Compression and reinforce variation with convolutional neural networks for hyperspectral image classification
Introduction
The classification of Hyperspectral images (HSI) is a necessary process for different earth observation applications [1], such as war areas [2], military [3], [4], environmental monitoring [5], agriculture [6], small object detection [7], [8], [9], food quality [10], medical [11], [12], and others. HSI can extract spectral data from hundreds of surface object continuous spectrum segments. The spatial resolution of HSI data sets has substantially improved due to the rapid development of remote-sensing technology, which vastly improves the ability of HSI data sets to express distinct objects appropriately.
As described in [1], there are several critical challenges with HSI classification tasks. For example, hyperspectral data has hundreds of band values, and the information between the spectral bands is usually redundant, resulting in a large data dimension and a high computing demand. More so, the presence of mixed pixels causes significant interference in the categorization of HSI, as a single pixel frequently correspondings to numerous object categories and is commonly misclassified. Furthermore, manually labeling HSI samples are expensive, resulting in a tiny number of off-the-shelf labeled samples.
In high-dimensional data analysis, visualization, and modeling, dimensionality reduction methods (DRM) are commonly employed as preprocessing. DRM seeks to increase the performance of estimated accuracy, visualization, and comprehension of learned knowledge in general. DRMs can generally be divided into feature extraction and feature (band) selection [13], [14], [15], [16], [17]. The DRM is one of the most critical HSI processes, aiming to reduce model complexity and overfitting, and these new lower dimensions of features represent the original ones. The feature extraction approach reduces the dimensionality through particular mathematical processes to generate a new subset of features that are a part of the original dataset and retain only the pertinent data that can improve the final goal while discarding the rest. On the other hand, feature selection algorithms (FSA) select a subset of features most relevant to the problem to improve computational efficiency and reduce generated model errors by deleting unrelated features or noise. FSA methods have three types, filter, wrapper, and embedded [14], [18], [19].
Principal component analysis (PCA), linear discriminant analysis (LDA), independent component analysis (ICA), kernel PCA (KPCA) [5], [20], [21], and region-aware latent features fusion-based clustering [18] are examples of feature extraction methods. The most common compressing method in HSI is PCA. PCA uses the correlation between features to find data patterns. It aims to identify the highest variance directions in high-dimensional data and project them onto a subspace with the same or fewer dimensions as the original. On the other hand, KPCA is nonlinear unsupervised features extraction, the kernelized version of PCA [5], [10], [22], [23], [24]. LDA is a linear supervised feature extraction method that aims to minimize class variations. The general concept behind LDA is very similar to PCA.
In contrast, PCA attempts to find the orthogonal component axes of maximum variance. LDA aims to find the feature subspace that optimizes class separability. Thus, it increases computational efficiency, reduces overfitting, and highlights the quality of the classification. It has been widely used to classify agricultural and food products and other applications based on hyperspectral data [25], [26], [27]. ICA is a linear, supervised feature extraction; considered a further step of PCA and a powerful tool for extracting source signals or valuable information from the original data. Compared to the PCA and LDA, ICA optimizes higher-order statistics such as kurtosis (non-Gaussian), yielding independent components [20].
In FSA, filter methods nominate features according to specific predefined criteria before feeding to a learning model such as minimum-redundancy maximum-relevance (mRMR), trivariate mutual information-clonal selection algorithm, distance-based criteria, consistency-based criteria, and manifold learning-based criteria [28], [29]. The wrapper method chooses and evaluates the candidate features through a chosen training model. So, the research algorithm of the best subset of features is basically “wrapped” around the model. This feature selection method is considered costly due to its computational complexity and the long execution time. It is better in classification than in the filter methods, but filter methods are faster, less complex, and better chosen for high dimensional datasets compared to wrapper methods. Recursive feature elimination (SVM-RFE) is one of the wrapper methods. The SVM-REF employs the weight vector as a ranking criterion to select the features that lead to the most considerable margin of class separation. The embedded methods use the advantages of filters and wrappers methods like the least absolute shrinkage and selection operator and the partial least square [14], [19], [30], [31].
