Underwater wireless communication is subject to considerable limitations, caused by variable channels, latency, and low bandwidth. Classical resource allocation methods are inadequate to accommodate these rapidly changing circumstances. This work proposes a Deep Learning–Assisted Adaptive Resource Allocation (DL-ARA), which combines a hybrid CNN-LSTM model with a multi-step prediction factor for the channel, along with meta-reinforcement learning algorithms, specifically Soft Actor-Critic (SAC) and Model-Agnostic Meta-Learning (MAML). The link quality, node energy, and state of the queue predicted values are used to dynamically allocate the power, bandwidth, and time slots. In experiments, DL-ARA achieved a maximum throughput of 85.3 kbps, energy consumption of approximately 1.8 × 10⁻3 J/bit, and a packet error rate of 4.7%, significantly exceeding the baselines of static water-filling and standard DRL. Adaptation time was reduced from 12.0 s (DQN) to 3.2 s. Overall, these results show that DL-ARA offered a robust, energy-efficient, and low latency solution, establishing DL-ARA as a strong candidate for future real-time underwater communication systems..