Filter bank-based multicarrier (FBMC) systems based on offset QAM (FBMC/OQAM) have recently attracted increased interest in several applications due to their enhanced flexibility, higher spectral efficiency, and better spectral containment compared to conventional OFDM. They suffer, however, from an inter-carrier/inter-symbol interference that complicates signal processing tasks such as channel estimation. Most of the methods reported thus far rely on the assumption of (almost) flat subchannels to more easily tackle this problem, addressing it in a way similar to OFDM. However, this assumption may be often quite inaccurate, due to the high freq. selectivity of the channel and/or the small number of subcarriers employed to cope with frequency dispersion in fast fading. In such cases, severe error floors are exhibited at medium to high SNR values, which cancel the advantage of FBMC over OFDM. Moreover, the existing methods provide estimates of the subchannel responses, most commonly in the frequency domain. The goal of this paper is to revisit this problem through an alternative formulation that focuses on the estimation of the channel impulse response itself and makes no assumption on the degree of frequency selectivity of the subchannels. The possible gains in estimation performance offered by such an approach are investigated through the design of optimal (in the MSE sense) preambles, of both the full and sparse types, and of the smallest possible duration of only one pilot FBMC symbol. Existing designs for flat subchannels are then shown to result as special cases. Longer preambles, consisting of two consecutive pilot FBMC symbols, are also analyzed. The simulation results demonstrate significant improvements from the proposed approach for both mildly and highly frequency selective channels. Most notably, no error floors appear anymore over a quite wide range of SNR values.