Spatiotemporal heterogeneity in soil CO2 efflux (FS) underlies one of our greatest gaps in understanding global carbon (C) cycles. Though scientists recognize this heterogeneity, FS sampling schemes often average across spatial heterogeneity or fail to capture fine temporal heterogeneity, and many ecosystem models assume flat terrain. Here, we test the idea that simple, remotely sensible terrain variables improve regression models of spatiotemporal variation in FS. We used automatic chambers that, for the first time, capture FS in complex temperate forest terrain at fine temporal resolution with 177,477 hourly FS measurements at 8 locations from ridgetop to valley along planar and swale hillslopes, across three years ranging from dry to record wet precipitation. In two of these years, we measured FS weekly at 50 additional locations distributed across the 8-ha catchment. Growing season FS estimates were 1.25 times greater when sampling hourly versus weekly. At ridgetops, growing season FS increased by an average of 463 gC m-2 180 day-1 (75.9%) from dry to wet years, while valleys decreased by 208 gC m-2 180 day-1 (-20.1%). This bidirectional response to interannual moisture was identified in distinct Random Forest models of FS for convergent (water accumulating) or non-convergent (water shedding) hillslope positions. We hypothesize that different FS constraints drive these opposing responses—water availably to biota limits FS from ridgetops while slow oxygen diffusion limits FS from wet valleys. Accounting for hillslope position and shape reduces variance of FS estimates in complex terrain, which could improve FS sampling, C budgets, and modeling.