An AI-ready remote sensing dataset for high-resolution forest disturbance mapping.

Area of Science:

• Forestry Science
• Remote Sensing
• Artificial Intelligence

Background:

• European forests face escalating threats from natural disturbances (insect outbreaks, pathogens, windthrow), exacerbated by extreme weather and salvage logging.
• Current monitoring methods using medium-resolution satellite imagery often miss high-spatial-detail events, and manual reporting is inefficient.

Purpose of the Study:

• To introduce a novel, high-resolution dataset for classifying forest disturbance types using deep learning.
• To provide a valuable resource for improving forest management and climate adaptation strategies.

Main Methods:

• Development of a dataset comprising ~17,500 image patches (500x500 pixels at 0.2m resolution) from German digital orthophotos.
• Inclusion of five spectral channels (RGB, near-infrared) and object height, with segmentation masks for disturbance classes.
• Application of a deep learning model for classification and ablation analysis to assess channel importance.

Main Results:

• The deep learning model achieved an overall accuracy of 88.2% in classifying forest disturbances.
• Near-infrared and object height channels were identified as the most informative for disturbance detection.
• The dataset facilitates high-resolution, automated monitoring of forest disturbances.

Conclusions:

• The presented dataset is a significant advancement for deep learning-based forest disturbance monitoring.
• High-resolution data, particularly near-infrared and object height, are crucial for accurate classification of threats like bark beetle damage and windthrow.
• This resource can enhance forest management practices and climate change adaptation efforts.