Integrating UAV-borne LiDAR and deep learning for large-scale detection of productive macaw palms (Acrocomia aculeata)

Pronounced heterogeneity in fruit production among macaw palms (Acrocomia aculeata) poses significant challenges for optimizing management and scaling commercial plantations in Brazil. This study integrates high-density UAV-borne LiDAR with deep learning to enable automated, large-scale assessment of fruiting status in macaw palm. LiDAR data with point cloud densities ranging from 653 to 1,625 points m⁻² were collected across four commercial stands in southeastern Brazil. Field surveys georeferenced fruiting and non-fruiting palms, enabling segmentation of individual point clouds that were converted into two-dimensional height-colored images for model training and validation. Three convolutional neural network (CNN) architectures (ResNet-18, VGG-16, and DenseNet-121) were evaluated, with ResNet-18 achieving the highest classification accuracy (75%) and most stable training convergence. YOLOv8 was applied to canopy height models for automated palm detection and integrated with ResNet-18 to generate plantation-scale maps of fruiting status. YOLOv8 achieved a 92.5% F1-score, with reduced performance for shorter palms (<4.8 m). Classification performance was strongly dependent on point cloud density: high-density data (>1,173 points m⁻²) showed higher accuracy for fruiting palms but increased misclassification of non-fruiting palms, while low-density data (653 points m⁻²) exhibited the opposite bias. The complete workflow achieved 64.6% overall accuracy. These results demonstrate the potential of UAV-borne LiDAR and deep learning for precision management of macaw palm plantations and highlight the need to standardize LiDAR acquisition parameters, particularly point cloud density, before operational deployment.