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

Damini Raniga, Amarasingam Narmilan, Juan Sandino, Ashray Doshi, Johan Barthélemy, Krystal Randall, Sharon A. Robinson, Felipe Gonzalez, Barbara BollardSensors 2024 · DOI

TL;DR

This paper introduces an innovative workflow for monitoring the health of fragile Antarctic vegetation using drone-based multispectral imagery and AI. By comparing two machine learning models, XGBoost and U-Net, the study demonstrates the potential of these technologies for accurate, non-invasive ecological assessment in extreme environments.

What Problem Does This Paper Address?

Antarctica's unique ecosystems, particularly its moss and lichen communities, are highly vulnerable to climate change and ozone depletion. Traditional monitoring methods are labor-intensive, costly, and can disturb these delicate environments. While Unmanned Aerial Vehicles (UAVs) and Artificial Intelligence (AI) offer promising non-invasive solutions, their application in Antarctic vegetation monitoring, especially using multispectral imagery and deep learning, remains limited. This research aims to bridge this gap by developing and validating a robust methodology for classifying moss and lichen health status, crucial for conservation efforts.

Methods

The study proposes a novel workflow for vegetation monitoring, focusing on two supervised machine learning (ML) classifiers: Extreme Gradient Boosting (XGBoost) and U-Net.

  • XGBoost: This gradient boosting algorithm was used for initial predictions. Its pipeline involved data collection, spectral index calculation, feature extraction, model configuration, training, feature importance analysis, and deployment.
  • U-Net: A convolutional neural network (CNN) architecture known for semantic segmentation, U-Net was employed in two ways:
  • Method 1: Utilized the original labeled data, similar to XGBoost.
  • Method 2: Incorporated XGBoost's predictions as additional input features for the U-Net model. This "two-stage ensemble" approach aimed to leverage the strengths of both models.

The process involved extensive data pre-processing, including orthomosaic generation, georeferencing, and meticulous data labeling based on ground truth points. Spectral indices were calculated and used as features for the models.

Datasets

The study utilized data collected from Antarctic Specially Protected Area (ASPA) 135, near Casey Station, East Antarctica, between January and February 2023.

  • Aerial Data: Acquired using a custom-built BMR3.9RTK UAV equipped with a MicaSense Altum multispectral camera and a Sony Alpha 5100 high-resolution RGB camera. A DJI Mini 3 Pro was also used for ultra-high resolution RGB data.
  • Ground Truth Data: 68 ground truth points were collected on foot using GNSS RTK for accurate georeferencing.
  • Classes: The models were trained to classify five classes: Healthy Moss, Stressed Moss, Moribund Moss, Lichen, and Non-Vegetation.

Key Insights and Ideas

  • The Power of Ensemble Learning: Method 2, which combined XGBoost predictions with U-Net, significantly outperformed Method 1 (U-Net alone), highlighting the synergistic benefits of a two-stage approach for complex segmentation tasks.
  • Data Quality is Paramount: The study underscores the critical role of training data quantity and quality. Limited ground truth data for certain classes (like healthy moss) led to lower recall and F1-scores, demonstrating the challenge of achieving high performance with sparse datasets.
  • Spectral Indices Still Matter: While deep learning models can learn complex features, spectral indices derived from multispectral imagery remain valuable inputs, particularly for identifying vegetation health indicators. The red band, MSAVI, and kurtosis were identified as top features for XGBoost.
  • UAVs as a Game-Changer for Antarctic Research: The research reinforces the utility of UAVs for non-invasive, high-resolution monitoring in extreme environments, offering a more efficient and sustainable alternative to traditional methods.

Key Results

  • XGBoost Performance: Achieved robust performance with precision, recall, and F1-scores exceeding 85% overall. Individual class performance varied, with Non-Vegetation achieving perfect scores, while Healthy Moss had lower recall due to limited samples.
  • U-Net Performance (Method 1 vs. Method 2):
    • Method 1: Achieved decent results, but with limitations, particularly for Healthy Moss.
  • Method 2: Showed substantial improvements across key metrics. For instance, precision for Healthy Moss increased from 74% to 94%, and recall for Stressed Moss improved from 69% to 86% compared to Method 1. Overall, Method 2 demonstrated enhanced precision, recall, and F1-scores across all classes.
  • Overall Accuracy: The U-Net model with Method 2 achieved an overall accuracy of approximately 93.23% in K-fold cross-validation.
  • Class Distribution: The study provided percentage distributions for each class, showing variations between XGBoost and U-Net predictions, offering insights into their differing segmentation capabilities.

Limitations and Future Work

  • Resolution Limitations: The resolution of the Altum multispectral camera limited the identification of finer moss details. Higher resolution data, like that from the DJI Mini 3 Pro, would improve segmentation accuracy, especially for intertwined moss classes.
  • Training Data Scarcity: The limited availability of ground truth data was a significant constraint, particularly for deep learning models like U-Net, hindering their ability to discern complex spectral patterns.
  • Future Work:
    • Capture higher-resolution RGB and multispectral data for more accurate labeling and segmentation.
  • Investigate the impact of specific spectral indices on different vegetation classes.
  • Explore other ML algorithms for improved segmentation.
  • Further refine ML techniques for polar environments, potentially by leveraging classical ML (like XGBoost) for initial predictions to augment DL model training.
  • Enhance CNN classifier performance and visually verify polygon labeling with higher-resolution RGB data.

Summary updated

BibTeX
@article{raniga2024monitoring,
  author = {Damini Raniga and Amarasingam Narmilan and Juan Sandino and Ashray Doshi and Johan Barthélemy and Krystal Randall and Sharon A. Robinson and Felipe Gonzalez and Barbara Bollard},
  title = {{Monitoring of Antarctica's Fragile Vegetation Using Drone-Based Remote Sensing, Multispectral Imagery and AI}},
  journal = {Sensors},
  year = {2024},
  doi = {10.3390/s24041063},
}
Figure