A Review of Space Target Recognition Based on Ensemble Learning

Abstract

The increasing number of space debris and space-active targets makes the space environment more and more complex. Space target recognition, a crucial component of space situational awareness, is of paramount importance to space security. Firstly, this paper elucidates the fundamental principles of ensemble learning, analyzes its characteristics and fusion method, and provides a comprehensive comparison of three common ensemble learning methods. Secondly, this paper analyzes the basic attributes and characteristics of space targets and categorizes the hierarchy of space target recognition. Again, the paper reviews the advances in the application of ensemble learning in space target recognition, focusing on three aspects: space target recognition datasets, the ensemble of traditional machine learning models, and ensemble deep learning. Subsequently, classical machine learning and ensemble learning algorithms are tested on a self-built space target simulation dataset, and we find that Stacking performs well on this dataset. Finally, the paper discusses future research directions.

1. Introduction

With the continuous advancement of space technology, the space environment has become increasingly complex. Space target recognition, as an integral part of space situational awareness (SSA) and on-orbit servicing [1], plays a crucial role in both defense and civil domains. Rapidly and accurately detecting and identifying non-cooperative spacecraft and space debris is essential for planning effective trajectories [2], enabling repairs of in-orbit satellites, and debris removal, thus preventing potential collisions and enhancing the efficiency of orbital resource utilization.

The automatic recognition of space targets relies on machine learning and deep learning methods that utilize existing prior data for inference and prediction through explicit feature extraction and model construction. Traditional machine learning models require fewer computational resources but are limited in their generalization ability and robustness, especially when handling complex, high-dimensional data, leading to restricted recognition accuracy. Deep learning models alleviate the reliance on feature engineering by facilitating end-to-end learning, enabling direct prediction from raw data. Nevertheless, the substantial parameter size of deep neural networks leads to increased inference times [3]. The recognition accuracy and speed of individual machine learning or deep learning are constrained by the scale of data and the number of model parameters.

Ensemble learning (EL), as an integral part of machine learning, enhances the robustness and accuracy of systems by constructing and combining the predictions of multiple basic learners [4]. It effectively overcomes the drawbacks of traditional machine learning and deep learning approaches, being applicable to both supervised and unsupervised learning [5]. In the task of space target recognition, ensemble learning demonstrates notable advantages. A former president of the Association for the Advancement of Artificial Intelligence (AAAI), Thomas G. Dietterich, categorized ensemble learning alongside scalable machine learning, reinforcement learning, and probabilistic networks as the four major research directions in machine learning [6]. Additionally, the effectiveness of ensemble learning has been analyzed from statistical, computational, and representational perspectives [7]. Moreover, theoretical foundations such as bias–variance–covariance decomposition, boundary theory [8], strong correlation [9], and random discrimination [10] further support the effectiveness of ensemble learning [11].

The early theoretical exploration of ensemble learning dates back to 1979, when Dasarathy first proposed the concept of combining multiple classifiers into a composite classifier system [12]. During this initial phase, research primarily focused on the fundamental principles and theoretical framework of ensemble learning, laying a solid foundation for subsequent studies. In 1990, Hansen and Salamon [13] demonstrated that combining multiple neural network models could reduce the variance of classifiers. Following this, in 1991, Jacobs [14] introduced the concept of a mixture of experts by integrating different expert models.

Subsequently, numerous classic ensemble algorithms were proposed, gradually establishing the basic procedures and framework of ensemble learning. In 1992, Wolpert [15] introduced the Stacked Generalization (Stacking) model, which enhances model performance by learning from the predictions of basic models. In 1996, Freund et al. [16] proposed the AdaBoost algorithm, which combines multiple weak classifiers into a strong classifier by dynamically adjusting the weights of the classifiers. In the same year, Breiman [9] introduced the Bagging (Bootstrap Aggregating) algorithm, which generates multiple training datasets through bootstrap sampling and trains corresponding basic learners. In 1997, Dagging, a variant of Bagging, was proposed; this generates non-overlapping datasets through random sampling [17]. Besides data-level operations, researchers have also innovated at the feature level by using sampled features as input to generate different basic learners [18]. Entering the 21st century, Breiman [19] proposed the Random Forest algorithm, marking a significant breakthrough in the field of ensemble learning. Subsequently, Rodriguez et al. [20] introduced the Rotation Forest algorithm, which randomly splits the training features into several non-overlapping subsets, applies Principal Component Analysis (PCA) to each subset, and uses axis rotation to form new features for training basic learners.

Modern integration frameworks are more diverse. In recent years, the combination of ensemble learning with algorithms such as deep learning [21], semi-supervised learning [22], and evolutionary algorithms [23] has emerged as a new research hotspot. The widespread availability of big data and the enhanced accessibility of high-performance computing resources have catalyzed the emergence of Ensemble Deep Learning (EDL) [21]. Fast Ensemble Deep Learning (FEDL), also known as snapshot, utilizes the non-convex characteristics of deep neural networks and the ability of Stochastic Gradient Descent (SGD) to escape local minima, thereby reducing the time overhead of training multiple deep learners in Ensemble Deep Learning [24]. The concepts of semi-supervised learning and ensemble learning emerged almost simultaneously. Combining the two [22] can address the issue of insufficient labeled samples in ensemble learning and reduce the generalization error of semi-supervised learning [25]. The conflict between the diversity of basic learners and the accuracy or robustness of solutions is a typical optimization problem. Evolutionary algorithms, characterized by their wide applicability, robustness, and ability to achieve global optimization [23], have been combined with ensemble learning by researchers to propose Evolutionary Ensemble Learning (EEL) algorithms. Consequently, the development of ensemble learning can be summarized into three stages, as illustrated in Figure 1.

The structure of this paper is as follows. In Section 1, we discuss the fundamentals of ensemble learning. Section 2 introduces the relevant aspects of space target recognition while Section 3 provides a comprehensive review of the application advances of ensemble learning in target recognition, covering dataset construction, traditional machine learning ensembles, and Ensemble Deep Learning approaches. Section 4 presents experiments on space target recognition using simulated datasets, and Section 5 concludes the paper with a summary and future directions.

2. Basic Principles of Ensemble Learning

2.1. Diversity in Ensemble Learning

A necessary and sufficient condition for an ensemble learning system to outperform any individual basic learner is that the classifiers must be both accurate and diverse [13]. Thus, the key to designing an effective ensemble learning system lies in selecting high-quality and diverse basic learners.

• Diversity in ensemble learning
  Diversity measurement methods primarily rely on the statistical analysis of the results. These methods can be categorized into pairwise diversity measures and non-pairwise diversity measures [26]. Pairwise diversity measures evaluate the diversity between pairs of basic learners and then analyze the overall diversity of the ensemble system using their averages. Non-pairwise diversity measures directly compute diversity metrics for the ensemble system. Theoretically, there is no clear correlation between the diversity of basic learners and the accuracy of the ensemble system [27]. Further theoretical research is needed to better define diversity and explore its relationship with ensemble accuracy. A summary of diversity measurement metrics in ensemble learning is presented in

  Table 1. Metrics for diversity in ensemble learning.

