Quadtree-based terrain visualization methods have been used in a lot of applications. However, because most procedures are performed on the CPU, the rendering speed is slow in comparison to methods using GPU. In this paper, we present a quadtree-based terrain visualization method working on the GPU with specially designed data structure, error-texture and LOD-texture, and block-based acceleration method. In preprocessing step, we calculate errors in world space and store them to error-texture. In rendering step, we examine projected errors of error-texture and choose the detail level, then store the projected errors to LOD-texture. View frustum culling is performed as block unit using the values of error-texture and LOD-texture. This method reduces CPU load and performs time consuming jobs such as LOD selection and view frustum culling.