Semi-natural habitats (SNH) play key roles for arthropod natural enemy communities in agricultural landscapes. Positive relationship between landscape complexity and biological pest control is now well known and is assumed to mainly come from the fact that natural enemies use semi-natural habitats for overwintering. However, the respective role of each type of semi-natural habitats in the landscape in shaping natural enemy communities and pest control remains poorly studied. Moreover, the relative importance of environmental variables in structuring these communities remains largely unexplored. The main purpose of this study was to provide an insight into the types of SNH natural enemies use for overwintering as well as the effects of local and landscape characteristics in structuring their overwintering communities. Overwintering natural enemy communities were sampled in 7 types of SNH (i.e., forest interior (FI), South-facing forest edge (FES), North-facing forest edge (FEN), dry unmanaged grassland (UGD), wet unmanaged grassland (UGW), managed grass strip (CAP grass strip) either dominated by monocotyledonous plants (MGM) or by dicotyledonous plants (MGD)). Abundance, species richness as well as community composition of each group of enemies were then explained by local and landscape parameters to assess their relative importance. In our study, overwintering natural enemy communities differed markedly among types of SNH. Explanatory variables proved to have a decreasing influence in shaping natural enemy community compositions from the local (i.e. in the emergence trap perimeter, in 3 m- and 15 m-radius circular zones around it) to the immediat landscape (within 30 m- and 60 m-radius circular zones) and then the mid-distant one (within 120 m-, 250 m- and 500 m-radius circular zones). We particularly found that management intensity and vegetation height were very strong drivers of natural enemy diversity at the local scale. Managed CAP grass strips turned out as the main source habitat of beneficials in the spring while forests acted quite negatively on local abundances of most of the beneficials studied. On the opposite, medium arable land and grassland surface areas proved to be favourable for them as a whole in the immediat landscape, while in the mid-distant landscape, fallows favoured aphidophagous hoverflies and arable lands did so for spiders. Our results highlight the need for a more precise description of SNH in the landscape if we are to mechanistically understand the role of compositional landscape heterogeneity on zoophagous arthropod populations and to give relevant guidelines to design landscapes favouring natural biological pest control.