From Orthomosaics to Raw UAV Imagery: Enhancing Palm Detection and Crown-Center Localization

1. Problem Statement

Accurate mapping of individual trees, specifically palms (Arecaceae), is critical for ecological monitoring and forest management in tropical ecosystems. While Unmanned Aerial Vehicle (UAV) imagery is widely used, current workflows rely heavily on orthomosaics (stitched images). This approach presents several limitations:

• Artifacts and Resolution Loss: Stitching processes introduce artifacts and often reduce resolution.
• Preprocessing Bottlenecks: Extensive pre- and post-processing is required, hindering real-time or edge deployment.
• Localization Errors: In dense forests, overlapping canopies cause systematic offsets between the center of a predicted bounding box and the actual crown center of the tree. This reduces the accuracy of spatial distribution modeling and downstream ecological analyses.

The study addresses the need to move from stitched orthomosaics to raw UAV imagery to improve visual fidelity and computational efficiency, while also investigating whether explicit crown-center annotations can overcome the limitations of standard bounding-box centroids.

2. Methodology

Data Collection and Datasets

The study utilized data from the Fundación para la Conservación de los Andes Tropicales (FCAT) reserve in Ecuador.

• Hardware: DJI Phantom 4 RTK drone with a 1-inch CMOS sensor, flying at 90m altitude.
• Two Datasets:
  1. Orthomosaic Patches: 1,500 images (800×800 px) cropped from large-scale orthomosaics. Contains 8,842 bounding boxes.
  2. Raw Patches: 880 images (912×912 px) from raw UAV captures. Contains 5,850 bounding boxes and 5,430 explicit crown-center annotations.
  • Annotation Protocol: Crown centers represent the geometric center of the visible crown, distinct from bounding box centers, which is crucial for handling occlusion and truncation.

Experimental Setup

The researchers evaluated two primary Research Questions (RQs) using state-of-the-art deep learning models:

• RQ1 (Domain Transfer): How does performance vary between training/testing on orthomosaics vs. raw imagery (within-domain and cross-domain)?
• RQ2 (Localization Precision): Do explicit crown-center annotations improve localization accuracy compared to deriving centers from bounding boxes?

Models Used:

• Detection: RT-DETR (v1, v2) and YOLO variants (v8 through v12).
• Keypoint/Center Localization: YOLO Pose models (v8, v11, v12) trained to predict crown centers directly.
• Metrics: COCO-style metrics (mAP, mAR) for detection; Object Keypoint Similarity (OKS) based mAP, Precision, Recall, and F1-score for localization.

3. Key Contributions

1. New Dataset & Protocol: Introduction of a raw UAV dataset with dual annotations (bounding boxes + crown centers), enabling fine-grained evaluation of localization accuracy in dense forests.
2. Cross-Domain Analysis: A comprehensive study comparing orthomosaic and raw imagery domains, revealing trade-offs between deployment accuracy and cross-domain generalization.
3. Pipeline Extension: An extension of the existing PRISM pipeline to integrate crown-center localization, providing a more ecologically meaningful reference for spatial analysis.

4. Key Results

Detection Performance (RQ1)

• Raw Imagery Superiority for Deployment: Models trained and tested on Raw Patches achieved the highest performance. For instance, YOLOv11 reached 0.627 mAP and 0.707 mAR on raw data.
• Orthomosaic Limitations: Models trained on orthomosaics performed significantly worse when tested on raw images (e.g., YOLOv11 dropped to ~0.526 mAP), indicating that raw imagery provides higher-quality training signals for field deployment.
• Cross-Domain Generalization: Conversely, models trained on Orthomosaics transferred better to raw images than the reverse. Orthomosaic-trained models achieved 0.545 mAP on raw test data, whereas raw-trained models dropped to 0.380 mAP on orthomosaics. This suggests orthomosaics offer better robustness for generalization due to their diverse (albeit noisy) annotations.

Localization Accuracy (RQ2)

• Crown-Center vs. Box Centroid: Explicitly annotating and predicting crown centers yielded massive improvements over using bounding box centroids.
  • mAP Improvement: Increased from ~0.32–0.34 (Box Center) to >0.75 (Crown Center).
  • F1-Score Improvement: Increased from ~0.55–0.60 to >0.80.
• Visual Evidence: Qualitative analysis (Figure 2) showed that in dense/occluded canopies, bounding box centroids were consistently displaced, while predicted crown centers aligned closely with ground truth.

5. Significance and Future Work

• Ecological Impact: The shift to raw imagery and crown-center localization enables more precise tree positioning, which is essential for accurate spatial distribution modeling and biodiversity monitoring.
• Operational Efficiency: By bypassing the stitching process, this approach supports real-time or near-real-time monitoring on edge devices, a capability traditional orthomosaic pipelines lack.
• Future Directions: The authors propose combining mixed-domain training with raw image inference to leverage the generalization of orthomosaics and the fidelity of raw data. Future work will focus on lightweight models for onboard UAV inference and extending these outputs to broader palm spatial modeling.

Conclusion: The study demonstrates that while orthomosaics are useful for robust cross-domain training, raw UAV imagery is superior for actual field deployment. Furthermore, moving beyond bounding boxes to explicit crown-center annotations is critical for resolving localization errors in complex, dense forest environments.