A non-negligible source of systematic bias in cosmological analyses of galaxy
surveys is the on-sky modulation caused by foregrounds and variable image
characteristics such as observing conditions. Standard mitigation techniques
perform a regression between the observed galaxy density field and sky maps of
the potential contaminants. Such maps are ad-hoc, lossy summaries of the
heterogeneous sets of co-added exposures that contribute to the survey. We
present a methodology to address this limitation, and extract the spurious
correlations between the observed distribution of galaxies and arbitrary stacks
of single-epoch exposures. We study four types of galaxies (LRGs, ELGs, QSOs,
LBGs) in the three regions of the DESI Legacy Surveys (North, South, DES),
which results in twelve samples with varying levels and type of contamination.
We find that the new technique outperforms the traditional ones in all cases,
and is able to remove higher levels of contamination. This paves the way for
new methods that extract more information from multi-epoch galaxy survey data
and mitigate large-scale biases more effectively.