A novel method for multispectral image pansharpening based on high dimensional model representation
Introduction
There have been great improvements in remote sensing imaginery technologies over the years. However, researchers still encounter difficulties due to lack of high spatial resolution of satellite sensors. To overcome the absence of high resolution image data in imaging instruments, most of the remote sensing imaging systems capture two different images of the same geographical location at the same time: multispectral image (MS) and panchromatic image (PAN). An MS image contains rich spectral information with several bands but also low spatial resolution. Contrarily, PAN image has high spatial resolution with only one spectral band. Pansharpening is an important technique to generate high resolutional multispectral image (HMS) by fusing MS and PAN images of the same scene to enhance different applications such as earth monitoring, object/crop classification, target detection, and segmentation (Loncan et al., 2015, Bhandari and Rahul, 2019).
Pansharpening methods vary according to the fusing process and can be classified into four major categories. The first category covers component substitution (CS) based methods. CS based methods replace the components in an MS image with a PAN image, and therefore the MS image is transformed into a different color space. The different details of the PAN image and the bands of the transformed MS image are extracted allowing us to merge these details into a low resolutional MS image. One of the CS based methods relies on Intensity-Hue-Saturation (IHS) transform which is also named as IHS pansharpening (Carper et al., 1990, Al-Wassai et al., 2011). Another method modifies IHS sharpening by finding the gain coefficients of the bands of the MS image adaptively and transferring the edges of the PAN image to the MS image to increase the spectral enhancement (Rahmani et al., 2010). IHS based methods are easy to apply hence they are widely used; however, they may cause spectral distortion. Another well-known CS based technique uses principal component analysis (PCA) which is a linear transformation that utilizes principal components of the MS image and replacing them with the components of the PAN image (Kwarteng & Chavez, 1989), but it may produce distortion caused by local dissimilarities between the components of the transformed MS image and the PAN image (Thomas, Ranchin, Wald, & Chanussot, 2008). Gram-Schmidt (GS) decomposition (Laben & Brower, 2000) and partial replacement adaptive method (PRACS) (Choi, Yu, & Kim, 2011) are also widely used CS based methods. These methods, which are also based on transformation and substitution similar to the abovementioned CS based methods, improve the spatial resolution of MS images but may not preserve spectral information with the MS image.
Multiresolution analysis (MRA) based pansharpening methods are in the second category and mainly use the separation of low and high frequencies through a multiresolution decomposition (Ranchin & Wald, 2000). These methods obtain spatial information from the high-frequency component of the PAN image to construct integration with the upsampled MS image. MRA based methods vary in terms of extraction of spatial details like Laplacian pyramid (LP) (Aiazzi et al., 1999), and wavelet transform methods (Nunez et al., 1999, King and Wang, 2001). MRA based fusion is combined with some other algorithms such as genetic algorithms and particle swarm optimization (Garzelli and Nencini, 2006, Jamal and Faez, 2011) to minimize missing spatial information. Compared to CS based methods, these MRA based methods provide better spectral representation. Nevertheless, they may suffer from distortion in terms of spatial enhancement.
The third category consists of model based methods. Variational optimization models are included in this category (Ballester et al., 2006, Palsson et al., 2013, Duran et al., 2017). Although spectral distortion is minimized by variational methods, these methods suffer from blurring effects and texture contorts. Bayesian statistics based pansharpening methods have been proposed as a model based approach for making good on spectral/spatial trade-off (Fasbender et al., 2008, Zhang et al., 2012, Wang et al., 2018). These methods need some prior knowledge and tuning of hyperparameters which can affect sharpening, and therefore may lead to failures in performance. Most of the model based pansharpening methods rely on linear models to acquire spatial information.
