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

AI summary

Overview

Research area: Robotics — multi-sensor Simultaneous Localization and Mapping (SLAM), aerial robotics, and agricultural autonomy, specifically a dataset and benchmark paper.

Technical level: Intermediate. The paper is a dataset/benchmark contribution that is readable without deep SLAM mathematics, but it assumes familiarity with sensor fusion terminology (LiDAR-inertial, radar-inertial, odometry, ATE).

Scope (one sentence): The paper introduces AgriLiRa4D, a multi-modal UAV dataset with LiDAR, 4D Radar, IMU, and FINS_RTK ground truth collected over flat, hilly, and terraced farmland, and benchmarks four existing multi-sensor SLAM algorithms on it.

What This Paper Is About

Agricultural UAVs need reliable localization for tasks such as spraying, surveying, and inspection, but GNSS degrades under vegetation occlusion and multipath, and pure visual SLAM fails on repetitive, texture-poor crops with dynamic illumination and vegetation motion. Existing multimodal SLAM datasets mostly cover indoor, urban, or semi-natural scenes, while existing agricultural datasets are largely ground-based, orchard-based, or lack the dense-foliage, low-altitude conditions that UAVs actually face. The goal of this paper is to fill that gap with a real-world, synchronized, multi-sensor agricultural UAV dataset plus a robustness benchmark, so that multi-modal localization methods can be evaluated under genuinely difficult farm conditions.

Key Contributions

1. AgriLiRa4D dataset. A large-scale agricultural UAV dataset covering three terrain types (flat, hilly, terraced) and two operation modes (boundary and coverage), with synchronized 3D LiDAR, 4D Radar, and IMU measurements plus centimeter-level position and high-precision orientation ground truth from a FINS_RTK system, along with complete intrinsic and extrinsic calibrations.
2. SLAM benchmark. A comprehensive evaluation of representative multi-sensor fusion SLAM algorithms spanning LiDAR–Inertial (FAST-LIO2, Faster-LIO), Radar–Inertial (EKF-RIO), and Radar–LiDAR–Inertial (GaRLIO) configurations, reporting quantitative localization accuracy and robustness.
3. Multi-modal fusion analysis. An analysis of how sensor modality, terrain geometry, vegetation, altitude, and speed affect pose estimation, including how Radar–LiDAR–IMU integration changes consistency and robustness across crop types, terrain slopes, and flight regimes.

Main Findings

• LiDAR–inertial methods are the most stable. FAST-LIO2 and Faster-LIO achieve the lowest ATE in the majority of sequences and operate reliably across all farmland types, showing only moderate degradation at higher altitudes or faster flight speeds.
• Radar–inertial odometry is the least stable. EKF-RIO fails on nearly all flat farmland sequences and drifts heavily even where it completes runs in hilly and terraced farmland. The paper attributes this to sparse and noisy 4D Radar returns over flat crops and grass, sensitivity to downwash-induced vegetation motion, and unreliable Doppler cues when spatial structure is weak.
• Radar–LiDAR–inertial performance is terrain-dependent. GaRLIO frequently fails in hilly farmland, is inconsistent in flat terrain, but is most stable and competitive in terraced farmland, where it occasionally matches or slightly outperforms the LIO baselines in rotation accuracy. The paper attributes the mixed behavior to a design/domain mismatch, since GaRLIO was developed and tested primarily on Unmanned Ground Vehicle platforms with near-ground viewpoints, balanced LiDAR–Radar overlap, and slower motion.
• Vertical drift dominates. Positional error accumulates predominantly along the Z (altitude) axis (illustrated for sequence NJHillC04), because airborne LiDAR observes the environment from near-horizontal viewpoints and 4D Radar has coarser elevation resolution, leaving altitude estimation more dependent on IMU integration.
• Terraced farmland is easiest; flat farmland is hardest. Terraced slopes (around 45°) with tea grown along contour lines provide strong, well-structured geometric contours and informative Doppler patterns, allowing even weaker modalities to complete trajectories. Flat sorghum fields offer minimal 3D geometric variation, sparse LiDAR features, and weak Radar returns.
• The dataset spans a wide difficulty spectrum. The benchmark results across sensing modalities and farmland types show AgriLiRa4D is discriminative, exposing failures that are not visible in structured or semi-structured environments.

Methodology in Plain English

The authors built a sensing payload on a TopXGun FP300E agricultural UAV: a RoboSense Airy 3D LiDAR with an integrated IMU, a Mindcruise A1 4D Radar mounted beneath the LiDAR to maximize field-of-view overlap, and a TJ-FINS70D FINS_RTK module for reference trajectories. Data are logged on an onboard ARM computer based on the RK3588 processor running ROS, with sensors connected over Gigabit Ethernet and synchronized using IEEE 1588 Precision Time Protocol.

