Multimodal collaborative UAV framework for single rubber tree parsing

Area of Science:

• Forestry Science
• Computer Vision
• Remote Sensing

Background:

• Accurate individual-tree segmentation is crucial for rubber plantation management.
• Challenges include overlapping crowns, varying point density, and diverse tree morphology.
• Existing methods struggle with the complexities of densely planted stands.

Purpose of the Study:

• To introduce MDA-SegNet, a multimodal point-cloud instance segmentation framework.
• To enhance individual-tree segmentation in densely planted rubber plantations.
• To improve resource inventory and agro-forestry management.

Main Methods:

• MDA-SegNet integrates orthophoto and LiDAR data using a Multimodal Deformable Encoding (MDE) module.
• A Z-Order Selective Mamba (ZOS-Mamba) module addresses vertical dependencies and uneven point density.
• An Adaptive Lemming Optimization Clustering (ALOC) module refines segmentation for overlapping canopies.

Main Results:

• MDA-SegNet demonstrated superior performance on a custom rubber tree dataset (RT-Set) and public forest datasets.
• Achieved an F-score of 87.32% and Recall of 82.90% on RT-Set, outperforming state-of-the-art models.
• Showcased strong robustness and cross-domain generalizability with higher F-score and mIoU.

Conclusions:

• MDA-SegNet provides reliable individual-tree segmentation for dense rubber plantations.
• The framework effectively handles challenges like crown overlap and morphological variations.
• Offers significant improvements for precision agriculture and forest management applications.