In response to the insufficient modeling of spatiotemporal heterogeneity under a single scale in water-floating garbage trajectory prediction, which often leads to high uncertainty in results, a multi-scale adaptive graph convolutional model (MAGC-Trajectory) was proposed. First, an adaptive gated graph convolution module was developed. Static adjacency relationships imposed by spatiotemporal priors were cross-domain integrated with data-driven dynamic topological structures. This enhanced the model’s ability to capture temporal dependencies of garbage drift and fluctuations of its trajectories. Next, a multi-scale spatiotemporal interaction module was designed. Spatial features were decoupled along different temporal scales. They were then fused in a weighted manner with temporal features. This strengthened the representation of spatiotemporal heterogeneity in garbage trajectories. Meanwhile, an improved nonlinear learning layer was introduced. A learnable adaptive activation function was employed to reinforce global fusion of multiscale spatiotemporal features. This resulted in high-order drift trajectory features with a unified representation. Lastly, a probabilistic prediction layer was constructed. Mean-variance estimation was utilized to estimate the trajectories’ distribution interval. This quantified predictive uncertainty and thus provided more robust predicted trajectories. Experiments were conducted on the water-floating garbage trajectory dataset. Results showed that, compared with the baseline model, the proposed approach reduced MAE by 0.000 2 and RMSE by 0.000 5. These improvements provide actionable support for decision-making in mitigating water-floating garbage pollution within the study area.