研究目的
To reconstruct full polarimetric (full-pol) SAR data from single polarimetric (single-pol) SAR images using deep neural networks to balance the trade-offs between information richness, system complexity, and resolution.
研究成果
The proposed deep neural network method effectively reconstructs full-pol SAR data from single-pol images, showing good agreement with real data in both qualitative and quantitative analyses. It enables traditional PolSAR applications without prior assumptions, demonstrating robustness and potential for expansion to other non-full-pol data.
研究不足
The method may have errors in low-intensity areas and for rare scatterers like ships due to insufficient training samples. Generalization to completely different terrains might require additional training data.
1：Experimental Design and Method Selection:
The method uses a two-stage deep neural network framework. A feature extractor network based on convolutional neural networks (CNN) extracts hierarchical multi-scale spatial features from grayscale single-pol SAR images. A feature translator network maps these spatial features to normalized polarimetric features for reconstructing the full-pol data.
2：Sample Selection and Data Sources:
Two L-band full-pol SAR images from NASA/JPL UAVSAR over San Diego (SD) and New Orleans (NO) in the USA are used. Training data is from SD, and testing data is from NO to evaluate generalization.
3：List of Experimental Equipment and Materials:
SAR images from UAVSAR, computational resources for deep learning.
4：Experimental Procedures and Operational Workflow:
Preprocess VV-pol intensity images by logarithm and normalization. Use VGG16 pre-trained network for feature extraction, interpolate features, form hyper-column descriptors, and apply a fully-connected network for translation. Train and test the networks.
5：Data Analysis Methods:
Quantitative evaluation using mean absolute error (MAE) and Bartlett distance; qualitative analysis through visual comparison and applications like Freeman decomposition and unsupervised classification.
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
