Typical remote sensing image interpretation consists of large number of long-lived independent computation jobs such as calibration, correction, and transformation and computation. These time consuming jobs are suitable for execution in desktop grid with high performance commodity PCs for its lower cost and high management ability. In this paper, a framework of parallel remote sensing image interpretation in desktop grid is proposed. The framework consists of modules such as job partitioning and scheduling, load balancing, communication, and image processing. We also provided an implementation based on java multi threaded programming. The computation results on real application of soil moisture estimation workload show that the average speedup is 6.1 in a 9-node heterogeneous desktop grid environment and failure rate of jobs is less than 5.6%.