We have a cloud mask but we don't have reference data set to assess its accuracy. Without generating a reference dataset by ourselves is it possible to assess the cloud mask's accuracy?? For example; can we compare the spectral signatures of the rasters which were taken a cloud-free day and the mask applied raster? or can we look the spectral signature values correlation between 2 raster? (one is cloud-free, second is the mask applied raster?
As far as Sentinel-2 is concerned (as suggested by the tags), reference cloud masks exist. See for instance https://zenodo.org/record/1460961#.YSyWLkw69aQ
Still with Sentinel-2, it is also possible to compare two different cloud masks methodologies by comparing the smoothness of reflectance time series. We used such a criterion to compare options in atmospheric correction in the following reference .