were more reliable than CFSR or ERA-interim.
Maybe more significantly, they both actually report entirely different things. Satellites measure radiances which can be converted into more sensible variables, while reanalyses use a mathematical representation of the atmosphere and observations. So comparing these are a lot like comparing apples and oranges.
Some other thoughts you may want to consider:
- Is balance more important than truth? Reanalyses tend to be more meteorologically complete, therefore balanced. Satellites actually observe the earth system, so they will be more grounded in reality.
- What temporal resolution do you want? Reanalysis outputs are usually once every few hours, while the temporal resolution of satellites vary a large amount, in both number of observations per hour and when observations are taken (such as geostationary vs polar orbiting).
- What horizontal resolution do you want? Satellites may have a horizontal resolution modelers can only dream of (metaphorically speaking), but at the cost of the number of regions (for polar orbiting satellites) or consistency (for geostationary satellites).
- For what amount of time do you want data? The ERA-20C dataset can go back to 1900 and the NOAA/CIRES dataset goes back to 1850, while the first meteorological satellite was put up in 1959.