  Categories	Metrics	Description	Equation	Range	Value at Maximum Diversity
  Pairwise Diversity Measures	Q [28]	Classification consistency	$Q_{ij}={\displaystyle \frac{N^{11}{N}^{00}-N^{10}{N}^{01}}{N^{11}{N}^{00}+N^{10}{N}^{01}}}$	[−1, 1]	−1
  	$\rho$ [29]	Classification consistency	${\rho }_{ij}=$${\displaystyle \frac{N^{11}{N}^{00}-N^{10}{N}^{01}}{\sqrt{\left(N^{11}+N^{10}\right)\left(N^{01}+N^{00}\right)\left(N^{11}+N^{01}\right)\left(N^{10}+N^{00}\right)}}}$	[−1, 1]	−1
  	diss	Disagreement count	${dis}_{ij}={\displaystyle \frac{N^{10}+N^{01}}{N}}$	[0, 1]	1
  	DF	Joint error rate	${DF}_{ij}={\displaystyle \frac{N^{00}}{N}}$	[0, 1]	0
  Non-pairwise Diversity Measures	KW [29]	Label Variance	$KW={\displaystyle \frac{\sum _{j=1}^{N}l\left(z_j\right)\left(M-l\left(z_j\right)\right)}{N{M}^2}}$	[0, 1]	1/4
  	Kappa [30]	Inter-classifier consistency	$\kappa =1-{\displaystyle \frac{\frac{1}{M}\sum _{j=1}^{N}l\left(z_{j0}\right)\left(M-l\left(z_j\right)\right)}{N\left(M-1\right)\stackrel{̿}{p}\left(1-\overline{p}\right)}}$	[−1, 1]	−1
  	E	Degree of consensus or disagreement	$E={\displaystyle \frac{1}{N}}{\displaystyle \sum _{j=1}^N}{\displaystyle \frac{1}{\left(M-⌈\frac{M}{2}⌉\right)}}min\left\{l\left(z_j\right),M-l\left(z_j\right)\right\}$	[0, 1]	1
  	$\theta$	Variance in proportion of correct classification	$\theta =Var\left(X\right)$	[0, 1]	0
  	GD [31]	Classification consistency	——	0 or 1	1
  	PCDM [32]	Correct classification proportion	——	——	——

  .
  The following symbols are defined: M denotes the number of the basic learner; $n_i$ and $n_j$ (i, j = 1,2,…, N, i ≠ j) represent two basic learners. $N^{11}$ is the number of samples correctly classified by both $n_i$ and $n_j$; $N^{10}$ is the number of samples correctly classified by $n_i$ but incorrectly classified by $n_j$; $N^{01}$ is the number of samples incorrectly classified by $n_i$ but correctly classified by $n_j$; $N^{00}$ is the number of samples incorrectly classified by both $n_i$ and $n_j$. The total number of samples N is given by $N=N^{11}+N^{10}+N^{01}+N^{00}$. $z_j$ (j = 1,2, …, N) is the $j^{th}$ sample out of N samples; $r\left(z_j\right)$ is the number of basic learners that correctly classify the $z_j$.

2.

Methods for enhancing diversity

Enhancing diversity in ensemble systems typically involves addressing three primary dimensions: data, algorithmic parameter, and model architecture.

At the data level, the primary methods for enhancing diversity include perturbations to input samples and features. Common techniques for perturbing input samples include resampling, sequential sampling, and hybrid sampling methods. Resampling methods generate different training subsets while sequential sampling methods sample based on the results of the previous learning iteration. Hybrid sampling methods [33] improve model diversity and performance by integrating global and local data characteristics. Perturbing input features involves randomly selecting subsets from the initial attribute set to train diverse basic learners. Rokach et al. [34] proposed a general framework for feature set splitting to optimize decision tree models. For time-series data, Yang et al. [35] introduced a splitting ensemble method based on Hidden Markov models to address issues of model initialization and selection. Additionally, perturbing output representations is another important technique. Methods such as randomizing outputs [36] or using Error-Correcting Output Codes (ECOCs) [37] can enhance predictive accuracy.

At the algorithmic parameter level, diversity among basic learners is enhanced by modifying their parameter sets to generate distinct models. In the C4.5 decision tree algorithm, adjusting the confidence factor can generate different decision boundaries [38]. Similarly, multi-core learning generates diversity by tuning the parameters and their combination weights. Nen M et al. [39] explored how different parameter configurations in multi-kernel learning algorithms can produce basic learners with diverse characteristics. Additionally, in neural networks, altering the number of nodes and the network topology can generate different basic learners.

At the model structure level, diversity can be enhanced by adjusting the internal or external structures of basic learners. For example, heterogeneous ensembles introduce different types of models to improve both the diversity and performance of the ensemble system. The Mixture of Experts (MoE) model employs a ‘gating network’ to determine the weights of multiple expert models, combining their outputs through a weighted sum based on these weights [14]. Additionally, Wang et al. [40] found that incorporating a moderate number of decision trees into neural network ensemble systems can further enhance their overall performance.

Selecting appropriate basic learners is critical in ensemble learning as they must be capable of capturing the complexity of the data without overfitting. Furthermore, computational efficiency, stability, and complementarity are key factors to consider in the selection process. Theoretical support and empirical evaluations can help determine the most suitable basic learners. Additionally, the choice of basic learners should be optimized based on the specific problem, data characteristics, and available resources.

2.2. Fusion Method

The fusion of basic learners is the final step in constructing an ensemble system [41]. Currently, the primary fusion methods include voting, averaging, and learning-based approaches. Examples of some fusion methods are shown in Figure 2.

• Voting method
  Voting methods are primarily used for classification tasks and can be categorized into majority voting, plurality voting, and weighted voting. Majority voting selects a class as the final prediction if it receives more than half of the total votes from the basic learners. Plurality voting, on the other hand, chooses the class with the highest number of votes as the final result. In cases where multiple classes receive the same highest number of votes, a random selection is made. Weighted voting takes into account the predictive performance of each basic learner, and the voting weight of each basic learner is positively correlated with its accuracy.
• Averaging method
  Averaging methods are commonly used for regression problems to combine the outputs of basic learners. Simple averaging involves taking the mean of all predictions from the basic learners. In practice, weighted averaging is often more effective; the weights of the basic learner can be dynamically adjusted during the training process to optimally combine the outputs of different basic learners and improve the overall predictive accuracy.
• Meta-learning method
  Meta-learning methods, exemplified by stacked generalization, involve using the predictions of basic learners as new features for a secondary learner (meta-learner). This meta-learner learns from the basic learners’ predictions to uncover latent relationships among them, thereby enhancing the model’s generalization ability and predictive accuracy.

In summary, the fundamental process for constructing ensemble systems can be categorized into data preparation, the generation of training sets, the determination of input attributes, the selection of basic learners, the establishment of combination strategies, model evaluation, etc. This basic process is shown in Figure 3.

2.3. Common Ensemble Methods

2.3.1. Bagging

Bagging, proposed by Breiman [9], generates multiple sub-datasets through bootstrap sampling with replacement. Each base learner is trained on a subset, and their predictions are aggregated via voting or averaging. Bagging reduces variance and prevents overfitting, making it particularly effective for high-variance models like decision trees. Variants such as the enhanced version proposed by Tuysuzoglu et al. [42], OverBagging [43], and Neighborhood Balanced Bagging [44] address challenges in imbalanced datasets.

Random Forest, a popular implementation of Bagging, further enhances performance by randomly selecting features before training each tree. Rotation Forest [20] enhances Random Forest by incorporating PCA while combining Random Forest with Fisher Discriminant Analysis and stratified sampling [45] has been shown to address classification challenges in high-dimensional datasets. The process of Bagging is illustrated in Figure 4a.

2.3.2. Boosting

Boosting transforms weak learners into strong learners by iteratively adjusting sample weights to minimize the error function [7]. Unlike Bagging, Boosting employs sequential training, progressively improving classification accuracy. The basic process of Boosting is illustrated in Figure 4b.

