研究目的
To develop an enhanced deep learning framework for accurate water body extraction from very high-resolution remote sensing images, addressing challenges such as blurred boundaries and noise.
研究成果
The proposed Deep U-Net-CRF-RR method effectively extracts water bodies from high-resolution remote sensing images, reducing noise and preserving boundaries, outperforming existing methods in both quantitative and qualitative assessments, and showing promise for practical applications in water resource management.
研究不足
The method requires extensive training data and computational resources; it may be sensitive to parameter settings in CRF and RR, and performance depends on the quality of ground truth labels, which can have inaccuracies.
1：Experimental Design and Method Selection:
The methodology involves a modified Deep U-Net architecture for initial classification, followed by a fully connected conditional random field (FCCRF) model for refinement, and regional restriction (RR) using superpixels to enhance consistency. Mean-field inference is used for efficient computation.
2：Sample Selection and Data Sources:
Data from GaoFen-2 (GF-2) and WorldView-2 (WV-2) satellites, covering lakes, rivers, and ponds in Liuzhou, China. Images are from the near-infrared band with spatial resolutions of 0.8 m and 0.5 m, orthorectified and labeled by China's National Geography Census.
3：8 m and 5 m, orthorectified and labeled by China's National Geography Census.
List of Experimental Equipment and Materials:
3. List of Experimental Equipment and Materials: Computer with Inter(R) Xeon(R) E5-2687W CPU at
4：00 GHz, 00-GB RAM, NVIDIA GRID M60-8Q (8 GB) GPU. Software:
Keras library with TensorFlow backend.
5：Experimental Procedures and Operational Workflow:
Preprocess images by subtracting mean value, divide into 2048x2048 patches, apply data augmentation (rotation, flipping). Train Deep U-Net using binary cross-entropy loss and Adam optimizer. Use FCCRF with Gaussian kernels for pairwise potential, and RR with superpixels for correction. Perform mean-field inference for final segmentation.
6：Data Analysis Methods:
Evaluate using confusion matrix metrics: false alarm rate (PFA), missed alarm rate (PMD), Kappa coefficient (KC), and percentage correctly classified (PCC). Compare with CV-based, MRF-based, SegNet, and DeepWaterMap-3 methods.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
