Traditional multiple elimination methods face challenges in processing large-scale datasets due to poor scalability and low resource utilization. This paper proposes an efficient heterogeneous acceleration framework for the surface and interbed multiple elimination (SAIME) method over CPU-GPU Platforms. Our achievements include: (1) A 3-stage pipeline parallel strategy combining double-buffered I/O prefetching and CUDA streams is proposed, achieving 91% I/O bandwidth utilization through computation-communication-storage overlap. (2) A memory-optimized prediction operator using shared memory tiling and mask precomputation achieves a 107 \(\times \) acceleration for the kernel function. (3) A dynamic load-balancing strategy enabling concurrent CPU/GPU task execution is designed. Results show that Hybrid-SAIME obtains a 5.17–5.91 \(\times \) end-to-end speedup over baseline SAIME. This work provides valuable references for storage-computation-network co-optimization in next-generation seismic imaging systems.