Individual tree crown delineation and tree species classification with hyperspectral and LiDAR data

Area of Science:

• Ecology
• Remote Sensing
• Data Science

Background:

• Ecological applications increasingly rely on remote sensing data.
• A data science challenge was established to enhance the utility of remote sensing in ecology.
• The challenge focused on individual tree crown (ITC) delineation, field-to-ITC alignment, and tree species classification.

Purpose of the Study:

• To present the methods and results of team Fondazione Edmund Mach (FEM) in a data science challenge.
• To improve the use of remote sensing data for ecological applications, specifically focusing on individual trees.

Main Methods:

• Individual Tree Crown (ITC) delineation utilized a region growing method on hyperspectral near-infrared bands, optimized using the Jaccard score.
• Alignment between field data and delineated ITCs employed Euclidean distance based on position, height, and crown radius.
• Tree species classification involved a support vector machine classifier, using selected hyperspectral bands and canopy height model, with band selection via sequential forward floating selection and Jeffries Matusita distance.

Main Results:

• Team FEM achieved first place in ITC delineation (Task 1) and alignment (Task 2), and second place in classification (Task 3).
• The Jaccard score for delineated crowns was 0.3402, demonstrating effective delineation of both small and large crowns.
• Alignment was accurate for all test samples, and classification achieved 88.1% overall accuracy, 75.7% kappa accuracy, and 61.5% mean class accuracy, though biased towards dominant species.

Conclusions:

• The proposed methods demonstrated high performance in individual tree crown delineation and alignment from remote sensing data.
• The support vector machine approach provided robust tree species classification, despite some bias towards prevalent species.
• Team FEM's success highlights the potential of advanced data science techniques for ecological remote sensing applications.