Abstract: We investigate the performance of the parametric maximum likelihood component separation method in the context of the cosmic microwave background (CMB) B-mode signal detection and its characterization by small-scale CMB suborbital experiments. We consider high-resolution (FWHM = 8') balloon-borne and ground-based observatories mapping low dust-contrast sky areas of 400 and 1000 square degrees, in three frequency channels, 150, 250, 410 GHz, and 90, 150, 220 GHz, with sensitivity of order 1 to 10μK per beam-size pixel. These are chosen to be representative of some of the proposed, next-generation, bolometric experiments. We study the residual foreground contributions left in the recovered CMB maps in the pixel and harmonic domain and discuss their impact on a determination of the tensor-to-scalar ratio, r. In particular, we find that the residuals derived from the simulated data of the considered balloon-borne observatories are sufficiently low not to be relevant for the B-mode science. However, the ground-based observatories are in need of some external information to permit satisfactory cleaning. We find that if such information is indeed available in the latter case, both the ground-based and balloon-borne experiments can detect the values of r as low as ~0.04 at 95 per cent confidence level. The contribution of the foreground residuals to these limits is found to be then subdominant and these are driven by the statistical uncertainty due to CMB, including E-to-B leakage, and noise. We emphasize that reaching such levels will require a sufficient control of the level of systematic effects present in the data.