MIMN-DPP: Maximum-information and minimum-noise determinantal point processes for unsupervised hyperspectral band selection

Highlights：

• We propose a new double-graph model for effectively measuring the correlation between bands of hyperspectral image.

• A new maximum information and minimum noise (MIMN) criterion is proposed for unsupervised band selection. MIMN criterion can select a band subset with rich information and low noise that can supplement for robustness.

• Combining k-Determinantal Point Process (k-DPP) with MIMN, a new MIMN-DPP search method is proposed for selecting the band subset with low-redundancy and low noise based on double-graph model.

• MIMN-DPP method will result in high classification accuracy and improved robustness despite of the random sampling in k-DPP.

摘要

•We propose a new double-graph model for effectively measuring the correlation between bands of hyperspectral image.•A new maximum information and minimum noise (MIMN) criterion is proposed for unsupervised band selection. MIMN criterion can select a band subset with rich information and low noise that can supplement for robustness.•Combining k-Determinantal Point Process (k-DPP) with MIMN, a new MIMN-DPP search method is proposed for selecting the band subset with low-redundancy and low noise based on double-graph model.•MIMN-DPP method will result in high classification accuracy and improved robustness despite of the random sampling in k-DPP.

论文关键词：Hyperspectral images (HSI),Unsupervised band selection,Maximum information and minimum noise (MIMN) criterion,Determinantal point processes (DPP)

论文评审过程：Received 23 May 2019, Revised 20 October 2019, Accepted 18 January 2020, Available online 21 January 2020, Version of Record 31 January 2020.