The Underwater Wireless Optical Communication (UWOC) has gained significant attention due to its high data transmission rates. However, the complexity of the channel, caused by absorption and scattering in water, poses challenges for accurate channel estimation (CE) and signal detection (SD). This paper proposes an approach for UWOC channel preclassification and direct detection in optical orthogonal frequency-division multiplexing (OOFDM) systems using a memory-convolution-enhanced bidirectional network (LCNN-BiLSTM). The LCNN-BiLSTM is trained offline on simulated data generated from UWOC channels in various types of water environment. Under unknown channel conditions, the long short-term memory neural network (LSTM) identifies the channel, while the LCNN-BiLSTM compensates for signal distortions, implicitly estimating the channel state information (CSI) and directly recovering the transmitted data. This paper investigates the effects of parameters such as pilot signals, water types, communication distance, and data rates on the bit error rate (BER). The simulation results demonstrate the superior performance of the proposed system in complex UWOC channel environments.