Deep Nonnegative Dictionary Factorization for Hyperspectral Unmixing

Wenhong Wang,Hongfu Liu

doi:10.3390/rs12182882

Abstract

As a powerful blind source separation tool, Nonnegative Matrix Factorization (NMF) with effective regularizations has shown significant superiority in spectral unmixing of hyperspectral remote sensing images (HSIs) owing to its good physical interpretability and data adaptability. However, the majority of existing NMF-based spectral unmixing methods only adopt the single layer factorization, which is not favorable for exploiting the complex and structured representation relationship of endmembers implied in HSIs. In order to overcome such an issue, we propose a novel two-stage Deep Nonnegative Dictionary Factorization (DNDF) approach with a sparseness constraint and self-supervised regularization for HSI unmixing. Beyond simply extending one-layer factorization to multi-layer, DNDF conducts fuzzy clustering to tackle the mixed endmembers of HSIs. Moreover, self-supervised regularization is integrated into our DNDF model to impose an effective constraint on the endmember matrix. Experimental results on three real HSIs demonstrate the superiority of DNDF over several state-of-the-art methods.

Highlights

Hyperspectral remote sensing images (HSIs) capture hundreds of narrow spectral channels, and provide detailed spectral information and spatial distribution status about the target ground objects
Existing deep Nonnegative Matrix Factorization (NMF)-based methods commonly focus on factorizing the coefficient matrix to explore the abstract features of the data [17], which is not favorable for efficiently utilizing the complex hierarchical and multi-layers structured representation information between the endmembers and the mixed pixels included in HSIs
To overcome the above issues, we propose a novel Deep Nonnegative Dictionary Factorization (DNDF) model in this paper, by which the single-layer NMF is used in succession to decompose the learned dictionary layer by layer within the pre-training stage, a fine-tuning stage is employed to further optimize all the matrix factors obtained in the previous stage by reconstructing the original data via the production of these factors

Summary

Introduction

Hyperspectral remote sensing images (HSIs) capture hundreds of narrow spectral channels, and provide detailed spectral information and spatial distribution status about the target ground objects. With their rich spectral information for distinguishing different objects, HSIs have received many successful applications, such as segmentation [1], classification [2], target detection [3], tracking [4], and recognition [5]. Existing deep NMF-based methods commonly focus on factorizing the coefficient matrix to explore the abstract features of the data [17], which is not favorable for efficiently utilizing the complex hierarchical and multi-layers structured representation information between the endmembers and the mixed pixels included in HSIs

Objectives

Methods

Findings

Conclusion