High-resolution UAV imagery-based instance segmentation for automatic counting of rubber trees across phenological...

Area of Science:

• Agricultural remote sensing
• Computer vision
• Deep learning for ecological monitoring

Background:

• Accurate rubber tree counting is vital for yield estimation and sustainable agriculture.
• Deep learning models require reliable canopy segmentation for precise analysis.
• Existing methods struggle with overlapping crowns, background vegetation, and seasonal variations in rubber tree plantations.

Purpose of the Study:

• To develop a high-precision rubber tree canopy segmentation model (CSAF) for UAV imagery.
• To address challenges in segmentation accuracy caused by complex field conditions.
• To introduce a comprehensive dataset (RT-Set) for benchmarking rubber tree segmentation and counting.

Main Methods:

• CSAF model incorporating Boundary Continuity Modelling Module (BCMM), Physical Morphology Constraint Module (PMCM), and Cross-Temporal Learning Adaptively Module (CTLAM).
• BCMM utilizes Fourier transform and state-space modeling for boundary refinement.
• PMCM employs Poisson diffusion prior for morphological constraints, and CTLAM uses state-feedback for cross-temporal adaptation.
• Development of the RT-Set dataset with 5281 high-resolution images covering the full growth cycle of rubber trees.

Main Results:

• CSAF achieved an AP50 of 76.63% on the RT-Set dataset, outperforming existing methods.
• The model demonstrated competitive performance against mainstream approaches like DetecTree2 and Cascade Mask R-CNN on public datasets.
• CSAF exhibited high accuracy and stable performance in rubber tree counting experiments.

Conclusions:

• CSAF offers a robust solution for high-precision rubber tree canopy segmentation and counting.
• The proposed model effectively handles complex field conditions and inter-seasonal variations.
• The RT-Set dataset provides a valuable benchmark for advancing research in rubber tree monitoring.