In underwater sensor networks (UWSNs), the unique characteristics of acoustic channels have posed great challenges for the design of medium access control (MAC) protocols. The long propagation delay problem has been widely explored in recent literature. However,the long preamble problem with acoustic modems revealed in real experiments brings new challenges to underwater MAC design. The overhead of control messages in handshaking-based protocols becomes signiﬁcant due to the long preamble in underwater acoustic modems. To address this problem, we advocate the receiver-initiated handshaking method with parallel reservation to improve the handshaking efﬁciency. Despite some existing works along this direction, the data polling problem is still an open issue. Without knowing the status of senders, the receiver faces two challenges for efﬁcient data polling: when to poll data from the sender and how much data to request. In this paper, we propose a trafﬁc estimation-basedreceiver-initiated MAC(TERI-MAC)to solve this problem with an adaptive approach. Data polling in TERI-MAC depends on an online approximation of trafﬁc distribution. It estimates the energy efﬁciency and network latency and starts the data request only when the preferred performance can be achieved. TERI-MAC can achieve a stable energy efﬁciency with arbitrary network trafﬁc patterns. For trafﬁc estimation, we employ a resampling technique to keep a small computation and memory overhead. The performance of TERI-MAC in terms of energy efﬁciency, channel utilization, and communication latency is veriﬁed in simulations. Our results show that, compared with existing receiver-initiated-based underwater MAC protocols, TERI-MAC can achieve higher energy efﬁciency at the price of a delay penalty. This conﬁrms the strength of TERI-MAC for delay-tolerant applications.