AdaBoost, a classic binary classification algorithm, has been extended to multi-class problems (e.g., AdaBoost.M1 and AdaBoost.M2). Gradient Boosting [46] optimizes models incrementally by minimizing the negative gradient of the loss function with efficient implementations like LightGBM [47] and XGBoost [48]. Stochastic Gradient Boosting, proposed by Friedman [49], improves generalization by randomly selecting feature subsets in each iteration. CatBoost [50], akin to XGBoost and LightGBM, employs a weighted cross-entropy loss function to effectively address class imbalance. Variants such as AdaCost [51], BABoost [52], and RareBoost [53] adjust sample weights while SMOTEBoost [54], RUSBoost [55], and PCBoost [56] integrate data synthesis methods to improve performance on imbalanced data.

2.3.3. Stacking

Stacking combines predictions from diverse basic learners as inputs for a meta-learner, which optimizes the final prediction. Unlike Bagging and Boosting, Stacking employs a hierarchical training approach, transforming basic model predictions into a new dataset for meta-learner training. This method excels in complex tasks by leveraging the strengths of multiple models. However, its performance heavily depends on the selection of basic and meta-models, which can be optimized using techniques like Ant Colony Optimization [57] or Genetic Algorithms [58]. The fundamental process of Stacking is illustrated in Figure 4c.

In conclusion, various ensemble learning algorithms exhibit distinct structures and characteristics. In practical applications, it is essential to evaluate and compare these algorithms based on the specific features of the data and the objectives of the task. A comparison of common ensemble learning algorithms is provided in

Table 2. Comparison of common algorithms for ensemble learning.

Method	Core Idea	Training Method	Unique Advantages	Limitations	Application
Bagging	Generates multiple sub-datasets via bootstrap sampling; trains basic learners in parallel.	Parallel training	Reduces variance, prevents overfitting, and is suitable for high-variance models.	Less effective for high-bias models; higher computational cost.	[59]: Improving the performance of genetic programming. (https://github.com/marcovirgolin/2SEGP, accessed on 25 March 2025.) [60]: Object detection in complex traffic scenes.
Boosting	Iteratively adjusts sample weights to transform weak learners into strong learners.	Sequential training	Reduces bias and improves accuracy; effective for imbalanced data.	Prone to overfitting; sensitive to noisy data.	[61]: Genomic variant classifier. (https://github.com/djparente/polyboost, accessed on 25 March 2025.) [62]: Image quality assessment. (https://github.com/Dounia18/EGB, accessed on 25 March 2025.)
Stacking	Combines predictions from multiple base learners as new features to train a meta-learner.	Hierarchical training	Enhances generalization by leveraging diverse base learners; suitable for complex tasks.	High computational cost; complex model selection; risk of overfitting.	[63]: Prediction of hypertension. [64]: Splice site prediction. (https://github.com/OluwadareLab/EnsembleSplice, accessed on 25 March 2025.)

.

3. Space Target Recognition

Space targets primarily refer to artificial objects such as satellites, space stations, and protective shields and also include various space debris and cosmic objects entering the Earth’s outer space. Space target recognition involves the detection and tracking of these targets, extracting their feature information, and subsequently achieving the classification and comprehensive understanding of various space targets [65]. The overall block diagram of space target recognition methods is illustrated in Figure 5.

3.1. Basic Attributes and Characteristics of Space Targets

3.1.1. Basic Attributes of Space Targets

Space targets possess unique fundamental attributes that determine their specific operational mechanisms and functionalities. These attributes also provide multi-dimensional information sources for space target recognition and lay the foundation for the application of ensemble learning methods.

The orbital attributes of space targets refer to their path characteristics as they move through space. These characteristics are crucial for understanding the movement patterns of space targets, predicting their positions, and planning related space missions. Based on the distribution of their orbits, space targets can be categorized into Low Earth Orbit (LEO), High Earth Orbit (HEO), Geostationary Earth Orbit (GEO), and Critical Earth Orbit (CEO) targets. In addition to orbit distribution, orbital attributes also include whether a target is controllable as uncontrolled targets gradually deviate from their original trajectories. The attitude attributes of space targets primarily involve their orientation and stability in space. Precision in attitude control not only enhances communication and scientific observation but also optimizes the accuracy of navigation imaging, ensuring the reliable operation of spacecraft in complex environments. Material attributes of space targets are also part of their fundamental characteristics. Space targets such as satellites are typically constructed from durable materials designed to withstand extreme conditions in space including extreme temperature variations, radiation, and micrometeoroid impacts. Different targets often use different materials, resulting in distinct spectral curves that can be used for space target recognition. Geometric attributes refer to the shapes and sizes of space targets. For example, satellites are generally composed of components such as solar panels, sensors, communication equipment, and propulsion systems. Depending on their intended use and adopted attitudes, their composition and shape may vary accordingly.

In addition to the aforementioned attributes, the fundamental attributes of space targets also include photometric and radiometric properties. The recognition of space targets involves recognizing these attributes. Identifying different attributes of satellites typically requires processing multi-modal data. Ensemble learning methods, by combining the predictions of multiple models, can better adapt to target recognition under various environmental conditions, thereby enhancing the overall performance of a system.

3.1.2. Characteristics of Space Targets

Based on the aforementioned analysis of the attributes of space targets, the problem of space target recognition has the following characteristics compared to the recognition of conventional targets such as aircraft and ships.

• Limited sample size: Space targets move at high speeds in space, and the imaging conditions are relatively stringent. As a result, both space-based and ground-based equipment can capture only a limited number of images.
• Poor image quality: Distant space targets often appear small and dim. When atmospheric interference or unstable cloud conditions are present, space targets can easily be obscured by background noise.
• Diverse targets: Space contains a wide variety of targets including satellites, space debris, and planets. Each type of target has distinct shapes, sizes, materials, and motion attributes.
• Dynamic changes: Space targets are typically in a state of dynamic change, such as in the orbital variations of satellites and the drift of space debris.

3.2. Hierarchy of Space Target Recognition

Space target recognition is a complex multi-level task. By hierarchically categorizing the recognition process, the workflow can be more systematically organized and optimized, enabling the selection of appropriate algorithms and techniques for different levels of tasks. The hierarchy is illustrated in Figure 6.

The low-level feature extraction layer is the foundation of the recognition process, being primarily responsible for extracting fundamental physical and geometric features from raw observational data. These features typically represent the most basic attributes of the target such as shape, edge, texture, and spectral characteristics. The extraction of low-level features marks the starting point of the entire recognition process and directly influences the accuracy of subsequent high-level feature extraction and semantic understanding. At this level, the key tasks include acquiring preliminary position information, extracting geometric features, identifying spectral characteristics, and analyzing motion attributes.

The middle-level feature selection layer is primarily tasked with feature fusion and selection, combining the multiple features obtained from the low-level feature extraction layer to create a more representative and robust feature representation. Feature fusion can be achieved through various methods such as rule-based combination or deep learning-based feature fusion networks. Additionally, this layer reduces the dimensionality of features using dimensionality reduction techniques, thereby decreasing computational complexity, enhancing processing efficiency, and improving the expression capability and interference resistance of features. This provides higher-quality feature inputs for high-level semantic understanding.

The high-level semantic understanding layer represents the final stage of the recognition process, being primarily responsible for transforming the fused high-level features from the middle layer into concrete semantic information. This layer involves tasks such as target classification and recognition, target component segmentation, target tracking and state estimation, and target behavior analysis. Target classification and recognition map the extracted high-level features to specific category labels such as satellites, space debris, or planets. Target component segmentation involves decomposing the visual or physical representation of a space target into distinct components or substructures. Target tracking and state estimation continuously monitor the motion state of the target, predicting its future trajectory and position. Target behavior analysis identifies specific behavior patterns and intentions of the target such as anomaly detection and interactions between targets.