Over the past decades, automatic feature representation and extraction using machine learning techniques have gained popularity over handcrafted techniques for HSI classification [32], [33].
For instance, invariant attribute profiles [34] and texture profiles [35] were effective techniques for extracting spatial–spectral features from HSI. In addition, methods such as sparse representation, known as subspace-based learning and manifold learning [36], [37], have proven their ability to capture the high-dimensional structure of HSI by mapping the high-dimensional original space to low-dimensional subspace. However, the methods mentioned above are limited in data fitting and representation ability [38], [39]. In recent years, deep learning models (DL) have superseded the methods mentioned above on many levels, including feature extraction or representation, feature selection, and classification [40].
Section snippets
Literature review
Many DL models have been proposed to address the problems of traditional feature representation and HSI classification. The convolution neural network (CNN) and its variants are well-known DL models based on hierarchical feature learning and classification [41]. CNNs are widely used in HSI classification problems. They are commonly composed of a stack of the convolutional layer with different kernel sizes and activation functions to represent and extract features. CNNs can be used to build
Proposed methodologies
This section provides complete information on the operation of the compression and reinforced variation (CRV) method and explains this study model’s structure, multi-hybrid deep learning (MHDL).
Experimental
The proposed model in this study and preprocessing operation were used to train three commonly used hyperspectral image datasets: the Indian Pines, Pavia-University, and Kennedy Space Center.
Conclusion
The HSI dataset has a large number of classes and bands; therefore, these classes typically share the same values across all classes. This study proposed a novel feature selection method called compression and reinforced variation (CRV) to reduce the dimension of HSI. Furthermore, the structure learning model of this study, multi-hybrid deep learning (MHDL), enhanced the extraction of spectral–spatial features by using hybrid layers of CNN and kernel size, and it provided more stable results
CRediT authorship contribution statement
Dalal AL-Alimi: Formal analysis, Methodology, Software, Visualization, Writing – original draft. Zhihua Cai: Conceptualization, Formal analysis, Supervision, Writing – review & editing. Mohammed A.A. Al-qaness: Conceptualization, Formal analysis, Project administration, Validation, Writing – review & editing. Abdelghani Dahou: Formal analysis, Validation, Writing – original draft. Eman Ahmed Alawamy: Formal analysis, Validation, Writing – review & editing. Sakinatu Issaka: Validation, Writing –
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.
Acknowledgments
This work was supported by National Natural Science Foundation of China (Grant No. 62150410434).
References (68)
- et al.
Improved nitrogen retrievals with airborne-derived fluorescence and plant traits quantified from VNIR-SWIR hyperspectral imagery in the context of precision agriculture
Int. J. Appl. Earth Obs. Geoinf.
(2018) - et al.
An overview on spectral and spatial information fusion for hyperspectral image classification: Current trends and challenges
Inf. Fusion.
(2020) - et al.
Detecting newly grown tree leaves from unmanned-aerial-vehicle images using hyperspectral target detection techniques
ISPRS J. Photogramm. Remote Sens.
(2018) - et al.
Multiple instance hybrid estimator for hyperspectral target characterization and sub-pixel target detection
ISPRS J. Photogramm. Remote Sens.
(2018) - et al.
Classical and emerging non-destructive technologies for safety and quality evaluation of cereals: A review of recent applications
Trends Food Sci. Technol.
(2019) - et al.
Feature selective projection with low-rank embedding and dual Laplacian regularization
IEEE Trans. Knowl. Data Eng.
(2019) - et al.
Unsupervised feature selection via latent representation learning and manifold regularization
Neural Netw.
(2019) - et al.
Hyperspectral band selection via region-aware latent features fusion based clustering
Inf. Fusion
(2022) - et al.
