Use of Very High Spatial Resolution Commercial Satellite Imagery and Deep Learning to Automatically Map Ice-Wedge...

Area of Science:

• Geosciences
• Remote Sensing
• Artificial Intelligence

Background:

• Ice-wedge polygons (IWPs) are crucial permafrost indicators, but their large-scale mapping is challenging.
• Automated methods are needed to efficiently characterize IWPs across vast and complex Arctic regions.

Purpose of the Study:

• To develop and assess a high-throughput mapping workflow for automatic IWP characterization using deep learning (DL) and satellite imagery.
• To evaluate the DL-CNN model's performance and interoperability across different tundra types and image complexities.
• To refine understanding of opportunities and challenges for regional-scale IWP mapping.

Main Methods:

• A high-throughput mapping workflow centered on deep learning (DL) convolutional neural network (CNN) algorithms was developed.
• A region-based CNN object instance segmentation algorithm (Mask R-CNN) was applied to detect and classify IWPs in the North Slope of Alaska.
• Quantitative error statistics and visual inspections were used to validate IWP detection accuracies.

Main Results:

• The DL-CNN workflow achieved high detection accuracies (89%–96%) and classification accuracies (94%–97%) across various tundra types.
• The mapping workflow demonstrated low absolute mean relative error (AMRE) values (0.17–0.23) in discerning IWPs.
• Model performance was robust across heterogeneous tundra cover types, indicating successful interoperability.

Conclusions:

• The developed high-throughput mapping workflow shows robust performance for IWP characterization in diverse tundra landscapes.
• Increasing training sample variability is important for transfer-learning strategies in heterogeneous environments.
• This automated approach offers a promising solution for regional-scale IWP mapping applications.