Hyperspectral Image Classification Using Hybrid Swarm Feature Selection and Ensemble Classifier

Hyperspectral imaging has garnered significant interest for its capability to deliver comprehensive information about the composition and characteristics of objects within an image. Unlike RGB imaging, hyperspectral imaging captures data in narrow spectral bands allowing for a more comprehensive understanding of the scene, while traditional methods often struggle with high-dimensionality and overfitting in hyperspectral image classification, this study introduces a novel hybrid approach combining BPSO and BGWO, which exploits the strengths of each method to optimize feature selection, improving classification accuracy across various datasets.

In the classification process [1], each pixel in the image is assigned to classes based on its unique spectral signature. This signature represents the reflection or absorption pattern that characterizes each object or material [2]. Hyperspectral image classification has applications, including monitoring, land use mapping, agriculture, and mineral exploration [3]. However, hyperspectral data are complex and high-dimensional since they contain numerous narrow bands that were captured during the scanning process [4]. Addressing this problem, researchers have proposed several methods for feature selection to reduce dimensionality of data while improving classification accuracy [5]. The richness of the measured spectrum, with its high dimensionality, proves valuable for pixel classification, enabling the discrimination of different landscapes within the image scene [6]. However, land cover classification in hyperspectral imaging (HSI) data remains a prominent and challenging topic, primarily due to the “curse of dimensionality.” This challenge stems from the labor-intensive process of obtaining a limited number of ground-truth labeled pixels (training data), which must be appropriately distributed across various classes [7]. The commonly low number of collected ground-truth labels as compared to high spectral bands often leads to an unreliable estimate of classifier parameters hence make it prone to problems of over-fitting or under-fitting; thus, reducing classifier performance [8].

To enhance classification accuracy and efficiency, the remote sensing community has prioritized addressing data redundancy, driving the development of advanced feature selection techniques [9]. In data mining and pattern recognition applications, feature selection is essential, particularly for high-dimensional datasets. In order to develop a model, it entails choosing a subset of independent features [10].

The goals of feature selection encompass faster training of classification models, reduced model complexity, improved accuracy, and minimized overfitting. Thus, feature selection serves as a valuable means of providing prior knowledge of hyperspectral imaging and selecting relevant similarity measures with estimated weights for the classification method [11].

Classification plays a vital role in remote sensing by distinguishing objects within imagery into various classes, facilitating the creation of secondary products for mapping, monitoring, and analyzing land cover and land use changes. In hyperspectral image classification, the main goal is to assign each pixel in the scene to a specific category, a critical step in applications such as identifying land-cover types for Earth monitoring [12].

Over the last two decades, numerous supervised methods have been proposed for hyperspectral image classification. In the initial stages, spectral classifiers relied solely on spectral information, with Support Vector Machine (SVM) classifiers being a typical example. However, SVM's low sensitivity to high dimensionality has driven the development of various SVM-based classifiers tailored to handle the spectral classification of hyperspectral imagery [13]. In light of these challenges and advancements in hyperspectral image processing, our study proposes a novel method that combines feature extraction algorithms, specifically Scale-Invariant Feature Transform (SIFT) and Gray-Level Run Length Matrix (GLRLM), with feature selection algorithms, Binary Particle Swarm Optimization (BPSO), and Binary Grey Wolf Optimizer (BGWO). Our strategy aims to identify the ideal feature set that maximizes classification efficiency.

To evaluate the efficacy of our proposed methodology, we employ the Indian Pines dataset, containing hyperspectral images of farmland in Indiana, USA. Comparative analysis with other sophisticated algorithms utilizing GLRLM, SIFT classifiers, and Support Vector Machine (SVM) classifiers demonstrates the superiority of our proposed strategy, yielding notable gains in F1 scores across different categories. To reach this goal, we proposed in this research a new and more reliable technique in the feature selection process that helps in significantly improving the classification of hyperspectral images in various remote sensing and Earth observation applications. The swarm algorithms were hybridized by combining the outputs of two algorithms to improve the binary particle swarm. (BPSO) and Binary Gray Wolf Optimizer (BGWO), which is done through the AND operator, where the results showed a significant improvement in classification accuracy.

The proposed system, illustrated in Figure 1, consists of three main processes: feature-extraction, feature-selection, and classification. This study evaluates the effectiveness of two feature extraction methods (GLRLM and SIFT) alongside a novel feature selection approach that combines two swarming techniques with a collective classifier to enhance classification accuracy. SIFT identifies key points in images invariant to scale and rotation, making it ideal for robust feature extraction across varying spectral bands. Conversely, GLRLM captures texture features by counting the lengths of consecutive runs of pixels at specific gray levels, providing complementary spatial information for classification. The representation is general for most systems used in categorizing hyperspectral images.

