In precision agriculture, where data is used to automate and optimize agricultural tasks for maximum yield, large-scale mapping and digital modeling of the land of interest are essential. This digital backbone enables tasks, such as yield estimation, targeted spraying, and phenotyping, while also providing the foundation for autonomous navigation, perception, and robotic operation. However, in modern commercial orchards, with their large scales, high tree density, and structured layouts, achieving consistent mapping and accurate 3D modeling remains challenging.
Drone-based remote-sensing imagery (RSI), aided by global navigation satellite systems (GNSS), can generate accurate aerial maps. However, aerial RSI cannot reliably capture important details beneath dense tree canopies. On the other hand, ground-based robots using LiDAR-inertial odometry can produce high-fidelity 3D point clouds that capture both individual tree structures and orchard row layouts. Despite these advantages, weakened satellite signals under dense foliage cause their estimated trajectories to accumulate drift over long distances, reducing localization accuracy.
Previous studies have explored multi-sensor fusion techniques that utilize aerial maps to calibrate trajectories for ground robots and suppress long-term drift. Conventional matching techniques, however, are highly sensitive to differences between aerial imagery and LiDAR data. Their accuracy is further affected by seasonal appearance changes, inconsistent textures, and repetitive canopy patterns, limiting global consistency for practical applications.
To address these limitations, a research team led by Professor Kyeong-Hwan Lee from the Department of Convergence Biosystems Engineering at Chonnam National University in South Korea developed a novel cross-modal fusion framework, integrating low-altitude drone RSI with ground robot LiDAR-inertial measurement unit (IMU) odometry (LIO). "Our system utilizes deep learning-based cross-modal alignment to connect these two sources of information, identifying the same tree rows, canopy patterns, and open spaces in aerial images and in the robot's laser measurements," explains Prof. Lee. "This way, the aerial map can be used as a reliable geographic reference, placing the ground robot's detailed measurements accurately within the larger orchard map." Their study was made available online on March 26, 2026, and published in Volume 16, Issue 2 of the Artificial Intelligence in Agriculture journal on June 01, 2026.
The team first acquired RSI data using a drone platform, capturing high-resolution aerial images of an apple orchard in a pre-planned flight pattern. These images were processed to create a structured RSI-tiled database, and from this an RSI local-map was created for a small area within the orchard. In the same orchard, a ground robot equipped with a LiDAR-IMU system collected LiDAR point cloud data. This was then processed to create a structured 2D Bird's-Eye View (BEV) local-map for the same area as the RSI local-map, encoding essential 3D structural information, including canopy height variations, surface reflectance, and local structural density.
To align the two data sources, the 2D BEV was then matched with the corresponding RSI local-map using a cross-view LIO-RSI fusion matching network. This deep learning model, featuring dual-branch feature extraction, a transformer-based cross-attention module, and a multi-scale flow refinement module, utilized pixel-level structural cues to reconstruct and accurately align the LiDAR map with the aerial map, substantially reducing accumulated drift.
The aligned data were then incorporated into a pose-graph estimation framework to create a geographic information system-driven multi-layer orchard model. This digital model can be used to extract phenotypic traits, including tree height, as well as information about tree health.
In tests covering approximately 1.3 kilometers of orchard travel, the fusion framework achieved localization accuracy on the order of a few centimeters, demonstrated robustness to seasonal variations, and suppressed long-term drift more effectively than conventional approaches. Moreover, the system is suitable for deployment on embedded devices for real-time operation.
"By integrating what a robot sees on the ground with an aerial map, our system can help agricultural robots work reliably in orchards, supporting practical tasks such as crop inspection, targeted spraying, mowing, transportation, and harvesting," concludes Prof. Lee. "In the future, such information could lead to living digital models of farms and orchards, helping them adapt to seasons and environmental and societal pressures. Ultimately, this will help farmers produce food more efficiently and sustainably."
