Weakly supervised forest canopy extraction and multi-dimensional joint canopy entropy for quantifying canopy...

Area of Science:

• Forestry and Remote Sensing
• Ecological Informatics
• Geospatial Analysis

Background:

• Accurate forest Canopy Structural Complexity (CSC) assessment is vital for ecological monitoring.
• Existing canopy extraction methods require extensive labeled LiDAR data, limiting large-scale UAV surveys.
• There's a need for efficient canopy extraction and CSC quantification methods for unlabeled forest point clouds.

Purpose of the Study:

• To develop a weakly supervised canopy extraction strategy for UAV LiDAR data using limited labeled plots.
• To introduce a novel multi-dimensional joint canopy entropy index for comprehensive CSC quantification.
• To evaluate the performance and generalization of the proposed methods on large-scale, unlabeled forest datasets.

Main Methods:

• Proposed a weakly supervised learning approach with pseudo-label generation and self-supervised constraints for canopy extraction.
• Developed a multi-dimensional joint canopy entropy index integrating global (projection-based) and local (graph-based) canopy features.
• Conducted experiments using limited labeled plots for training/evaluation and numerous unlabeled plots for generalization analysis.

Main Results:

• The weakly supervised canopy extraction achieved 85.37% mIoU and 92.64% OA on labeled plots, outperforming fully supervised methods.
• The method demonstrated robust performance in complex forest scenes with varied topography and canopy structures, minimizing false positives.
• The multi-dimensional joint canopy entropy index effectively quantified CSC in large-scale unlabeled data, achieving a normalized score of 0.76, surpassing existing indices.

Conclusions:

• Weakly supervised learning is a viable and effective strategy for forest canopy extraction from UAV LiDAR data, reducing reliance on large labeled datasets.
• The proposed multi-dimensional joint canopy entropy index provides a more comprehensive and reliable measure of forest CSC.
• These advancements enable accurate ecological assessments and spatiotemporal canopy dynamics analysis in large-scale, unlabeled forest environments.