3D sensing and perception for precision weed management: a review of modalities, reconstruction pipelines, and context-aware detection

• LiDAR-RGB fusion gives most reliable outdoor 3D weed detection in dense crop canopies. • SfM with MVS delivered the lowest reconstruction error in dense crop scenes. • Neural implicit modeling and volumetric fusion captured fine leaf and stem structures. • Generalization fails across fields with crop variety, soil, occlusion, and weed stages. • Multi-site RGB-D with SEP labels improve sensing across diverse weeds and crop types. Weeds are known for rapidly invading crops, draining vital resources, and causing major yield losses, yet it remains difficult to control them without herbicides. Intensive herbicide use leaves residues, promotes weed resistance, and poses risks to the environment and human health. Manual weed removal is effective but labor-intensive and unsustainable in large-scale farming. Precision weed management offers a sustainable alternative but relies on accurate three-dimensional (3D) detection and localization to identify weeds within dense canopies. Capturing depth, shape, and position enables precise targeting, minimal crop injury, and supports autonomous control. This review seeks to investigate how existing 3D reconstruction technologies and perception techniques contribute to precision weed control. It integrates studies across sensors such as light detection and ranging (LiDAR), stereo vision, and monocular RGB, and algorithmic pipelines including Structure-from-Motion (SfM), Multi-View Stereo (MVS), volumetric fusion, and neural implicit models. Literature was gathered through selective searches in leading scientific databases using Boolean keyword combinations and strict inclusion criteria for weed-focused 3D applications. Comparative and functional analysis then measured sensor capability, algorithmic accuracy, and readiness for field deployment. Key challenges identified were limited annotated datasets, poor transfer across crop stages, occlusion in dense canopies, and high computational demands that slow field deployment. Further advances in sensor fusion, lightweight neural models, real-time mapping, and standardized metrics will enable field-ready 3D vision technologies for weed management.