Recognizing protected and anthropogenic patterns in landscapes using interpretable machine learning and satellite...

Area of Science:

• Environmental science
• Remote sensing
• Machine learning

Background:

• Accurate land cover mapping is vital for environmental research.
• Defining land cover classes, like naturalness and human influence, is challenging but essential for biodiversity and climate monitoring.
• Machine learning approaches are increasingly used for land cover analysis.

Purpose of the Study:

• To develop an interpretable machine learning approach for mapping naturalness and human influence using satellite imagery.
• To utilize territorial protected and anthropogenic areas as proxies for naturalness and human influence.
• To enhance the interpretability and consistency of machine learning models in remote sensing.

Main Methods:

• Training a weakly-supervised convolutional neural network (CNN).
• Applying attribution methods like Grad-CAM and occlusion sensitivity mapping.
• Proposing a novel image-to-image network architecture with a task-specific head connected by a full-resolution feature layer.

Main Results:

• The developed approach effectively maps patterns of naturalness and human influence.
• Intermediate layer activations were analyzed to establish consistent relationships with attributions.
• Attributions were found to be consistent across different scenes, enabling large-scale analysis.

Conclusions:

• The interpretable machine learning approach is a promising tool for observing and assessing naturalness and territorial protection.
• The novel network architecture facilitates detailed analysis and consistent large-scale remote sensing data analysis.
• This method contributes to improved land cover mapping for conservation and environmental monitoring.