Abstract: Reverse time migration (RTM) is a powerful tool for depth-domain imaging to reconstruct geologically complex deep targets. Deep seismic imaging, particularly through wave-equation-based RTM, is computationally intensive owing to the large-scale migration velocity model from shallow to deep zones. When near-surface velocities are low or a high-frequency wavelet is used, conventional coarse grid sampling often proves inadequate, leading to numerical dispersion and consequent low imaging accuracy. Conversely, fine grid sampling achieves high-precision imaging at the sacrifice of significantly increased computational costs and memory requirements. We propose a RTM method based on adaptive spatial grids. The approach consists of three key steps: adaptive grid partitioning of the migration velocity model, finite-difference wave-equation simulation of forward-time source wavefield and reverse-time receiver wavefield based on adaptive spatial grids, and transformation back to the conventional-grid coordinate system to obtain the final RTM image. Automatic grid refinement in shallow low-velocity zones effectively alleviates dispersion and enhances imaging quality, and adaptive grid coarsening in deep high-velocity zones reduces grid count, computational load, and memory requirements. Model and field data tests demonstrate the accuracy and computational efficiency of this method.