AgriLiRa4D: A Multi-Sensor UAV Dataset for Robust SLAM in Challenging Agricultural Fields

Flat Farmland

Hilly Farmland

Terraced Farmland

LiDAR

4D Radar

Abstract

Multi-sensor Simultaneous Localization and Mapping (SLAM) is essential for Unmanned Aerial Vehicles (UAVs) performing agricultural tasks such as spraying, surveying, and inspection. However, real-world, multi-modal agricultural UAV datasets that enable research on robust operation remain scarce. To address this gap, we present AgriLiRa4D, a multi-modal UAV dataset designed for challenging outdoor agricultural environments. AgriLiRa4D spans three representative farmland types—flat, hilly, and terraced—and includes both boundary and coverage operation modes, resulting in six flight sequence groups. The dataset provides high-accuracy ground-truth trajectories from a Fiber Optic Inertial Navigation System with Real-Time Kinematic capability (FINS_RTK), along with synchronized measurements from a 3D LiDAR, a 4D Radar, and an Inertial Measurement Unit (IMU), accompanied by complete intrinsic and extrinsic calibrations. Leveraging its comprehensive sensor suite and diverse real-world scenarios, AgriLiRa4D supports diverse SLAM and localization studies and enables rigorous robustness evaluation against low-texture crops, repetitive patterns, dynamic vegetation, and other challenges of real agricultural environments. To further demonstrate its utility, we benchmark four state-of-the-art multi-sensor SLAM algorithms across different sensor combinations, highlighting the difficulty of the proposed sequences and the necessity of multi-modal approaches for reliable UAV localization. By filling a critical gap in agricultural SLAM datasets, AgriLiRa4D provides a valuable benchmark for the research community and contributes to advancing autonomous navigation technologies for agricultural UAVs.

Dataset

The 33 sequences of our dataset were collected across three representative farmland terrains—flat plains, hilly regions, and mountainous terraces—located in Nanjing, China. The dataset is organized into six sequence groups based on terrain type and scanning mode (boundary or coverage), namely NJFlatB, NJFlatC, NJHillB, NJHillC, NJTerrB, and NJTerrC.
For all sequences except NJTerrB and NJTerrC, the UAV flew at a constant altitude with respect to the take-off point. In contrast, for the mountainous-terrain sequences NJTerrB and NJTerrC, the UAV maintained a fixed height Above Ground Level (AGL) to ensure flight safety and stable sensor coverage over rapidly varying elevation. Multiple combinations of flight altitudes and speeds were employed to introduce different levels of SLAM difficulty. Each sequence additionally begins with a short stationary or hovering segment to facilitate IMU initialization.

Flat Farmland

Hilly Farmland

Terraced Farmland

Scene	Sequence	Scanning Mode	Altitude (m)	Speed (m/s)	Path Length (m)
Flat Farmland	NJFlatB01	boundary	5	3	434.77
	NJFlatB02	boundary	5	8	464.21
	NJFlatB03	boundary	10	3	456.32
	NJFlatB04	boundary	10	8	462.18
	NJFlatB05	boundary	15	3	465.89
	NJFlatB06	boundary	15	8	454.21
	NJFlatC01	coverage	5	8	805.65
	NJFlatC02	coverage	10	3	801.17
	NJFlatC03	coverage	10	8	798.96
	NJFlatC04	coverage	15	3	822.23
Hilly Farmland	NJHillB01	boundary	8	3	490.61
	NJHillB02	boundary	8	8	493.07
	NJHillB03	boundary	13	3	480.98
	NJHillB04	boundary	13	8	484.60
	NJHillB05	boundary	18	3	483.84
	NJHillB06	boundary	18	8	488.41
	NJHillC01	coverage	8	3	776.47
	NJHillC02	coverage	8	8	783.31
	NJHillC03	coverage	13	3	761.55
	NJHillC04	coverage	13	8	768.14
	NJHillC05	coverage	18	3	756.07
	NJHillC06	coverage	18	8	769.94
Terraced Farmland	NJTerrB01	boundary	3	3	204.91
	NJTerrB02	boundary	6	3	207.21
	NJTerrB03	boundary	6	6	209.71
	NJTerrB04	boundary	9	3	211.95
	NJTerrB05	boundary	9	6	215.72
	NJTerrC01	coverage	3	3	311.23
	NJTerrC02	coverage	3	6	307.53
	NJTerrC03	coverage	6	3	311.24
	NJTerrC04	coverage	6	6	300.84
	NJTerrC05	coverage	9	3	313.64
	NJTerrC06	coverage	9	6	317.48

Calibration

Accurate LiDAR–Radar extrinsic calibration is crucial for multi-sensor fusion and consistent cross-modal point cloud alignment. We formulate this calibration as a 3D–3D registration problem, aligning the coordinate frames of both sensors within a unified reference. The CAD-derived translation and rotation (see Figure above) serve as initial priors, which are subsequently refined through a manual calibration procedure using multiple corner reflectors placed in the environment. This refinement significantly improves alignment precision compared to the raw CAD configuration, providing a reliable geometric basis for downstream multi-sensor SLAM.
The figure below is the visualization of the LiDAR and 4D Radar point clouds used to assess the extrinsic calibration. Height-colored LiDAR points and white 4D Radar points are visualized in a common frame following extrinsic alignment, with side and top-down views illustrating the spatial consistency across scenarios

Ground

Hovering

Scene

Side View

Top-down View

Mapping Results

Scene	LiDAR	4D Radar
Flat Farmland
Hilly Farmland
Terraced Farmland

BibTeX


@misc{zhan2025agrilira4dmultisensoruavdataset,
      title={AgriLiRa4D: A Multi-Sensor UAV Dataset for Robust SLAM in Challenging Agricultural Fields}, 
      author={Zhihao Zhan and Yuhang Ming and Shaobin Li and Jie Yuan},
      year={2025},
      eprint={2512.01753},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2512.01753}, 
}