Convolutional neural network for high-resolution wetland mapping with open data: Variable selection and the...

Area of Science:

• Environmental science
• Remote sensing
• Geospatial analysis

Background:

• Accurate, up-to-date wetland maps are essential for landscape-scale conservation.
• Current mapping methods often lack resolution, rely on commercial data, or are geographically limited.
• Automated, generalizable, and repeatable mapping approaches are needed.

Purpose of the Study:

• To develop and evaluate a deep learning model for automated, high-resolution wetland mapping at landscape scale.
• To assess the model's performance across diverse geographies and with varying input data.
• To demonstrate the utility of free, remotely sensed data for wetland mapping.

Main Methods:

• A U-Net deep learning architecture was trained using multispectral imagery (NAIP, Sentinel-2) and LiDAR-derived data (intensity, geomorphons).
• The model was trained to map wetlands at 1-meter spatial resolution.
• Model performance was evaluated using accuracy, precision, recall, and AUC metrics, including cross-geography transferability tests.

Main Results:

• The model achieved high accuracy (94%), precision (96.5%), and AUC (95.2%) for wetland mapping at 1-meter resolution.
• The model demonstrated robustness, correctly predicting wetlands even with imperfect training data.
• Cross-geography application showed reduced performance, but limited retraining significantly improved results, indicating spatial generalizability.

Conclusions:

• High-resolution wetland mapping is feasible using free data and deep learning models.
• LiDAR and geomorphon data enhanced model accuracy, but simpler models can be used when these are unavailable.
• This approach offers a powerful tool for informing wetland conservation, restoration, and development decisions.