Research on land cover classification of multi-source remote sensing data based on improved U-net network

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Computer Vision

Background:

• Remote sensing images (optical, SAR) are crucial for land cover classification.
• Single-source data often limits classification accuracy.
• The U-Net network, while common, has limitations in accuracy and parameter efficiency.

Purpose of the Study:

• To improve land cover classification accuracy by combining optical and SAR data.
• To enhance the U-Net network architecture for better feature extraction and reduced parameters.
• To address limitations of traditional U-Net in classifying small-area terrains and overall accuracy.

Main Methods:

• An improved U-Net architecture was developed for land cover classification.
• The enhanced network integrates optical and SAR image bands.
• Modifications include a convolutional block attention mechanism, stride-2 convolutions instead of pooling, and Leaky ReLU activation.

Main Results:

• The proposed method achieved classification accuracies of 0.8905 (optical), 0.8609 (SAR), and 0.908 (combined).
• Mean Intersection over Union (mIoU) values were 0.8104, 0.7804, and 0.8667 for the respective datasets.
• The enhanced U-Net demonstrated improved performance over the traditional U-Net.

Conclusions:

• Combining optical and SAR data with the enhanced U-Net significantly improves land cover classification.
• The architectural modifications effectively capture spatial and channel features while reducing network complexity.
• This approach offers a more robust solution for accurate ground information monitoring using remote sensing data.