The discrete wavelet transform (DWT)-based Set Partitioning in Hierarchical Trees (SPIHT) algorithm is widely used in many image compression systems. The time-consuming computation of the 9/7 discrete wavelet decomposition is usually the bottleneck of these systems. In order to perform real-time Reed-Solomon channel decoding and SPIHT+DWT source decoding on a massive bit stream of compressed images continuously down-linked from the satellite, we propose a novel graphic processing unit (GPU)-accelerated decoding system. In this system the GPU is used to compute the time-consuming inverse DWT, while multiple CPU threads are run in parallel for the remaining part of the system. Both CPU and GPU parts were carefully designed to have approximately the same processing speed to obtain the maximum throughput via a novel pipeline structure for processing continuous satellite images. As part of the SPIHT decoding system, the GPU-based inverse DWT is about 158 times faster than its CPU counterpart. Through the pipelined CPU and GPU heterogeneous computing, the entire decoding system approaches a speedup of 83x as compared to its single-threaded CPU counterpart. The proposed channel and source decoding system is able to decompress 1024x1024 satellite images at a speed of 90 frames per second.