Iterative training sampling and active learning approaches to hyperspectral image classification
Loading...
Links to Files
Permanent Link
Author/Creator
Author/Creator ORCID
Date
2020-01-01
Type of Work
Department
Computer Science and Electrical Engineering
Program
Engineering, Electrical
Citation of Original Publication
Rights
Access limited to the UMBC community. Item may possibly be obtained via Interlibrary Loan through a local library, pending author/copyright holder's permission.
This item may be protected under Title 17 of the U.S. Copyright Law. It is made available by UMBC for non-commercial research and education. For permission to publish or reproduce, please see http://aok.lib.umbc.edu/specoll/repro.php or contact Special Collections at speccoll(at)umbc.edu
This item may be protected under Title 17 of the U.S. Copyright Law. It is made available by UMBC for non-commercial research and education. For permission to publish or reproduce, please see http://aok.lib.umbc.edu/specoll/repro.php or contact Special Collections at speccoll(at)umbc.edu
Abstract
As one of fundamental tasks in remote sensing, hyperspectral image classification (HSIC) has attracted considerable interest. However, two challenging issues arise in HSIC. One is inconsistent classification, which is mainly caused by random training sampling (RTS) of selecting training samples. The other is mixed pixels classification problem which cannot be resolved by conventional pure pixel-based classifier.The first part of this dissertations is to develop a new framework for training sample selection, called iterative training sampling (ITS) which aims to improving the traditional RTS, while reducing the classification inconstancy at the same time. The ITS can be implemented in conjunction with any arbitrary spectral-spatial (SS) classification systems, referred to as ITS spectral-spatial (ITS-SS) classification where ITS-SS expands data cubes iteratively by adding new spatial-filtered (SF-ed) classification maps via feedback loops to the current being processed data cube then regenerates a new set of training samples from expanded data cube to perform classification through an iterative process.
What is more is that ITS can be further combined with active learning (AL), to derive a joint semi-supervised HSIC technique, to be called iterative training sample augmentation by AL (ITSA-AL). The central idea of ITSA-AL is to include training sample augmentation by AL in the iterative feedback process implemented in ITS at each iteration so that the a posteriori probability maps not only can be utilized to augment the current training samples but also can be fed back to expand the current data cube to include new a posteriori spectral-spatial information simultaneously.
In general, a hyperspectral data sample is a mixture of multiple material signatures including background (BKG) which cannot be resolved by traditional pure pixel-based classifier. As a second part of this dissertations, we present a kernel-based approach to hyperspectral mixed pixel classification (HMPC) which includes two nonlinear mixed pixel classifiers, kernel constrained energy minimization (KCEM) and kernel linearly constrained minimum variance (KLCMV). Since KCEM/KLCMV produce real-valued abundance fractional maps for classification, the traditional discreate value-based evaluation tool is not directly applicable. In this case, the commonly used hard classification measures, average accuracy (AA) and overall accuracy (OA) can be further generalized to real-valued mixed class abundance fractional map-based soft classification measures via 3D receiver operating characteristic (3D ROC) analysis-derived detection measures.
Finally, in order to relax the high computational complexity resulting from huge number of spectral bands, a novel concept called iterative band sampling (IBSam) is proposed. The central idea of IBSam is to re-sample bands in each iteration then feeds the SF-ed maps produced by different bands back to expand data cube in iterative manner. Extensive experiments are conducted to demonstrate the utility of IBSam where IBSam not only significantly reduce computing time but also produced the results that comparable to the results obtained by full bands.