They collected flights in Nanjing, China across three real farm sites: flat sorghum fields, gently sloped hilly farmland (about 15°) with sorghum and grass, and steep terraced tea farmland (about 45°). They varied flight altitude and speed to create different difficulty levels, and ran both "boundary" and "coverage" scanning patterns, producing six sequence groups. All sequences except the terraced ones were flown at a constant altitude relative to the take-off point; the terraced sequences held a fixed height above ground level. Each sequence starts with a short stationary or hovering segment to let the IMU initialize.

Sensor extrinsics were taken from the UAV CAD design files and then refined through a manual calibration procedure using corner reflectors, evaluated by visual inspection of aligned LiDAR and Radar point clouds in static and hovering scenarios. Ground truth from FINS_RTK is provided in three forms: pose in the ENU frame relative to take-off, pose in the FLU body frame, and geodetic latitude/longitude/altitude, with orientation following the FRD convention.

For evaluation, they ran four open-source algorithms (FAST-LIO2, Faster-LIO, EKF-RIO, GaRLIO) and measured the RMSE of Absolute Trajectory Error in translation (meters) and rotation (degrees) against FINS_RTK ground truth using the evo toolkit. A run counts as a failure if translational ATE relative to path length exceeds 15%, or if rotational ATE exceeds 45°. Other candidate algorithms (LIO-SAM, LiLi-OM, DRIO, 4D-IRIOM, DR-LRIO, AF-RLIO) were excluded because of missing official open-source implementations, reliance on modalities not present in the dataset (visual, depth, or GNSS), or hardware/engineering requirements that hinder reproducible large-scale evaluation.

Why This Matters

Impact on research. Existing multimodal UAV benchmarks are dominated by structured, indoor, urban, or semi-natural scenes with stable features, and existing agricultural datasets are largely ground-platform based, orchard-based, or lack LiDAR and full 6-DoF ground truth. AgriLiRa4D supplies a real farmland benchmark with synchronized LiDAR, 4D Radar, and IMU data plus FINS_RTK references, allowing robustness claims about multi-modal SLAM to be tested under dense foliage, repetitive textures, vegetation motion, and illumination variation. The results also surface failure modes, such as Z-axis drift and Radar sparsity over flat crops, that motivate new algorithm designs rather than incremental tuning.

Real-world applications.

• Autonomous pesticide spraying and crop-dusting, where GNSS alone is insufficient near canopies and downwash perturbs the scene.
• Crop monitoring and phenotyping surveys over large fields, including flat and hilly terrain.
• Inspection and large-scale farmland surveying, including terraced terrain where safe flight requires consistent height above ground level.
• Precision-farming navigation where continuous pose estimation is needed when GNSS becomes unreliable.

Industry relevance. The author list includes TopXGun Robotics, an agricultural drone manufacturer, and the platform and sensors come from commercial agricultural UAV and sensor vendors, indicating the dataset targets practical deployment. Reliable GNSS-independent localization is a prerequisite for certifying and scaling autonomous agricultural flight, so a public benchmark with ground truth at 2 cm + 1 ppm position accuracy and 0.01°/0.05° orientation accuracy (1σ, roll/pitch and yaw respectively) gives vendors and researchers a common yardstick. The dataset is downloadable at https://zhan994.github.io/AgriLiRa4D.

Future Directions

• Broaden crop and season coverage. The paper itself notes that most existing agricultural datasets involve a single crop species, a single season, or a limited geographic region, with few providing long-term, multi-season, or multi-terrain coverage; extending AgriLiRa4D along those axes is a natural next step.
• Add missing modalities and evaluate more algorithms. Visual, depth, and GNSS data are absent, which is why methods such as LIO-SAM, LiLi-OM, DRIO, 4D-IRIOM, DR-LRIO, and AF-RLIO were not benchmarked; adding those sensors would widen the comparable method pool.
• Improve altitude and Z-axis estimation. Because vertical drift dominates and depends on IMU integration when geometry or Radar elevation cues are weak, methods that better constrain elevation with airborne LiDAR or 4D Radar are an open problem.
• Adapt ground-vehicle and indoor-designed pipelines to aerial farmland operation. GaRLIO's mixed results and EKF-RIO's indoor validation point to a need for domain adaptation or retraining of Radar-centric methods for high-altitude, fast-motion, vegetation-perturbed flight.

Target Audience

SLAM and sensor-fusion researchers who need a challenging outdoor benchmark; UAV autonomy and navigation engineers working on GNSS-denied or GNSS-degraded operation; agricultural robotics practitioners and drone manufacturers evaluating localization robustness over real fields; and graduate students or newcomers to multi-modal odometry who want a well-documented dataset with ground truth, calibration, and baseline results to build on.

Authors’ 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. The dataset can be downloaded from: https://zhan994.github.io/AgriLiRa4D.