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

• Published in: Sensors
• Main Authors: Damini Raniga, Narmilan Amarasingam, Juan Sandino, Ashray Doshi, Johan Barthelemy, Krystal Randall, Sharon A. Robinson, Felipe Gonzalez, Barbara Bollard
• Format: Article in Journal/Newspaper
• Language: English
• Published: MDPI AG 2024
• Subjects: antarctic specially protected area (ASPA); machine learning; gradient boosting; convolutional neural network; unmanned aerial vehicle (UAV); lichen; Chemical technology; TP1-1185; Antarctic; Casey Station; East Antarctica; Antarc*; Antarctica
• Online Access: https://doi.org/10.3390/s24041063; https://doaj.org/article/a7bca5c7cf3f40ca9e85f353472030d0

Vegetation in East Antarctica, such as moss and lichen, vulnerable to the effects of climate change and ozone depletion, requires robust non-invasive methods to monitor its health condition. Despite the increasing use of unmanned aerial vehicles (UAVs) to acquire high-resolution data for vegetation analysis in Antarctic regions through artificial intelligence (AI) techniques, the use of multispectral imagery and deep learning (DL) is quite limited. This study addresses this gap with two pivotal contributions: (1) it underscores the potential of deep learning (DL) in a field with notably limited implementations for these datasets; and (2) it introduces an innovative workflow that compares the performance between two supervised machine learning (ML) classifiers: Extreme Gradient Boosting (XGBoost) and U-Net. The proposed workflow is validated by detecting and mapping moss and lichen using data collected in the highly biodiverse Antarctic Specially Protected Area (ASPA) 135, situated near Casey Station, between January and February 2023. The implemented ML models were trained against five classes: Healthy Moss, Stressed Moss, Moribund Moss, Lichen, and Non-vegetated. In the development of the U-Net model, two methods were applied: Method (1) which utilised the original labelled data as those used for XGBoost; and Method (2) which incorporated XGBoost predictions as additional input to that version of U-Net. Results indicate that XGBoost demonstrated robust performance, exceeding 85% in key metrics such as precision, recall, and F1-score. The workflow suggested enhanced accuracy in the classification outputs for U-Net, as Method 2 demonstrated a substantial increase in precision, recall and F1-score compared to Method 1, with notable improvements such as precision for Healthy Moss (Method 2: 94% vs. Method 1: 74%) and recall for Stressed Moss (Method 2: 86% vs. Method 1: 69%). These findings contribute to advancing non-invasive monitoring techniques for the delicate Antarctic ecosystems, showcasing the potential of ...