Low Earth Orbit (LEO) Satellite communication, with its strengths for wide networks coverages, strong flexibility, and good communication performance, is widely used in communication technologies such as emergency communication. However, due to geographical factors, there may be an uneven distribution of services for Internet of Things (IoT), resulting in congestion and networks load imbalance in certain links. To address this issue, a Traffic-Aware Routing Selection algorithm for LEO IoT is proposed. Firstly, a Convolutional Neural Network-Bidirectional Long Short Term Memory-Attention is utilized to predict the traffic of satellite. Then, the predictive traffic values are incorporated as observations into a Partially Observable Markov Decision Process and input into the Multi-Agent Deep Deterministic Policy Gradient algorithm model for reinforcement learning. Lastly, the effectiveness of the proposed algorithm is confirmed. It is noted that, compared with other algorithms, the proposed algorithm achieved a reduction of 11% to 53% in terms of maximum link utilization as well as a successful decrease in packet loss, thereby improving networks performance.