Automatic Extraction of Water and Shadow from SAR Images Based on a Multi-Resolution Dense Encoder and Decoder...

Area of Science:

• Remote Sensing
• Computer Vision
• Deep Learning

Background:

• Accurate extraction of water and shadow areas in Synthetic Aperture Radar (SAR) images is crucial for various applications but remains challenging with current automated methods.
• Existing techniques struggle to precisely delineate these features due to their complex characteristics in SAR data.

Purpose of the Study:

• To propose a novel framework, the Multi-Resolution Dense Encoder and Decoder (MRDED) network, for precise automatic extraction of water and shadow areas in SAR images.
• To integrate various deep learning architectures, including CNN, ResNet, DenseNet, GCN, and ConvLSTM, to enhance feature extraction capabilities.

Main Methods:

• The proposed MRDED network comprises three main components: Gray Level Gradient Co-occurrence Matrix (GLGCM) for low-level feature extraction, an Encoder utilizing ResNet for multi-resolution feature extraction, and a Decoder with Multi-level Features Extraction and Fusion (MFEF) and Score maps Fusion (SF).
• Two MFEF versions were developed: MFEF1 using ConvLSTM for Improved Chained Residual Pooling (ICRP) and MFEF2 employing Chained Residual Pooling (CRP).
• The framework fuses score maps from both MFEF versions and applies a Softmax function for final extraction.

Main Results:

• The MRDED framework achieved superior performance compared to eight other classification frameworks on large SAR image datasets.
• It demonstrated high accuracy with Pixel Accuracy (PA) of 80.12% and Intersection of Union (IoU) of 73.88% for water extraction.
• For shadow extraction, MRDED reached 88% PA and 77.11% IoU, and for background classification, 95.16% PA and 90.49% IoU.

Conclusions:

• The developed MRDED network offers a robust and accurate solution for automatic water and shadow extraction from SAR images.
• The integration of multi-resolution feature extraction and advanced fusion techniques significantly enhances classification performance.
• MRDED provides a promising advancement in the field of remote sensing image analysis.