Fusion of hyperspectral and panchromatic images using an average filter and a guided filter

Highlights

• An average filtering and guided filtering based fusion method is proposed.

• We utilize the average filter to extract the spatial information of the HS image.

• The PAN image is sharpened to enhance the spatial detail.

• The guided filter is utilized to transfer spatial information.

• The proposed method is effective for hyperspectral pansharpening.

Abstract

The fusion of hyperspectral and panchromatic images aims to generate a fused image with high spatial and high spectral resolutions. This paper proposes a novel hyperspectral pansharpening method using an average filter and a guided filter. Based on the traditional component substitution methods, we propose a new and simple method to extract the spatial information of the HS image by average filtering at first. Then to solve the significant spectral distortion, a guided filter is utilized to obtain more detailed spatial information from the PAN image which has been sharpened. The appropriate injection gains matrix is generated by selecting the optimal value of the tradeoff coefficient. The spatial detail is finally injected into each band of the interpolated HS image to achieve the fused image. Experimental results demonstrate that the proposed method can obtain more spatial information and preserve more spectral information in both subjective and objective evaluations.

Introduction

Image fusion is a process which can synthesize and extract the information of two or more different images to obtain improved information by using a certain algorithm [1]. Remote sensing image fusion is an important part of image fusion. Remote sensing image fusion aims to combine the information of different spectral and spatial resolutions images to achieve more useful information [2]. Low spatial resolution hyperspectral (HS) image and high spatial resolution panchromatic (PAN) image are provided by various sensors. The HS image is a low spatial resolution image but has high spectral resolution. The PAN image is a high spatial resolution image with low spectral resolution. Hence the fusion of hyperspectral and panchromatic images is a meaningful technology because it can produce a fused image with the high spectral resolution of the former and the high spatial resolution of the latter.

A large number of methods have been proposed for the fusion of hyperspectral and panchromatic images [3]. They can be broadly separated into five classes: component substitution (CS) algorithms, multiresolution analysis (MRA) algorithms, matrix factorization algorithms, Bayesian algorithms, and hybrid algorithms [4]. CS algorithms and MRA algorithms are traditional fusion methods. CS approaches convert the HS image into another data space where spectral and spatial information are separated. Then the spatial component is replaced by the PAN image [5]. The CS techniques contain algorithms such as the generalized intensity-hue-saturation (GIHS) [6], the principal component analysis (PCA) [7], [8], the Gram-Schmidt (GS) [9], and the GS Adaptive (GSA) [10] method. Those CS methods have three main advantages: 1. High fidelity of the spatial details [11], 2. Fast and simple implementation [6], and 3. Good robustness [11]. But the CS techniques also have a serious shortcoming. The methods can generate significant spectral distortion [12].

MRA approaches inject the spatial information of the PAN image into the HS image. MRA approaches include algorithms, such as the smoothing filter-based intensity modulation (SFIM) [13], the MTF-Generalized Laplacian Pyramid (MTF-GLP) [14], the MTF-GLP with High Pass Modulation (MTF-GLP-HPM) [15], and the decimated wavelet transform (DWT) [16] method. The advantages of the MRA methods are spectral consistency and temporal coherence [17]. The main shortcoming is the complicated implementation [3].

Matrix factorization approaches, Bayesian approaches, and hybrid approaches are proposed recently. Matrix factorization algorithms and Bayesian algorithms are model based methods. They perform well but are accomplished with high computational cost, e.g., the coupled nonnegative matrix factorization (CNMF) [18], the Bayesian HySure [19], the Bayesian Sparsity promoted Gaussian prior (Bayesian Sparse) [20], [21], and the Bayesian Naive Gaussian prior (Bayesian Naive) [22] method. Hybrid methods use concepts from different methods integrated into one method, such as the guided filter PCA (GFPCA) [23] method. The GFPCA method can preserve the spectral information well. However, the GFPCA method produces lots of blurs because the spatial information is not sufficiently integrated in the fused product.

This paper proposes a novel hyperspectral image fusion method with an average filter and a guided filter (AFGF) to solve the problems mentioned above. Based on the model of component substitution methods, we propose a simple and effective approach to retrieve the spatial information of the HS image by average filtering at first. Then to overcome the serious spectral distortion of component substitution methods, the guided filter is used for extracting more detailed spatial information from the PAN image, which has been sharpened. Finally, the appropriate injection gains matrix is presented to reduce the spectral and spatial distortion. Experimental results illustrate that the method using the enhanced PAN image can achieve better effects. Experimental results also demonstrate that the proposed fusion method performs superior in terms of subjective and objective evaluations.