4. Advances in Application of Ensemble Learning in Space Target Recognition

4.1. Space Target Recognition Dataset

Due to the high cost and difficulty of collecting space data, as well as the particularities of space targets and the sensitivity of space regions, it is challenging to obtain true visual information. Consequently, most space target datasets are generated through simulated or synthetic means [66]. This section enumerates seven datasets applicable to space target recognition and provides a summary in

Table 3. Space target recognition, classification, and pose estimation dataset.

Dataset	Year	Image Type	Resolution	Target	Category Count	Sample Count	Application	Disadvantage	Links
BUAA-SID 1.0 [67]	2010	Synthetic	320 × 240	Satellite	20	9200	Space target classification and identification	Limited number of satellites; synthetic images need improved realism	—
URSO [68]	2019	Synthetic	1280 × 960	Satellite	2	5000	Spacecraft attitude estimation	No segmentation annotations provided	https://zenodo.org/records/3279632 (Accessed on 25 March 2025)
SPEED [69]	2019	Synthetic & Real	1920 × 1200	Spacecraft	1	15,300	Non-cooperative satellite attitude estimation	Reliance on synthetic images for training, differing from real-space images	https://zenodo.org/records/6327547 (Accessed on 25 March 2025)
A Spacecraft Dataset [70]	2021	Synthetic & Real	1280 × 720	Satellite	—	3117	Spacecraft target detection, segmentation, and partial recognition	Limited dataset size; fixed image resolution	https://github.com/Yurushia1998/SatelliteDataset (Accessed on 25 March 2025)
SPEED+ [71]	2021	Synthetic & Real	1920 × 1200	Spacecraft	1	69,531	Visual spacecraft attitude estimation and relative navigation	Contains only single-target images	https://kelvins.esa.int/pose-estimation-2021/ (Accessed on 25 March 2025)
Perez [72]	2021	Synthetic	70,960	Satellite, Debris	5	—	Space debris detection and tracking; Satellite classification;	Single background; low variability in size and orientation	—
SPARK [73]	2021	Synthetic	—	Satellite, Debris	11	~150,000	Space target recognition;	Need for periodic updates due to increasing new space target types	https://cvi2.uni.lu/spark-2021/ (Accessed on 25 March 2025)

. The examples of images from the dataset are shown in Figure 7.

• BUAA-SID [67]
  The BUAA-SID dataset is a space target image database established by Beihang University (BUAA). The 3D models of space targets were created using 3ds Max and visible light simulation images were then generated. This dataset contains a large number of space target images based on real scenes, making it suitable for training and testing various space target recognition algorithms. However, the dataset does not include 3D models of the space targets. As the database was developed to meet specific project requirements, its content is not fully publicly available. The dataset includes images of 20 representative satellites, with 460 images generated for each satellite, totaling 9200 images. These images cover 230 sampling viewpoints.
• URSO [68]
  URSO (Unreal Rendered Spacecraft On-Orbit) is a space mission simulator designed to provide realistic images and depth masks of commonly used spacecraft orbiting the Earth. Its purpose is to create a visually realistic environment with labeled data for tasks such as space rendezvous and docking, as well as for debris removal. The dataset was constructed using Unreal Engine 4 (UE4) as the simulation engine, and the spacecraft models include Soyuz and Dragon. At an altitude within Low Earth Orbit, 5000 viewpoints were randomly sampled across the illuminated surface of the Earth. For each viewpoint, interface images and depth maps were generated, with image resolution set at 1080 × 960 pixels. The URSO dataset comprises 5000 images, which are intended for the training and evaluation of attitude estimation algorithms.
• SPEED [69]
  Deep learning relies on large annotated datasets, but obtaining ideal images of target spacecraft in space along with precise pose labels is challenging. To address these issues, the Space Laboratory (SLAB) at Stanford University and the Advanced Concepts Team (ACT) at the European Space Agency (ESA) organized the Satellite Pose Estimation Challenge (SPEC). The dataset for this challenge, named SPEED (Satellite Pose Estimation Dataset), is the first publicly available dataset for spacecraft pose estimation. It comprises 14,998 synthetic images and 300 real images, all of which are high-resolution grayscale images. This dataset is intended for training and evaluating deep learning models in space-related tasks.
• A Spacecraft Dataset [70]
  Dung et al. [70] introduced a spacecraft dataset consisting of both real and synthetic images, designed for tasks such as spacecraft detection, instance segmentation, and part recognition. In constructing the dataset, the spacecraft were decomposed into three components: solar panels, the main body, and antennas. Preprocessing was conducted to remove similar or duplicate images, followed by training a segmentation model to predict initial masks, which were then refined using Polygon-RNN++. The dataset consists of 3117 satellite and space station images captured from space, along with annotations for object bounding boxes, instance masks, and part masks. All images are of the resolution 1280 × 720.
• SPEED+ [71]
  While the SPEED dataset primarily consists of synthetic images, these images still have a gap compared to real space target images. The SPEED+ dataset is an extended and improved version of the original SPEED dataset, containing 60,000 synthetic images for training and 9531 spacecraft model simulation images. This dataset offers a sufficient quantity and quality of spacecraft images to evaluate and compare the robustness of on-board models.
• Perez [72]
  To explore the feasibility of using machine learning and deep learning methods for the identification and classification of satellites or space debris, Perez et al. [72] established an experimental setup to generate a series of image datasets for spacecraft and space debris. The dataset consists of both experimental and synthetic images. The experimental dataset comprises 60,460 experimental images generated in the visible and thermal infrared ranges for space target recognition and classification. These images include three types of satellites (Calypso, Cloudsat, and Jason-3) and one type of debris. The synthetic image dataset was created using a graphics engine combined with 3D CAD models. It contains three types of satellites (Calypso, Cloudsat, and Jason-3) and two different debris objects (treated as one category). The relevant parameters of the synthetic images can be adjusted as needed, enabling the generation of unlimited data. Additionally, extra imaging techniques such as blurring and saturation changes can be added during the generation process.
• SPARK [73]
  The University of Luxembourg has introduced SPARK (Spacecraft Recognition Leveraging Knowledge of Space Environment), a multi-modal image dataset for space targets. This dataset is generated in a real space-simulated environment, with each image containing target annotations in bounding boxes. It includes ten different satellites and five space debris objects (treated as one class), totaling 11 target categories. Each satellite has approximately 12,500 images while each debris object has 5000 images, resulting in roughly 150,000 RGB and depth images altogether.

Due to the particularity of space targets, constructing a space target recognition dataset poses numerous challenges. Space data collection is both costly and arduous. Owing to the complexity of the space environment and adverse imaging conditions, real or synthetic space target images often suffer from issues such as high background noise, small target sizes, and low brightness. Moreover, sub-optimal observation angles during data acquisition can obscure the satellite’s geometric features, thereby impacting recognition accuracy. Additionally, in the target annotation and segmentation phase, inadequately refined algorithms or preprocessing techniques may result in rough target masks that fail to precisely represent the target’s true shape and boundaries. Figure 8 exemplifies the challenges faced by space target images.

4.2. Ensemble of Traditional Machine Learning Models for Space Target Recognition

Traditional machine learning methods have been applied to tasks such as object classification, detection, and data selection. Techniques such as those of artificial neural networks, decision trees, and support vector machines (SVMs) have been used for space debris detection [74], galaxy morphological classification [75,76], and asteroid identification [77]. The ensemble of traditional machine learning methods, known for their accuracy and robustness, have increasingly been applied to the field of object recognition. Research in this area primarily focuses on basic learner selection, feature extraction, handling imbalanced data, and algorithm improvement. A summary of the current research status is presented in Table 4.