Deep learning (DL) based models have been proposed after the achievements of many different image analysis problems in recent years (Krizhevsky et al., 2017, Shen et al., 2017, Altan and Karasu, 2020). These techniques are a new generation of pansharpening methods and can be considered as the fourth category. They are mostly used to overcome the limitations of linear models (Masi et al., 2016, Wei et al., 2017). Convolutional neural networks (CNN) is a common tool for pansharpening problem (Dong et al., 2015, Scarpa et al., 2018), but the accuracy highly depends on the network architectures and this may induce the generalization problem. To overcome it, further studies used deeper networks but an increasing number of layers led to worsening training accuracy of the network (Wang et al., 2019). To tackle the previously mentioned problems, various deep learning pansharpening methods were proposed which were combined with traditional approaches such as using detail injections before or after the network (Deng et al., 2020, Liu et al., 2020) and Bayesian theory (Guo, Zhuang, & Guo, 2020). However, these methods may suffer from losing high-frequency details because the pre-interpolated MS image is joined with the PAN image as the training data. Non-negative matrix and tensor decomposition based pansharpening methods differ from the abovementioned four pansharpening categories. Matrix factorization based image fusion is an unmixing based solution and served to separate the signals into a dictionary matrix and a weight matrix (Yokoya, Yairi, & Iwasaki, 2011). Tensor decomposition based methods are the extension to matrix decomposition which takes multilinear correlations into account. Tensor decomposition was proposed for pansharpening after the achievements on the super-resolution problem (Li et al., 2018a, Bu et al., 2020). These methods accomplish improvement on the resolution, on the other hand, Tucker decomposition does not guarantee a unique solution as a tensor decomposition algorithm.
Our research aims to develop a novel and efficient CS based pansharpening algorithm for empowering spatial resolution quality of the fused image. To create an effective pansharpening technique, a new High Dimensional Model Representation (HDMR) based method, called Adaptive HDMR, is developed. HDMR is a well known orthogonal data decomposition technique for multivariate functions which was first proposed by Sobol (Sobol, 1993). It is a divide-and-conquer algorithm that allows us to express a multivariate function through functions with fewer variables (Tunga & Demiralp, 2008). This algorithm is also used as a multivariate data partitioning technique in data processing and modeling problems (Karaca and Tunga, 2018, Ozay and Demiralp, 2014, Tunga and Karahoca, 2015, Karcılı and Tunga, 2017). In another study, HDMR is proposed as a powerful decomposing method for image enhancement. The study demonstrates the physical meanings of HDMR components in terms of image processing (Tunga & Koçanoğulları, 2018).
The success of the HDMR method in image decomposition and enhancement led us to use HDMR as a pansharpening method. We investigated if it is possible to transform an MS image with both high spatial and spectral resolution required for the pansharpening problem with the help of HDMR components. The preliminary experiments showed that the classical HDMR expansion causes deformation on the spectral structure. To cope with this problem and develop an efficient pansharpening method, the HDMR components are multiplied by some scaling factors and a new structure which is called Adaptive HDMR, is built. The development of an HDMR based pansharpening method contributes significantly to spatial enhancement while preserving spectrality. When the above categories of pansharpening methods and the experimental results of this study are analyzed, four main contributions of the proposed method can be observed. (1) Adaptive HDMR preserves spectral structure better than the CS based methods. (2) Adaptive HDMR is a spatially superior method unlike most of MRA based methods. (3) Adaptive HDMR has fewer parameters than model based methods which makes it easier to implement. (4) Adaptive HDMR has low model complexity according to the number of used features when compared with the new deep learning based methods. The statistical results in this paper show that the proposed displays an impressive performance on eliminating the addressed gaps in the state-of-the-art pansharpening methods.
The paper is organized as follows. Section 2 explains the mathematical background of HDMR and its components. This section also provides information about the HDMR components of an MS image. Section 3 covers the details of the Adaptive HDMR and the steps of the algorithm proposed as a pansharpening technique. Section 4 includes information about the data used in experiments and performance metrics. Experimental results on different low resolutional MS images of our method and the results on the same MS image of different state-of-the-art pansharpening techniques are also compared in this section. Section 5 contains discussions about the findings as well as the effects of the model parameters on these results. Conclusion and potential future works are given in Section 6.
