We describe the implementation and evaluation of a ghost-point immersed boundary method (GPIBM) in a high-order finite-difference large-eddy simulation code. This study is motivated by the need to simulate flow in the atmospheric boundary layer over hills or cliffs, which require the capability to simulate flow in a Cartesian box with an uneven bottom boundary. The usual GPIBM technique, where the ghost points are populated by mirroring flow variables values at the image points, is not sufficient. This is because the underlying numerical schemes are global in nature and have an effective stencil width that is as large as the number of points in the domain. This requires that flow variables at all points in the solid region are populated with flow variables such that the solution extends smoothly from the solid to the fluid region with appropriate conditions enforced at the immersed boundary. A Laplacian smoothing technique is coupled with the usual GPIBM to achieve this extension of flow fields to all points in the solid region. The methodology is investigated for three problems: Flow over a sphere immersed in a Tayloy-Green vortex, a flow over a circular cylinder and flow over a cosine-shaped hill. Reasonably accurate results are obtained with the current implementation. These results point to the need for improved smoothing techniques so as to improve the order of accuracy and enable accurate comparisons with benchmark experimental and numerical results. © 2023, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.