Monitoring of Antarctica's Fragile Vegetation Using Drone-Based Remote Sensing, Multispectral Imagery and AI

Area of Science:

• Remote Sensing and Geospatial Analysis
• Artificial Intelligence in Ecology
• Antarctic Ecosystem Monitoring

Background:

• Antarctic vegetation (moss, lichen) is vulnerable to climate change and ozone depletion.
• Non-invasive monitoring methods are essential for assessing vegetation health.
• Limited use of deep learning (DL) and multispectral imagery for Antarctic vegetation analysis via unmanned aerial vehicles (UAVs).

Purpose of the Study:

• To highlight the potential of DL for Antarctic vegetation datasets.
• To introduce and compare a novel workflow using Extreme Gradient Boosting (XGBoost) and U-Net classifiers.
• To detect and map Antarctic moss and lichen health status.

Main Methods:

• Utilized high-resolution multispectral imagery from UAVs in Antarctic Specially Protected Area (ASPA) 135.
• Developed and compared two supervised machine learning (ML) models: XGBoost and U-Net.
• Trained models on five classes: Healthy Moss, Stressed Moss, Moribund Moss, Lichen, and Non-vegetated; U-Net employed two methods, one with original data and another incorporating XGBoost predictions.

Main Results:

• XGBoost achieved over 85% in precision, recall, and F1-score.
• The U-Net workflow incorporating XGBoost predictions (Method 2) significantly improved classification accuracy compared to using original data alone (Method 1).
• Notable U-Net Method 2 improvements include precision for Healthy Moss (94% vs. 74%) and recall for Stressed Moss (86% vs. 69%).

Conclusions:

• Deep learning and machine learning models show significant promise for non-invasive monitoring of Antarctic vegetation.
• The proposed workflow enhances the accuracy of vegetation classification in sensitive polar ecosystems.
• Findings support the use of UAVs, multispectral imagery, and AI for effective remote sensing in polar regions.