Section snippets
High dimensional model representation
HDMR is one of the methods that is used to decompose a given multivariate function into its orthogonal components. The basic formula of the method is given aswhere N is the number of independent variables describing the given multidimensional problem. HDMR expansion is finite and the right-hand terms are uniquely obtained using the following vanishing conditions (Tunga and Demiralp, 2012, Tunga, 2015)
The proposed method – adaptive HDMR
The proposed pansharpening method of this study is the Adaptive HDMR algorithm. This algorithm starts with obtaining the classical HDMR components of the upsampled MS image using Eqs. (9), (10), (11), (12). The spatial details of the MS image are revealed by this HDMR decomposition. However, as mentioned earlier, the higher-order HDMR terms of the MS image holds greater spatial details. The retention of the spatial details causes deformation on the spectral structure. To overcome the spectral
Experimental design and datasets
We set up three experiments using four-band digital imagery from different satellites (Li et al., 2018b, Hasanlou and Saradijan, 2016). Four band imagery typically contains red, green, blue, and near-infrared (NIR) bands.
The first experiment is carried out on an urban scene from Pleiades satellite. The sizes of the PAN and MS images for this data set are and respectively. The images of the second experiment are taken from Spot-6 and they have and sizes for
Discussion
We analyzed the effects of the scaling factors, s in terms of spectral fidelity and spatial improvement for all three remote sensing images. We have excluded and from the analysis due to the fact that Adaptive HDMR’s constant term, and one of the univariate components, , do not have significant effect on the results of the proposed pansharpening method. The effect of the other scaling factors can be observed in Fig. 6. Fig. 6a illustrates QNR results for all of the s in .
Conclusion
In this paper, we introduced a brand new pansharpening method based on HDMR. We decomposed the MS image into its HDMR components each of which comprises details of the target location. We developed a three-step algorithm: The first step of our proposed algorithm is inserting scaling factors into the HDMR expansion to enhance the effects of the HDMR components. This makes the most informative components the dominant ones in the expansion. In the second step, weighting factors are calculated for
CRediT authorship contribution statement
Evrim Korkmaz Özay: Conceptualization, Investigation, Visualization, Writing - original draft, Writing - review & editing. Burcu Tunga: Methodology, Investigation, Writing - review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (63)
A survey of pansharpening methods with a new band-decoupled variational model
ISPRS Journal of Photogrammetry and Remote Sensing
(2017)- et al.
An interpolation-based texture and pattern preserving algorithm for inpainting color images
Expert Systems with Applications
(2018) - et al.
Detecting GSM churners by using Euclidean Indexing HDMR
Applied Soft Computing
(2015) Multispectral fusion of multisensor image data by the generalized laplacian pyramid
IEEE 1999 International Geoscience and Remote Sensing Symposium
(1999)MTF-tailored multiscale fusion of high-resolution MS and PAN imagery
Photogrammetric Engineering & Remote Sensing
(2006)- Al-Wassai, F. A., Kalyankar, N. V. & Al-Zuky, A. (2011). The ihs transformations based image fusion. arXiv preprint,...
- et al.
Managing computational complexity using surrogate models: A critical review
Research in Engineering Design
(2020) Ensemble of surrogates and cross-validation for rapid and accurate predictions using small data sets
AI EDAM
(2019)Multispectral and panchromatic data fusion assessment without reference
Photogrammetric Engineering & Remote Sensing
(2008)- et al.