Selecting a subset of basic learners for an ensemble may perform better than using all basic learners [78]. Ensemble selection can be categorized into static ensemble selection (SSE) and dynamic ensemble selection (DSE). In SSE, the subset of basic learners selected is used for classification predictions of all test samples. The minimization of the margin distance [79] is often used as the basis for the selection of basic learners while also considering metrics such as basic learner performance [80], classification confidence [81], or confusion matrices [82] or employing meta-heuristic algorithms for selecting subsets of basic learners [83]. Compared to SSE, DSE selects specific basic learners for each test instance, potentially leading to higher accuracy. Fan et al. [84] combined K-medoids clustering with random reference classifiers (RRCs) to select the optimal combination of basic learners.

Appropriate input features can significantly enhance a model’s classification performance by removing irrelevant or redundant features, preventing excessively high feature dimensionality, and speeding up the model’s training process. There are two main approaches that combine feature selection with ensemble learning. The first involves using ensemble learning for feature selection, generating effective feature subsets. For instance, Gu et al. [85] employed a random sampling method to select two-dimensional convolutional features extracted by each convolutional kernel, achieving dimension reduction while obtaining effective feature combinations. The second approach leverages feature selection to construct diverse feature subsets for ensemble learning [86]. Li et al. [87] combined geometric features with wavelet moments, adjusting the weight distribution to improve AdaBoost for space target recognition. Xu et al. [88] integrated multi-feature extraction methods with XGBoost. Traditional models takes signals or image features extracted by sensors as inputs, often facing the challenge of high input feature dimensionality. Perez et al. [72], when detecting and classifying satellites and space debris, utilized PCA and cubic feature mapping to reduce feature dimensionality.

For class imbalance (CI) datasets, traditional machine learning algorithms tend to focus excessively on the majority class, leading to a decline in overall classification accuracy. Ensemble learning addresses class imbalance at algorithmic or data levels to enhance model performance. At the algorithmic level, cost factors are introduced to construct cost-sensitive algorithms. For instance, AdaBoost increases the weights of misclassified samples in each iteration, improving overall classification performance. At the data level, sampling techniques are employed to balance the dataset. SMOTE (Synthetic Minority Oversampling Technique) is commonly used to handle class imbalance by generating synthetic samples of the minority class. SMOTEBoost [54] combines SMOTE with Boosting, generating synthetic samples for the minority class in each iteration, thereby indirectly adjusting sample weights. Unlike traditional Boosting, SMOTEBoost not only focuses on misclassified samples but also increases the number of minority class samples by generating synthetic ones. Based on SMOTEBoost, researchers have developed RUSBoost [89], which is based on random undersampling; EUSBoost [90], which combines random undersampling with evolutionary undersampling; and PCBoost [56], which uses data synthesis methods. In addition to improving Boosting algorithms, Xiao et al. [91] enhanced the Bagging algorithm using K-means-based undersampling while Gao et al. [92] employed the SMOTE-Tomek hybrid sampling technique and Random Forest to tackle the issues of manual threshold setting and class imbalance in space satellite telemetry data.

In enhancing model accuracy and algorithm improvements, Tomasz et al. [93] trained a linear classifier for each sample and then integrated them, a method known as Exemplar-SVMs. This ensemble approach provides more exemplar metadata, which makes it useful for object recognition and scene understanding. Shi et al. [94] utilized Extreme Learning Machines (ELMs) for high-precision single-step prediction and corrected the prediction results from the time dimension and finally formulated the interpretation strategy based on the Bagging algorithm for different target categories. Zhou et al. [95] chose the C4.5 algorithm as the basic learner to achieve multi-source information fusion and maritime target recognition. Wang et al. [96] introduced the Maximum Information Coefficient (MIC) and Bootstrap sampling to generate multiple Bayesian networks. They used ensemble learning strategies to compute edge weights and iteratively optimize the network structure.

In the radar automatic target recognition problem, ensemble learning algorithms such as XGBoost can be employed to enhance the performance of radar systems [97,98].

Table 4. Overview of applications in ensemble of traditional machine learning methods.

Application	Reference	Year	Model	Dataset
Basic Learner Selection	[79]: SSE algorithm minimizing interval distance.	2008	Bagging	UCI
	[80]: Interval definition method considering the performance of basic learner.	2014	Bagging	UCI
	[81]: Interval definition method considering classification confidence.	2014	EP-CC	UCI
	[82]: Subspace partitioning method for basic learners based on confusion matrix.	2014	Bagging	UCI, Statlog
	[83]: Interval definition method integrating basic learner and sample weights.	2017	SA, SSE	Custom HRRP Dataset, UCI
	[84]: Selective ensemble technique based on K-medoids clustering and random reference classifiers.	2019	K-medoids, etc.	Custom HRRP Dataset, UCI
Feature Extraction	[85]: Feature selection based on two-dimensional convolutional kernel and random sampling.	2018	Bagging	MSTAR
	[87]: Feature extraction method combing geometric feature and wavelet moments.	2015	Boosting	Custom Space Target Dataset
	[88]: Small target detection method combing multi-feature extraction and XGBoost.	2023	Boosting	Measure Sea Clutter Data
	[72]: Feature dimensionality reduction based on PCA.	2021	SVM, PCA, DNN, etc.	Synthetic Space Target Dataset
Imbalance Data	[54]: SMOTEBoost based on synthetic minority class.	2003	Boosting	KDD Cup-99, etc.
	[89]: RUSBoost based on random undersampling.	2010	Boosting	SP3, etc.
	[90]: EusBoost based on random undersampling and evolutionary undersampling.	2013	Boosting	KEEL
	[56]: PCBoost based on data synthesis methods.	2012	Boosting	UCI
	[91]: Undersampling method based on K-means.	2021	Bagging	Glass5 Shuttle0vs4
	[92]: Recognition method for space satellite operation modes based on SMOTE-Tomek hybrid sampling and Random Forest.	2023	Bagging	Orbiting scientific satellite telemetry parameters
Algorithm Improvement	[93]: Ensemble of Exemplar-SVMs.	2011	Exemplar-SVM	PASCAL VOC
	[94]: Data classification strategy based on Bagging.	2018	Bagging	Simulated data, real satellite telemetry data
	[95]: Maritime target recognition based on multi-source information fusion.	2021	Bagging	Hypothetical case examples
	[96]: Bayesian network structure optimization strategy based on ensemble learning.	2021	Bagging	ASIA ALARM
	[97]: Radar target recognition based on XGBoost.	2023	Boosting	Custom radar dataset
	[98]: Small target detection based on XGBoost.	2023	Boosting	IPIX

Table 4. Overview of applications in ensemble of traditional machine learning methods.

