Enhancing surface drainage mapping in eastern Canada with deep learning applied to LiDAR-derived elevation data.

Area of Science:

• Agricultural Science
• Geospatial Analysis
• Remote Sensing

Background:

• Agricultural dykelands in Nova Scotia utilize land forming for surface drainage, crucial for understanding land use amid rising sea levels.
• Traditional methods for delineating and classifying agricultural fields are labor-intensive and time-consuming.
• Deep learning (DL) shows promise in image analysis but hasn't been widely applied to detecting agricultural surface drainage patterns.

Purpose of the Study:

• To develop and test a Mask R-CNN deep learning model for detecting land-formed fields in agricultural dykelands using LiDAR data.
• To improve upon traditional pixel-by-pixel classification with a method identifying cohesive pixel groups as objects.
• To assess the effectiveness of DL in mapping surface drainage patterns for agricultural land management.

Main Methods:

• Utilized Mask R-CNN, a deep learning model, for object detection and segmentation.
• Employed LiDAR-derived elevation data as input for the model.
• Focused on identifying cohesive groups of pixels representing land-formed fields.

Main Results:

• The Mask R-CNN model achieved a high performance with a mean Average Precision (mAP) of 0.89.
• Identified that 53% of Nova Scotia's dykelands are used for agriculture.
• Found that approximately 75% (6924 hectares) of these agricultural fields were land-formed.

Conclusions:

• Deep learning techniques applied to LiDAR data offer novel insights for surface drainage mapping.
• The developed DL model effectively detects land-formed fields, improving efficiency over traditional methods.
• This approach enhances precise agricultural land management capabilities in coastal regions facing environmental changes.