Compression and reinforce variation with convolutional neural networks for hyperspectral image classification

Highlights

• Propose hyperspectral image classification model using multi-hybrid deep learning model.

• Propose a feature selection method by increasing the variance and filtering data.

• Apply compression and reinforced variation to enhance data compression and classification.

• Evaluate the proposed approach using different HS images with extensive comparisons.

Abstract

In Hyperspectral images (HSI), dimensionality reduction methods (DRM) play a critical role in reducing the input data dimension and complexity. As much as the deep learning methods (DLM) have presented very aggressive achievements, preprocessing methods and DRM are very important to enhance the learning of DLMs. This study introduces a novel DRM called Compression and Reinforced Variation (CRV), which is used to reduce the input data dimension. The CRV minimizes the gap between the big and small related data in each class and omits the noise and redundant data. It selects the most informative features and normalizes them to enhance data distribution before inserting them into the learning model. The learning model of this study is multi-hybrid deep learning (MHDL) model to improve the extraction of multi-class HSI and spectral–spatial features. MHDL is a novel classification model that includes hybrid layers of conventional neural networks and batch normalization to avoid overfitting, normalizing the training, and extracting the spectral–spatial features for HSI. The proposed CRV provided highly efficient methods for reducing the HSI dimension and improving the classification accuracy of the MHDL model. In contrast to other conventional DRMs, CRV gave the highest accuracy in the shortest time. CRV-MHDL was also compared to seven existing classification models for three distinct datasets, and the findings demonstrated that the CRV-MHDL outperforms all of them by more than 2%. The code of this study is available at this link: https://github.com/DalalAL-Alimi/CRV.

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].