Application	Reference	Year	Model	Dataset
Basic Learner Selection	[79]: SSE algorithm minimizing interval distance.	2008	Bagging	UCI
	[80]: Interval definition method considering the performance of basic learner.	2014	Bagging	UCI
	[81]: Interval definition method considering classification confidence.	2014	EP-CC	UCI
	[82]: Subspace partitioning method for basic learners based on confusion matrix.	2014	Bagging	UCI, Statlog
	[83]: Interval definition method integrating basic learner and sample weights.	2017	SA, SSE	Custom HRRP Dataset, UCI
	[84]: Selective ensemble technique based on K-medoids clustering and random reference classifiers.	2019	K-medoids, etc.	Custom HRRP Dataset, UCI
Feature Extraction	[85]: Feature selection based on two-dimensional convolutional kernel and random sampling.	2018	Bagging	MSTAR
	[87]: Feature extraction method combing geometric feature and wavelet moments.	2015	Boosting	Custom Space Target Dataset
	[88]: Small target detection method combing multi-feature extraction and XGBoost.	2023	Boosting	Measure Sea Clutter Data
	[72]: Feature dimensionality reduction based on PCA.	2021	SVM, PCA, DNN, etc.	Synthetic Space Target Dataset
Imbalance Data	[54]: SMOTEBoost based on synthetic minority class.	2003	Boosting	KDD Cup-99, etc.
	[89]: RUSBoost based on random undersampling.	2010	Boosting	SP3, etc.
	[90]: EusBoost based on random undersampling and evolutionary undersampling.	2013	Boosting	KEEL
	[56]: PCBoost based on data synthesis methods.	2012	Boosting	UCI
	[91]: Undersampling method based on K-means.	2021	Bagging	Glass5 Shuttle0vs4
	[92]: Recognition method for space satellite operation modes based on SMOTE-Tomek hybrid sampling and Random Forest.	2023	Bagging	Orbiting scientific satellite telemetry parameters
Algorithm Improvement	[93]: Ensemble of Exemplar-SVMs.	2011	Exemplar-SVM	PASCAL VOC
	[94]: Data classification strategy based on Bagging.	2018	Bagging	Simulated data, real satellite telemetry data
	[95]: Maritime target recognition based on multi-source information fusion.	2021	Bagging	Hypothetical case examples
	[96]: Bayesian network structure optimization strategy based on ensemble learning.	2021	Bagging	ASIA ALARM
	[97]: Radar target recognition based on XGBoost.	2023	Boosting	Custom radar dataset
	[98]: Small target detection based on XGBoost.	2023	Boosting	IPIX

4.3. Ensemble Deep Learning for Space Target Recognition

With the accumulation of data and the advancement of artificial intelligence, deep learning has been widely applied in object recognition. This section primarily summarizes the current applications of one-stage and two-stage object detection algorithms. A summary of Ensemble Deep Learning models is provided in Table 5.

Currently, the Convolutional Neural Network (CNN) is one of the most commonly used networks in the field of computer vision, being widely applied to space target classification and detection tasks. In 2018, Yan Z [99] pioneered the application of CNNs in spacecraft detection while Wu T [100] proposed the T-SCNN for space target recognition, achieving high accuracy. Additionally, CNNs leveraging transfer learning and data augmentation have demonstrated notable success in satellite classification and pose estimation [101]. Xi et al. [102] employed a LeNet-based CNN for debris classification that is effective across a range of signal-to-noise ratios. Bapu et al. [103] introduced an adaptive CNN using N-gram techniques to recognize satellite images. Arthur et al. [104] reviewed ensemble techniques for object detection in high-resolution remote sensing images, selecting three ensemble learning strategies: TTA with Bayesian Dropout, FGE, and SWA-for vehicle detection in desert regions. Yaotian Zhang [105] and Yangbo Wu [106] proposed CNN + LSTM and Transformer + CNN deep learning networks for space target recognition, respectively. Pryidharshini [107] presented a novel approach using a Bi-LSTM-CNN architecture optimized with Bayesian optimization for the efficient detection and prediction of space debris types and sizes (RCS). Olivi L et al. [108] proposed an improved method, R-Stack-CNN, which incorporates the Random Forest learning of time development features to enhance the robustness of the original STACK-CNN. Jiang Tao et al. [109] developed a CNN-based algorithm named SDebrisNet for detecting space debris. AlDahoul et al. [110] proposed a multi-modal deep learning method that integrates a visual Transformer model based on RGB images with an end-to-end CNN model for detecting spacecraft and space debris.

The integration of deep neural networks with image object detection has led to seminal algorithms such as R-CNN [111], Fast R-CNN [112], and Faster R-CNN [113]. R-CNN was the first method to apply deep neural networks for object detection. In 2020, Yang X [114] proposed a hybrid R-CNN with partial semantic information for space target recognition, enabling the segmentation of primary components of detected satellites. Li Linze et al. [115] drew inspiration from R-FCN and Light-head R-CNN to improve Mask R-CNN, resulting in a 20% reduction in detection time. Yulang Chen et al. [116] built a new feature extraction structure on the R-CNN base, incorporating DenseNet, ResNet, and FPN to create the RSD model for satellite component detection. Regarding the selection of basic learners, J Huang et al. [117] defined diversity among models as the cosine distance between category-wise average precision vectors. Building upon J Huang’s work, Jinsu Lee et al. [118] considered both category and object size in diversity calculations when selecting models.

To address the time-consuming issues of the R-CNN architecture, Redmon et al. [119] proposed the regression-based YOLO algorithm. The performance of the base object detection model can be improved by using YOLO ensembles [120] or combining transfer learning and ensemble learning with YOLO [121]. In terms of improving the YOLO network structure, Xiangjiu Che et al. [122] integrated a CA module into the backbone of YOLOv5 and proposed a Bagging-based dynamic weighted ensemble method for YOLOv5. M Yuan et al. [123] introduced a lightweight YOLOv3 network based on knowledge distillation, which performed well on a spacecraft component detection dataset. Aprinaldi et al. [124] utilized the NVIDIA Jetson module with the YOLOv5 model, significantly enhancing the detection capabilities for small objects. Qiang Tang et al. [125] optimized the ELAN module of YOLOv7 using RepPoint, proposing the YOLOv7-R model.

Liu et al. [126] combined the anchor mechanism of Faster R-CNN with YOLO to propose the SSD (Single-Shot MultiBox Detector) algorithm. Jie Xu et al. [127] used the SSD as the backbone and integrated ensemble learning with contextual modeling and multi-scale feature representation to ensure both accuracy and real-time performance in detection. Angela Casado-Garcia et al. [128] applied test-time augmentation to generate inputs for ensemble algorithms, integrating different architectures of YOLO, the SSD, and Faster R-CNN. S Zhang et al. [129] proposed a novel single-shot detector, RefineDet++, which balances the recognition precision of two-stage methods with the efficiency of one-stage methods. Zhang H et al. [130] introduced DDS-SSD, which employs dilated convolution modules in the feature fusion to expand the receptive field of shallow features and uses deconvolution modules to increase the size of high-level feature maps.

In addition to the above algorithms, researchers have also integrated other deep neural networks. For instance, CoAtNets have been utilized for solving fused image classification problems [131] while Stacking has been applied to bounding box fusion tasks [132].

Although deep learning demonstrates outstanding performance, it is highly dependent on large datasets and involves complex hyperparameter tuning. In the field of space target recognition, there is still a lack of publicly available large-scale datasets. Therefore, future research could focus on developing space target databases or exploring few-shot detection methods.

Table 5. Overview of applications in ensemble of deep learning.

