Autonomous navigation in unstructured agricultural and forestry environments demands high-fidelity, real-time 3D mapping to support safe and efficient motion planning. This survey systematically reviews dense mapping techniques, spanning grid-based, octree, surfel, mesh, and neural implicit representations, and evaluates their suitability for mobile robots operating under variable terrain and foliage cover. We begin by examining common sensor modalities (monocular/stereo cameras, RGB-D, LiDAR, mmWave radar) and their integration into mapping frameworks such as TSDF/ESDF, OctoMap, Voxblox/Nvblox, and NDT-OM. Next, we detail adaptations of these methods for agricultural and forestry applications. Finally, we address the emerging challenge of estimating solid ground beneath vegetation and constructing traversability maps through statistical models, proprioceptive–exteroceptive fusion, radar sensing, and self-supervised deep learning. By comparing state-of-the-art algorithms and datasets, we identify key trade-offs between memory efficiency, map fidelity, and real-time performance. We conclude with recommendations for integrating uncertainty-aware and active-perception strategies, and outline future directions toward scalable, cloud-based, multi-robot mapping for precision agriculture and forestry management.