Detecting Pest-Infested Forest Damage through Multispectral Satellite Imagery and Improved UNet+

Area of Science:

• Forestry
• Remote Sensing
• Computer Vision

Background:

• Plant pests pose significant threats to agricultural and forest ecosystems.
• Accurate monitoring of pest-induced forest damage is vital for effective management strategies.
• Previous deep learning approaches for pest damage detection primarily used RGB imagery, neglecting valuable multispectral and vegetation index data.

Purpose of the Study:

• To improve forest pest infestation area segmentation by integrating multispectral, vegetation index, and RGB information into deep learning models.
• To develop and evaluate a novel image segmentation method utilizing UNet++ with an attention mechanism for detecting bark beetle and aspen leaf miner damage.
• To leverage Sentinel-2 satellite imagery for enhanced forest health monitoring.

Main Methods:

• A new deep learning model based on UNet++ with an scSE attention mechanism was proposed for image segmentation.
• ResNeSt101 was employed as the feature extraction backbone within the UNet++ architecture.
• A dataset was created using Sentinel-2 imagery and forest health damage data from British Columbia, Canada, incorporating 11 original bands and 13 vegetation indices.

Main Results:

• The integration of vegetation indices and multispectral data significantly enhanced the segmentation performance.
• The proposed UNet++ model with attention mechanism achieved an overall accuracy of 85.11%.
• The method demonstrated superior segmentation quality and more accurate quantitative indices compared to existing state-of-the-art techniques.

Conclusions:

• Multispectral imagery and vegetation indices are crucial for improving the precision of forest pest damage detection.
• The developed deep learning approach offers a more accurate and effective solution for monitoring forest pest infestations.
• This research provides a valuable tool for forest ecosystem management and conservation efforts.