Application	Reference	Year	Model	Dataset
CNN	[99]: CNN for spacecraft detection.	2018	CNN	Space Target Dataset
	[100]: Two-stage Convolutional Neural Network (T-SCNN) for space target recognition.	2019	T-SCNN	Space Target Dataset
	[101]: Satellite classification and pose regression using transfer learning and data augmentation.	2020	CNN	BUAA-SID
	[102]: LeNet-based CNN for debris classification.	2020	CNN	Space debris Simulation images
	[103]: N-gram-based adaptive Convolutional Neural Network.	2019	CNN	Google Earth images
	[104]: Ensemble learning methods for high-resolution satellite image object detection.	2022	Bagging, CNN	Desert environment images
	[105]: Space target recognition method combing CNN and LSTM.	2018	CNN, LSTM	Space target Simulation data
	[106]: Space target recognition method combing CNN and Transformer.	2023	CNN, Transformer	RCS simulation data
	[107]: Space debris detection and monitoring.	2025	Bi-LSTM-CNN	Satellites and debris in Earth’s orbit
	[108]: Meteor and space debris detection.	2024	Refined STACK-CNN	Mini-EUSO session 6
	[109]: Space debris detection.	2023	SDebrisNet	SDD
	[110]: Spacecraft and debris recognition.	2022	Transformer + CNN	SPARK
R-CNN Series	[114]: Two-stage object detection network based on space target characteristics.	2020	R-CNN	BUAA-SID, space debris data
	[115]: Improved Mask R-CNN for feature detection and recognition of non-cooperative space targets.	2020	Mask R-CNN	Simulated satellite images
	[116]: Satellite component detection model (RSD) based on R-CNN.	2020	R-CNN	Simulated satellite images
	[117]: Model diversity computation method.	2017	Faster R-CNN	COCO
	[118]: Model diversity computation method.	2018	ResNet	PASCAL VOC
YOLO Series	[120]: Butterfly automatic detection and classification algorithm based on YOLO ensemble.	2020	YOLOv3	Butterfly Image Dataset
	[121]: Drone image object detection method based on transfer learning and ensemble learning.	2021	YOLOv3	VisDrone, AU-AIR
	[122]: Chest imaging disease detection method based on dynamic weighted Bagging strategy.	2022	Bagging, YOLOv5	VinDr-CXR
	[123]: Lightwight YOLOv3-based spacecraft component detection.	2022	YOLOv3	Spacecraft Component Detection Dataset
	[124]: Target detection method based on transfer learning and YOLOv5.	2022	YOLOv5	RGBT
	[125]: Optimization of YOLOv7’s ELAN module using RepPoint.	2023	YOLOv7	BUAA-SID, etc.
SSD Series	[127]: Ensemble learning-based object detection method based on SSD.	2020	SSD	Pascal VOC, MS COCO
	[128]: Ensemble learning fusion strategy and test-time augmentations for object detection.	2020	SSD	Pascal VOC, Stomata, etc.
	[129]: Novel detector RefineDet++ based on single-shot detection.	2020	SSD	Pascal VOC, MS COCO
	[130]: Enhanced SSD with a new feature fusion module.	2021	SSD	Pascal VOC, MS COCO
Others	[131]: Ensemble of CoAtNet networks.	2023	Stacking, CoAtNets	Satellite Simulation Data
	[132]: Directed bounding box fusion method for object detection.	2023	OBBStacking	DOTA, FAIR1M

Table 5. Overview of applications in ensemble of deep learning.

Application	Reference	Year	Model	Dataset
CNN	[99]: CNN for spacecraft detection.	2018	CNN	Space Target Dataset
	[100]: Two-stage Convolutional Neural Network (T-SCNN) for space target recognition.	2019	T-SCNN	Space Target Dataset
	[101]: Satellite classification and pose regression using transfer learning and data augmentation.	2020	CNN	BUAA-SID
	[102]: LeNet-based CNN for debris classification.	2020	CNN	Space debris Simulation images
	[103]: N-gram-based adaptive Convolutional Neural Network.	2019	CNN	Google Earth images
	[104]: Ensemble learning methods for high-resolution satellite image object detection.	2022	Bagging, CNN	Desert environment images
	[105]: Space target recognition method combing CNN and LSTM.	2018	CNN, LSTM	Space target Simulation data
	[106]: Space target recognition method combing CNN and Transformer.	2023	CNN, Transformer	RCS simulation data
	[107]: Space debris detection and monitoring.	2025	Bi-LSTM-CNN	Satellites and debris in Earth’s orbit
	[108]: Meteor and space debris detection.	2024	Refined STACK-CNN	Mini-EUSO session 6
	[109]: Space debris detection.	2023	SDebrisNet	SDD
	[110]: Spacecraft and debris recognition.	2022	Transformer + CNN	SPARK
R-CNN Series	[114]: Two-stage object detection network based on space target characteristics.	2020	R-CNN	BUAA-SID, space debris data
	[115]: Improved Mask R-CNN for feature detection and recognition of non-cooperative space targets.	2020	Mask R-CNN	Simulated satellite images
	[116]: Satellite component detection model (RSD) based on R-CNN.	2020	R-CNN	Simulated satellite images
	[117]: Model diversity computation method.	2017	Faster R-CNN	COCO
	[118]: Model diversity computation method.	2018	ResNet	PASCAL VOC
YOLO Series	[120]: Butterfly automatic detection and classification algorithm based on YOLO ensemble.	2020	YOLOv3	Butterfly Image Dataset
	[121]: Drone image object detection method based on transfer learning and ensemble learning.	2021	YOLOv3	VisDrone, AU-AIR
	[122]: Chest imaging disease detection method based on dynamic weighted Bagging strategy.	2022	Bagging, YOLOv5	VinDr-CXR
	[123]: Lightwight YOLOv3-based spacecraft component detection.	2022	YOLOv3	Spacecraft Component Detection Dataset
	[124]: Target detection method based on transfer learning and YOLOv5.	2022	YOLOv5	RGBT
	[125]: Optimization of YOLOv7’s ELAN module using RepPoint.	2023	YOLOv7	BUAA-SID, etc.
SSD Series	[127]: Ensemble learning-based object detection method based on SSD.	2020	SSD	Pascal VOC, MS COCO
	[128]: Ensemble learning fusion strategy and test-time augmentations for object detection.	2020	SSD	Pascal VOC, Stomata, etc.
	[129]: Novel detector RefineDet++ based on single-shot detection.	2020	SSD	Pascal VOC, MS COCO
	[130]: Enhanced SSD with a new feature fusion module.	2021	SSD	Pascal VOC, MS COCO
Others	[131]: Ensemble of CoAtNet networks.	2023	Stacking, CoAtNets	Satellite Simulation Data
	[132]: Directed bounding box fusion method for object detection.	2023	OBBStacking	DOTA, FAIR1M

5. Experiments and Analysis

5.1. Experimental Dataset

The dataset used in this experiment comprises satellite simulation data generated by Blender. To authentically represent the geometric structures of satellites, 3D models of CloutSat, Jason-1, and Sentinel-6 were imported. Utilizing Blender’s advanced rendering capabilities, images were simulated under various angular and lighting conditions. For each satellite model, 1071 synthetic images were generated, totaling 3213 images, each with a resolution of 640 × 640 pixels. The dataset was divided into training and testing sets in a 7:3 ratio. Sample examples are shown in Figure 9.

To enhance the realism and diversity of the dataset and improve the generalization ability of the model, this paper used data enhancement methods such as geometric transformations, color adjustments, blur effects, and noise addition to process the original images. A total of 6000 satellite simulation images were eventually obtained. Ultimately, a total of 6000 satellite simulation images were obtained. These enhancements have provided a richer and more reliable dataset for subsequent satellite recognition tasks. Examples of the augmented data are shown in Figure 10.

The use of Blender for generating simulated satellite images enables rapid data generation and flexible control over experimental conditions. However, synthetic data inherently possess limitations that may impact model performance. Firstly, synthetic data struggle to fully replicate complex real-world physical phenomena, such as lighting, shadows, and material reflections, leading to potential discrepancies between simulated and real images. This can hinder the model’s ability to handle the complexities of real-world data. Secondly, synthetic data often lack common real-world anomalies, including noise, blurring, and occlusions, which may reduce the model’s robustness to such disturbances. Additionally, the 3D satellite models used in synthetic data generation may oversimplify geometric structures and surface textures, omitting critical details present in real satellites. This simplification can adversely affect the model’s recognition and classification capabilities in practical applications.

5.2. Experimental Design

To compare the recognition performance of various traditional machine learning algorithms, ensemble algorithms, and deep learning algorithms for space target recognition, this paper utilizes the simulated dataset to evaluate the performance of SVM, decision tree, k-nearest neighbors (KNN), multi-layer perceptron (MLP), AdaBoost, Random Forest, and Stacking algorithms. Additionally, different numbers of basic learners have been configured to analyze their impact on the ensemble learning algorithms. Finally, the paper compares and analyzes the performance of homogeneous and heterogeneous ensemble models.

