Low-cost accurate methods for estimating soil carbon (C) stocks are needed if terrestrial C offset markets are going to be implemented in the United States. Accurately measuring C stocks is often prohibitively expensive due to high spatial variability and analytical costs, therefore the development of cost-effective sampling designs and methods of inference are critical. We evaluated sampling optimization approaches for estimating soil C baseline levels for a dairy farm in Harford, NY with multiple land uses, including cultivation of silage corn, alfalfa hay, pasture, and forest. Three hundred and nineteen samples were collected in a spatially balanced design over a 232 hectare area to a depth of 30 cm. Secondary variables including soil type, elevation, slope, cropping history, and manure application rate were assessed for correlations with soil C and suitability for sampling stratification. Random, stratified, and systematic sampling arrangements at three sampling densities (n = 253, 160, 83), were compared to the full sampling grid (n = 319) using both design-based and modelbased approaches for soil C assessment. Total soil C stocks for the sampling area were estimated by three different approaches: i) spatial mean (SM) where total C stocks are calculated by the area-weighted average of the mean C stocks for each landscape unit; ii) ordinary kriging where the sum of the predicted values for the interpolation grid are used to determine total C-stocks; iii) (SSURGO) where average C stocks are based on estimates from the Soil Survey Geographic database with total C stocks calculated from the area-weighted average for each soil map unit. The systematic sampling arrangement was preferred over random or stratified arrangements because RMSE increased little with the reduced sample size, and the distribution of soil C stocks for the lowest sampling density closely resembled the full sampling grid. Landscape units defined by manure application rate explained the spatial variability of soil C-stocks better than any other categorical variable. Model-based approaches provided more reliable estimates for soil C stocks than design-based approaches. SM resulted in a higher RMSE than OK, 20.7 and 23.1 Mg ha-1 compared to 18.0 and 22.4 Mg ha-1, respectively. Additionally, when the sampling density was reduced from 319 to 83, OK estimates fluctuated less than SM, with mean and total soil C stocks for the entire farm differing by 2% from that of the full sampling grid. Estimates of total C stocks to 30 cm for the entire 232 ha sampling area ranged from 16217-20049 Mg. Modelbased approaches provided the most reliable estimates of soil C stocks. SSURGO based estimates consistently underestimated soil C stocks by 2.6-18.1 % compared to the full grid sampling, but given the low cost this approach may be of interest in some circumstances. ii