National wetland mapping using remote-sensing-derived environmental variables, archive field data, and artificial...

Area of Science:

• Environmental Science
• Remote Sensing
• Artificial Intelligence

Background:

• Effective wetland conservation relies on accurate, fine-scale national inventories, which are currently lacking.
• Wetlands pose mapping challenges due to their intricate spatial patterns and indistinct boundaries.
• Advancements in remote sensing, field data, and AI offer new possibilities for detailed wetland mapping.

Purpose of the Study:

• To create a fine-scale national wetland map for mainland France using AI.
• To leverage environmental variables from remote sensing and archival data for wetland identification.
• To assess the accuracy of the AI-derived map against existing global, European, and national wetland maps.

Main Methods:

• An artificial intelligence (AI) random forest model was trained using approximately 135,000 soil or flora plots.
• The model incorporated 5m topographical variables from an airborne digital terrain model (DTM) and geological map data.
• Model calibration used spatial cross-validation, optimizing for the precision-recall area under the curve (PR-AUC).
• Validation involved a ground survey of approximately 3000 plots along transects and comparison with nine existing wetland maps.

Main Results:

• The AI-derived suitability map achieved a PR-AUC of 0.76, effectively showing gradual wetland boundaries and fine-grained patterns.
• The resulting binary wetland map demonstrated significantly higher accuracy (F1-score 0.75, overall accuracy 0.67) than existing maps.
• The study successfully mapped wetlands across mainland France (approximately 550,000 km²).

Conclusions:

• The developed AI approach provides a valuable tool for fine-scale national wetland mapping.
• The high-resolution maps are crucial for targeted spatial planning and environmental management.
• These maps will support biodiversity conservation, water resource management, and carbon storage initiatives.