Recognition of COVID-19 disease from X-ray images by hybrid model consisting of 2D curvelet transform, chaotic salp swarm algorithm and deep learning technique
Chaos, Solitons & Fractals
(2020)
A variational model for P+ XS image fusion
International Journal of Computer Vision
A novel local contrast fusion-based fuzzy model for color image multilevel thresholding using grasshopper optimization
Applied Soft Computing
Hyperspectral and multispectral image fusion via graph Laplacian-guided coupled tensor decomposition
IEEE Transactions on Geoscience and Remote Sensing
The use of intensity-hue-saturation transformations for merging SPOT panchromatic and multispectral image data
Photogrammetric Engineering and remote sensing
Comparison of three different methods to merge multiresolution and multispectral data: Landsat TM and SPOT panchromatic
Photogrammetric Engineering & Remote Sensing
A new adaptive component-substitution based satellite image fusion by using partial replacement
IEEE Transactions on Geoscience and Remote Sensing (TGRS)
High dimensional model representation and its application varieties
Mathematical Research
Detail injection-based deep convolutional neural networks for pansharpening
IEEE Transactions on Geoscience and Remote Sensing
Image super-resolution using deep convolutional networks
IEEE Transactions on Pattern Analysis and Machine Intelligence
Bayesian data fusion for adaptable image pansharpening
IEEE Transactions on Geoscience and Remote Sensing
PAN-sharpening of very high resolution multispectral images using genetic algorithms
International Journal of Remote Sensing
Bayesian pan-sharpening with multiorder gradient-based deep network constraints
IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Quality assessment of pan-sharpening methods in high-resolution satellite images using radiometric and geometric index
Arabian Journal of Geosciences
A new pan-sharpening method using multiobjective particle swarm optimization and the shiftable contourlet transform
ISPRS Journal of photogrammetry and Remote Sensing
Indusion: Fusion of multispectral and panchromatic images using the induction scaling technique
Geoscience and Remote Sensing Letters
A wavelet based algorithm for pan sharpening landsat 7 imagery
IEEE Proceedings of International Geoscience and Remote Sensing Symposium
Imagenet classification with deep convolutional neural networks
Communications of the ACM
Extracting spectral contrast in Landsat Thematic Mapper image data using selective principal component analysis
Photogrammetric Engineering & Remote Sensing
Coupled tensor decomposition for hyperspectral pansharpening
IEEE Access
Cited by (15)
Contrast stretching based pansharpening by using weighted differential evolution algorithm
2022, Expert Systems with ApplicationsCitation Excerpt :There is not a standard approach for categorizing pansharpening methods. Component-Substitution (CS) (Choi, Yu, & Kim, 2011) and Multiresolution Analysis (MRA) based pansharpening methods are the two most common types of classical pansharpening methods (Özay & Tunga, 2021). The development of new techniques capable of producing PI with high spectral and spatial accuracy is an inevitable need for the commercial satellite imagery market.
Pansharpening approach via two-stream detail injection based on relativistic generative adversarial networks
2022, Expert Systems with ApplicationsCitation Excerpt :Recently, Jolicoeur-Martineau (2018) introduced Relativistic average GAN (RaGAN) to boost the generative accuracy of standard GANs, which yielded remarkable results in several applications (Kupyn, Martyniuk, Wu, & Wang, 2019; Wang et al., 2018; Yadav, Chen, & Ross, 2019). With the tremendous improvements of hardware, convolutional neural networks (CNNs) have recently become a powerful tool to deal with pansharpening and its related applications(see Aadil, Rahim, & ul Hussain, 2019; Aiazzi, Alparone, Baronti, Garzelli, & Selva, 2006; Masi et al., 2016; Meng et al., 2019; Nogales, Delgado-Martos, Melchor, & García-Tejedor, 2021; Özay & Tunga, 2021; Priego-Torres, Sanchez-Morillo, Fernandez-Granero, & Garcia-Rojo, 2020; Ronneberger et al., 2015; Shen et al., 2019; Singh et al., 2020; Thomas, Ranchin, Wald, & Chanussot, 2008; Vivone et al., 2014; Wang et al., 2018; Wei, Bioucas-Dias, Dobigeon, & Tourneret, 2015; Wei et al., 2017; Yadav et al., 2019; Yuan, Wei, Meng, Shen, & Zhang, 2018). The CNN-based methods depend on the large-scale data set training to learn a nonlinear functional mapping between the low spatial resolution MS images and the high spatial resolution MS images.
Triple-loss driven generative adversarial network for pansharpening
2024, IET Image ProcessingSWPanGAN: A hybrid generative adversarial network for pansharpening
2024, IET Image ProcessingA spatiotemporal transferable image fusion technique for GeoEye-1 satellite imagery
2023, Aerospace SystemsEfficient and Accurate Neural Fingerprints Obtained via Mean Curve Length of High Dimensional Model Representation of EEG Signals
2023, European Signal Processing Conference