Wetland classification based on depth-adaptive convolutional neural networks using leaf-off SAR imagery.

Area of Science:

• Remote Sensing
• Artificial Intelligence
• Wetland Ecology

Background:

• Deep learning (DL) shows promise for wetland classification using optical data.
• Classifying large-scale wetlands with radar data using DL requires further evaluation.
• Sentinel-1 Synthetic Aperture Radar (SAR) offers potential advantages over optical data for wetland mapping.

Purpose of the Study:

• To develop and evaluate an end-to-end depth-adaptive convolutional neural network (CNN) for large-scale wetland mapping.
• To assess the effectiveness of incorporating multi-land cover proximity information and SAR denoising techniques.
• To optimize the balance between model complexity and accuracy in wetland classification.

Main Methods:

• Developed a depth-adaptive CNN based on U-Net architecture for wetland classification.
• Utilized time-series leaf-off Sentinel-1 SAR imagery and ancillary data.
• Investigated the impact of multi-land cover proximity and CNN-based SAR denoising.

Main Results:

• The proposed DL method achieved high accuracy (OA=0.93, MIoU=0.60), outperforming traditional Random Forest (RF) methods (OA=0.89, MIoU=0.18).
• The DL approach demonstrated reduced computational cost compared to other state-of-the-art CNNs without sacrificing accuracy.
• Incorporating proximity information and SAR denoising enhanced classification accuracy, particularly for forested wetlands.

Conclusions:

• The novel DL method efficiently classifies wetlands by integrating denoised SAR imagery and ancillary data.
• This approach offers a valuable tool for operational wetland mapping at large spatial scales.
• The findings highlight the potential of combining DL and radar data for environmental monitoring.