研究目的
To investigate different polarimetric distance measures and their potential for use in Random Forests for the classification of PolSAR images, comparing them with hand-crafted features.
研究成果
Intensity is a strong cue for classification, but polarimetric information improves results significantly. Statistic and geodesic distance measures outperform hand-crafted features, with no significant differences among them except for the χ2-distance. Computational efficiency varies, so easy-to-compute distances can be used without loss in accuracy, providing a practical advantage for real-world applications.
研究不足
The study is limited to specific datasets (Oberpfaffenhofen and Berlin images), and the computational efficiency of some distance measures (e.g., geodesic distance) is demanding, which could be optimized. The differences in accuracy between most distance measures are insignificant, suggesting that simpler measures might be preferred for efficiency.
1：Experimental Design and Method Selection:
The study uses Random Forests adapted for PolSAR data, with internal node tests based on various distance measures between Hermitian matrices. The rationale is to evaluate the effectiveness of these measures in feature learning and classification without explicit feature extraction.
2：Sample Selection and Data Sources:
Two datasets are used: a fully polarimetric image over Oberpfaffenhofen, Germany (E-SAR, DLR, L-band) and a dual-polarimetric image over Berlin, Germany (TerraSAR-X, DLR, X-band). Both include urban and natural targets with manual annotations.
3：List of Experimental Equipment and Materials:
PolSAR data from E-SAR and TerraSAR-X sensors, computational resources for implementing Random Forests and distance measures.
4：Experimental Procedures and Operational Workflow:
The image data is divided into five folds; training samples are drawn from four folds, and the fifth is used for testing. Experiments are repeated five times. Distance measures are computed for node tests in the Random Forest, and classification accuracy is evaluated.
5：Data Analysis Methods:
Balanced accuracy (average per-class detection rate) is calculated over test samples, averaged over folds and runs. Statistical analysis compares the performance of different distance measures.
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
