Feature Selection for Cross-Scene Hyperspectral Image Classification Using Cross-Domain I-ReliefF

• Main Authors: Chengjie Zhang, Minchao Ye, Ling Lei, Yuntao Qian
• Format: Article
• Language: English
• Published: IEEE 2021-01-01
• Series: IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
• Subjects: Cross-domain I-ReliefF; cross-scene feature selection; hyperspectral images
• Online Access: https://ieeexplore.ieee.org/document/9446730/

In the classification of hyperspectral images (HSIs), too many spectral bands (features) cause feature redundancy, resulting in a reduction in classification accuracy. In order to solve this problem, it is a good method to use feature selection to search for a feature subset which is useful for classification. Iterative ReliefF (I-ReliefF) is a traditional single-scene-based algorithm, and it has good convergence, efficiency, and can handle feature selection problems well in most scenes. Most single-scene-based feature selection methods perform poorly in some scenes (domains) which lack labeled samples. As the number of HSIs increases, the cross-scene feature selection algorithms which utilize two scenes to deal with the high dimension and low sample size problem are more and more desired. The spectral shift is a common problem in cross-scene feature selection. It leads to difference in spectral feature distribution between source and target scenes even though these scenes are highly similar. To solve the above problems, we extend I-ReliefF to a cross-scene algorithm: cross-domain I-ReliefF (CDIRF). CDIRF includes a cross-scene rule to update feature weights, which considers the separability of different land-cover classes and the consistency of the spectral features between two scenes. So CDIRF can effectively utilize the information of source scene to improve the performance of feature selection in target scene. The experiments are conducted on three cross-scene datasets for verification, and the experimental results demonstrate the superiority and feasibility of the proposed algorithm.