研究目的
To study the potential of Tomographic Synthetic Aperture Radar (TomoSAR) data for generating semantic land cover maps in a supervised framework and compare it with polarimetric SAR data.
研究成果
TomoSAR data, when combined with carefully designed features including spatial and 3-D descriptors, outperforms polarimetric SAR data in land cover classification, particularly for challenging classes like city and road. This demonstrates the potential of TomoSAR for semantic mapping, with recommendations for future development of more features and advanced classification methods.
研究不足
The study relies on hand-crafted features and a specific dataset; future work could explore more sophisticated algorithms and additional features. The approach may be limited by the availability of multi-baseline data and computational resources.
1：Experimental Design and Method Selection:
The methodology involves extracting features from TomoSAR data, including covariance matrices and tomograms, and using a Random Forest classifier for supervised classification. Features are designed to capture pixel-wise, spatial, and 3-D contextual information.
2：Sample Selection and Data Sources:
The PolTom dataset from the E-SAR sensor (DLR) in L-Band Oberpfaffenhofen scene is used, with 14 images in VV polarization for TomoSAR and one fully polarimetric image for PolSAR. The dataset is manually annotated with reference labels for five classes: city, road, forest, shrubland, and field.
3：List of Experimental Equipment and Materials:
E-SAR sensor data, computational tools for feature extraction and classification.
4：Experimental Procedures and Operational Workflow:
Features are extracted from covariance matrices and tomograms, including statistical moments, cross-correlation coefficients, and 3-D geometric moments. Spatial features are added using patches and superpixels. The Random Forest classifier is trained on selected bands and evaluated on the remaining data.
5：Data Analysis Methods:
Classification accuracy is assessed using overall accuracy (OA) and balanced accuracy (BA), with feature importance analyzed.
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
