Abstract—Currently, the majority of manifold learning algorithms applied to SAR image feature extraction are unsupervised. The semi-supervised manifold learning becomes a research trend for it can make full use of class information and be more coincident with actual data. Considering the non-uniform distribution of high-resolution SAR data and the problem of manually set neighbor values, adaptive neighborhood selection was introduced into the Semi-supervised Discriminative Locality Alignment (SDLA) and adaptive neighborhood selection semi-supervised discriminative locality alignment (ANSSDLA) was proposed for building extraction of high-resolution SAR data. Then, single polarization TerraSAR-X data and fully polarization RADARSAT-2 data were taken as experiment data to validate the ANSSDLA algorithm. It was found that the ANSSDLA algorithm has strong adaptability by comparing the results of ANSSDLA and SDLA. In addition, the result of comparative experiments of multi-polarization data shows that cross polarization data is more suitable for building extraction.

 

Index Terms—synthetic aperture radar, manifold learning, adaptive neighborhood selection, semi-supervised discriminative locality alignment

Cite: Bo Cheng and Shiai Cui, "Adaptive Neighborhood Selection Semi-supervised Discriminative Locality Alignment Based Urban Building Areas Extraction from High-Resolution SAR Images," International Journal of Signal Processing Systems, Vol. 5, No. 1, pp. 12-17, March 2017. doi: 10.18178/ijsps.5.12-17