1.png

Figure 1. Proposed hyperspectral image classification

For instance, gray-level run length matrix (GLRLM) [23] algorithm is a common method used in computer vision or image analysis for feature extraction while Scale-Invariant Feature Transform (SIFT) [24] also helps achieve this objective. Hyperspectral images (HSI) are classified using these separately as well as in combination [25, 26]. These GLRLM and SIFT are both techniques that can be used for image recognition, where the former is a texture descriptor. By employing both GLRLM and SIFT, different areas of the image will be captured leading to a better representation of the same hence improved classification accuracy. These two algorithms are part of many optimization algorithms used for feature selection in hyperspectral image classification. BPSO on the other hand tries to simulate particles or small organisms moving around, using a simple mathematical model [27]. Conversely, BGWO is an evolutionary algorithm applied to solve mathematical problems [28].

Additionally, this research introduces another model that combines the results obtained from two swarm intelligence algorithms. One dimensional matrix represents sub-features that individual swarms are expected to find using particular algorithms. To ensure only effective features are selected, AND operation is applied between first swarm’s binary vector and second swarm’s vector. The integration between these swarm methods has been indicated by Figure 2 below.

2.png

Figure 2. Proposed hybrid swarms’ structure for feature selection

The AND operation between the binary vector of the first swarm (S1) and the vector of the second swarm (S2) is mathematically represented as:

Final_features $=s_1 \cap s_2$         (1)

where, Final features represent the final selected features.

Additionally, BPSO and BGWO have their respective update equations. In BPSO, the position and velocity of each particle are updated as follows [29].

$v_i^{t+1}=v_i^t \oplus\left(\right.$ pbest $\left._i^t \oplus x_i^t\right)$          (2)

$x_i^{t+1}=x_i^t+v_i^{t+1}$        (3)

where, $pbest_{\boldsymbol{i}}{ }^t$ is the best position that particle I could have found at time t, $x_i^t$ is the position of particle I at time t, $v_i^t$ is the velocity of particle I at time t, represents the bitwise XOR operation and is also used for the bitwise XOR operation between the particle positions and velocities.

In BGWO, the updated equations for the position of each grey wolf are as follows [30]:

$x_i^{t+1}=x_i^t \oplus\left(A \oplus D_i^t\right)$          (4)

$D_i^{t+1}=C \oplus\left(X^{* t} \oplus X_i^{t+1 t}\right)$         (5)

$x_i^{t+1}=clip\left(x_i^{t+1} l b \mathrm{ub}\right)$       (6)

where, Xi(t) represents the position of the grey wolf i at time t, $X^{* t}$ is the best position found by the grey wolves, A is the coefficient matrix, $D_i^t$ is the distance vector of the grey wolf i at time t, C is a random vector, lb and ub are the lower and upper bounds of the search space.

The hybrid approach aims to enhance the feature selection process in hyperspectral image classification, leading to improved accuracy and performance. By combining the strengths of BPSO and BGWO.

This study employed a combination of different machine learning methods to identify hyperspectral imaging (HSI) data. Specifically, a SVM, KNN algorithm, and Naive Bayes classifiers were utilized, as mentioned in the relevant sources [31-33]. By combining these classifiers, an ensemble classifier was created. A group of classifiers, often referred to as basis classifiers, are built using ensemble methods, which are learning algorithms. Predictions are then made on new data points by taking into account the weighted votes from these classifiers, as shown in Figure 3.

3.png

Figure 3. Basic outline of ensemble classifier

A novel approach to hyperspectral image (HSI) classification feature selection that enhances classification efficiency is presented in this study. The researchers have used SIFT and GLRLM techniques for feature extraction and then a hybrid approach was proposed which combined the outputs of two swarm algorithms, BPSO and BGWO, through AND Operator on both scales. According to Indian Pines, KSC, and Botswana datasets, classification results were improved significantly as compared to full spectral data and have consistently raised above 1% threshold. For Botswana, KSC and Indian Pines datasets respectively, The Kappa coefficient as a measure of classification precision achieves outstanding values of 94.47, 93.48, and 81.70. These empirical results prove the usefulness of the proposed feature selection method over other methods tested within this study.

In summary, the hybrid swarm approach combining BPSO and BGWO proved to be a highly effective method for feature selection in hyperspectral image classification. By leveraging the strengths of both algorithms, the method not only improved classification accuracy but also demonstrated robustness across various datasets. This approach holds potential for wide applications in fields such as environmental monitoring, where accurate land cover classification is crucial, and precision agriculture, where early crop health detection can lead to more sustainable farming practices.