Conceptual and empirical comparison of dimensionality reduction algorithms (PCA, KPCA, LDA, MDS, SVD, LLE, ISOMAP, LE, ICA, t-SNE)
Comput. Sci. Rev.
(2021) - et al.
Machine learning techniques for analysis of hyperspectral images to determine quality of food products: A review
Curr. Res. Food Sci.
(2021)
Deep learning classifiers for hyperspectral imaging: A review
ISPRS J. Photogramm. Remote Sens.
Estimating percentages of fusarium-damaged kernels in hard wheat by near-infrared hyperspectral imaging
J. Cereal Sci.
Maize seed classification using hyperspectral image coupled with multi-linear discriminant analysis
Infrared Phys. Technol.
Advances in data preprocessing for biomedical data fusion: An overview of the methods, challenges, and prospects
Inf. Fusion
X-ModalNet: A semi-supervised deep cross-modal network for classification of remote sensing data
ISPRS J. Photogramm. Remote Sens.
Learning to propagate labels on graphs: An iterative multitask regression framework for semi-supervised hyperspectral dimensionality reduction
ISPRS J. Photogramm. Remote Sens.
Towards resource-frugal deep convolutional neural networks for hyperspectral image segmentation
Microprocess. Microsyst.
Multiple convolutional layers fusion framework for hyperspectral image classification
Neurocomputing
Cascaded dual-scale crossover network for hyperspectral image classification
Knowledge-Based Syst.
Using near-infrared hyperspectral imaging with multiple decision tree methods to delineate black tea quality
Spectrochim. Acta Part A Mol. Biomol. Spectrosc.
HybridCNN based hyperspectral image classification using multiscale spatiospectral features
Infrared Phys. Technol.
Successes and challenges of factor analysis/target transformation application to visible-to-near-infrared hyperspectral data
Icarus
Multidimensional relation learning for hyperspectral image classification
Neurocomputing
Hyperspectral image classification using mixed convolutions and covariance pooling
IEEE Trans. Geosci. Remote Sens.
Wheat blast detection and assessment combining ground-based hyperspectral and satellite based multispectral data
Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci.
Hypersectral imaging for military and security applications: Combining myriad processing and sensing techniques
IEEE Geosci. Remote Sens. Mag.
Quantifying leaf-scale variations in water absorption in lettuce from hyperspectral imagery: a laboratory study with implications for measuring leaf water content in the context of precision agriculture
Precis. Agric.
Using hyperspectral image enhancement method for small size object detection on the sea surface
Hyperspectral imaging in medical applications
Hyperspectral imaging in the medical field: Present and future
Appl. Spectrosc. Rev.
Convolutional neural network based on bandwise-independent convolution and hard thresholding for hyperspectral band selection
IEEE Trans. Cybern.
A similarity-based ranking method for hyperspectral band selection
IEEE Trans. Geosci. Remote Sens.
Deep reinforcement learning for semisupervised hyperspectral band selection
IEEE Trans. Geosci. Remote Sens.
Feature selection and feature extraction in pattern analysis: A literature review
Cited by (4)
A multi-level deformable gated aggregated network for hyperspectral image classification
2023, International Journal of Applied Earth Observation and GeoinformationFHIC: Fast Hyperspectral Image Classification Model Using ETR Dimensionality Reduction and ELU Activation Function
2023, IEEE Transactions on Geoscience and Remote SensingA Multistrategy Evolutionary Multiobjective Optimization Method for Hyperspectral Endmember Extraction
2023, IEEE Transactions on Geoscience and Remote SensingSHBO-based U-Net for image segmentation and FSHBO-enabled DBN for classification using hyperspectral image
2023, Imaging Science Journal
The code (and data) in this article has been certified as Reproducible by Code Ocean: (https://codeocean.com/). More information on the Reproducibility Badge Initiative is available at https://www.elsevier.com/physical-sciences-and-engineering/computer-science/journals.