Transformer-based cross-view LiDAR–Orthomosaic fusion for geo-localization and digital Modeling in apple orchards

Precision agriculture increasingly relies on accurate large-scale localization and digital modeling for autonomous tasks in orchards. However, map drift and localization uncertainty under GNSS-limited agricultural settings remain challenging due to canopy occlusion and repetitive row geometry. This study introduces a novel cross-view fusion framework that integrates drone remote-sensing imagery (RSI) with ground LiDAR-IMU odometry (LIO) to achieve robust geo-localization and high-fidelity digital modeling in GNSS-denied orchard environments. We utilized a transformer-based matching network to learn modality-invariant structural cues and predict a pixel-level flow field between RSI and LIO structures. The estimated flow field was converted into geo-localization and loop-closure factors, which were integrated into the pose-graph for joint optimization. Additionally, a geo-spatial driven multi-layer orchard model was constructed based on RSI and LIO, which could extract phenotypes such as the height of fruit trees. We validated the framework in three representative areas of the apple orchard. On the multi-season dataset, the matching network achieved a mean endpoint error of 3.42 px (approximately 7 cm) and 94.5% of the samples within 5 px. The reconstruction pipeline achieved the root mean square error of the absolute pose error to 2.8–3.3 cm. The tree-height estimator built on the orchard model attained an R 2 of 0.95 and a mean absolute percentage error of 3.28%. The system is suitable for deployment on embedded devices for realtime operation. The framework provides centimeter-level global consistency in orchards, facilitates multi-season reconstruction and geo-spatial driven analysis, and serves as an AI-driven digital solution for next-generation orchard management. • Transformer-based cross-view fusion enables robust geo-localization and high-fidelity orchard-scale 3D modeling. • Pose-graph optimization with geo-localization and loop-closure factors sustains centimeter-level global consistency across seasons. • A GIS-driven multi-layer orchard model supports tree-scale phenotyping. • Embedded device deployment validates the performance and operational feasibility of the framework.