5.2.1. Feature Extraction

The Histogram of Oriented Gradients (HOG) is a widely used image feature descriptor in computer vision and object recognition tasks. It captures the shape and appearance information of objects by computing the gradient orientation histograms of local regions within an image. The core principle of HOG is that the appearance and shape of local objects in an image can be effectively described by the gradient or the directional distribution of edges. The basic process for extracting HOG features from an image is illustrated in Figure 11.

Given that space targets possess distinct geometric features and are subjected to varying lighting conditions, and considering that HOG effectively captures shape and contour information while being robust to changes in lighting and shadows, we have selected HOG features as the input for the model. However, in practical applications, HOG features are high-dimensional and computationally intensive. To address this, after extracting HOG features, we have employed PCA to reduce the feature dimensionality to 100 dimensions.

5.2.2. Parameter Settings

The experimental environment for this paper is shown in

Table 6. Experimental environment information.

Category	Detailed Information
CPU	8-core Ryzen 7 6800H 3.2 GHz
Memory	16 GByte
Storage	512 GByte
Operating system	Windows 11 Home Chinese Edition
System type	x64-based PC
Programming language	Python
Programming tool	Visual Studio Code

. The SVM, decision tree, KNN, AdaBoost, Random Forest, and Stacking algorithms have been implemented using the scikit-learn machine learning library.

As for the parameter settings of the basic algorithms, the SVM employs a linear kernel function and sets the penalty parameter for the error term to 1. The efficiency and simplicity of the linear kernel make it excel in handling linearly separable datasets while an appropriately set penalty parameter helps balance model complexity and training error. The KNN algorithm is configured with 5 neighbors, using uniform weights and Euclidean distance to compute similarities between samples. For the decision tree algorithm, uniform weights are chosen to avoid biases caused by class imbalance. Additionally, no maximum depth is imposed on the decision tree, and all features are considered for finding the best splits, allowing the decision tree to fully utilize the details in the data and maximize its fitting capacity. The MLPClassifier is configured with a multi-layer perceptron (MLP) architecture consisting of three hidden layers with 128, 64, and 32 neurons. The model employs the ReLU activation function for non-linear transformations and utilizes the Adam optimizer for efficient gradient-based optimization. Training is performed for a maximum of 20 iterations, balancing computational efficiency and model convergence.

The parameters for the ensemble learning models have been determined through grid search. In the AdaBoost model, decision trees have been chosen as the basic learners with a learning rate of 0.5, 200 basic learners, and a maximum depth of 4 for each decision tree. For the Random Forest model, the number of decision trees has been set to 200, with a maximum depth of 10 for each tree and a minimum of 1 sample per leaf node. The primary model in the Stacking model is Random Forest, and the logistic regression model has been chosen for the secondary model to prevent overfitting. The number of trees in the Random Forest is 100, and the maximum number of iterations for the logistic regression model has been set to 1000.

5.2.3. Performance Evaluation Metrics

In space target recognition experiments, precision, recall, F1-score, and accuracy are four core evaluation metrics used to comprehensively assess model performance.

Precision measures the proportion of correctly predicted positive samples among all samples predicted as positive, calculated thus:

$$precision={\displaystyle \frac{TP}{TP+FP}},$$

(1)

Here, TP (true positives) denotes samples correctly predicted as positive and FP (false positives) denotes samples incorrectly predicted as positive. A higher precision indicates fewer false alarms and more reliable predictions. Recall measures the proportion of correctly predicted positive samples among all actual positive samples, calculated thus:

$$Recall={\displaystyle \frac{TP}{TP+FN}},$$

(2)

Here, FN (false negatives) denotes samples incorrectly predicted as negative. A higher recall indicates fewer missed detections and better identification of positive samples. The F1-score is the harmonic mean of precision and recall, providing a balanced measure particularly useful in imbalanced class scenarios, calculated as Equation (3).

$$F1=2\times {\displaystyle \frac{precision\times recall}{precision+recall}},$$

(3)

Accuracy is calculated thus:

$$Accuracy={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}},$$

(4)

Here, TN (true negatives) denotes samples correctly predicted as negative, reflecting the overall correctness of predictions. However, accuracy may be misleading in imbalanced datasets as models may favor the majority class.

Precision, recall, and F1-score are often combined to provide a more comprehensive evaluation, especially in space target recognition tasks where both false alarms and missed detections are critical.

5.3. Experimental Result Analysis

5.3.1. Comparative Analysis of Model Classification Performance

From the experimental results in

Table 7. Performance of different models on simulated data.

Model	Precision	Recall	F1-Score	Accuracy Score	Execution Time
SVM	0.97512	0.97496	0.97504	0.97500	0.9712
Decision tree	0.93730	0.93725	0.93720	0.93722	0.3871
KNN	0.93004	0.92413	0.92508	0.92556	0.4252
MLP	0.98451	0.98444	0.98441	0.98444	2.4592
AdaBoost	0.98010	0.97910	0.97951	0.97944	26.7865
Random Forest	0.97903	0.97887	0.97895	0.97889	3.0597
Stacking	0.98740	0.98718	0.98729	0.98722	12.3631

(bold fonts are the best indicators), the Stacking model achieves the highest accuracy, with all metrics exceeding 0.98, but it has a longer execution time (12.3631 s) due to the need to train multiple base models and integrate them through a meta-model, resulting in high computational complexity. The MLP model exhibits accuracy close to Stacking (around 0.984) but also requires significant execution time (2.4592 s) as it involves multiple forward and backward propagation computations, making the training process time-consuming. In contrast, AdaBoost and Random Forest show slightly lower accuracy (around 0.98), but their execution times differ significantly. AdaBoost, which iteratively trains multiple weak classifiers, has the longest execution time (26.7865 s), while Random Forest, which builds decision trees in parallel, has a shorter execution time (3.0597 s). The SVM model demonstrates a balance between accuracy (0.975) and execution time (0.9712 s), making it suitable for high-dimensional linear problems. Decision tree and KNN have the shortest execution times (0.3871 s and 0.4252 s, respectively), making them suitable for real-time applications, but their lower model complexity results in poorer accuracy (around 0.93), limiting their performance on complex data. Overall, Stacking and MLP are ideal for high-precision tasks while decision tree and KNN are better suited for scenarios requiring high computational efficiency.

The detection results of each algorithm are shown in Figure 12. Misclassified samples are marked with red fonts and red detection boxes, and correctly classified samples are marked with green fonts and blue detection boxes. As can be seen from the figure, the detection accuracy values of the ensemble learning algorithms, deep learning algorithm, and SVM are higher while the error rates of decision tree and KNN are higher.

To better evaluate model performance, Receiver Operating Characteristic (ROC) curves were plotted for each category, as shown in Figure 13. This paper has employed a ‘one-vs-rest’ strategy to convert the multi-class classification problem into multiple binary classification problems.

The confusion matrix of different machine learning models is shown in Figure 14. All models demonstrate better performance for the CloutSat satellite while showing that poorer effectiveness is closely related to the shape and color characteristics of the satellites.

In summary, the performance differences among the models not only reflect the inherent characteristics of the models themselves but also depend on the nature of the task and the selected features. All three types of ensemble models have demonstrated strong performance, with Stacking achieving the best results in this experiment. The SVM model and MLP model have also exhibited excellent classification capabilities.

5.3.2. Analysis of Impact of the Number of Basic Learners on Classification Performance

This paper has set different numbers of basic learners to analyze the effect of the number of basic learners on the classification performance of ensemble learning models. As shown in Figure 15, the number of basic learners has a minimal effect on the Stacking model but a more significant impact on the Random Forest and AdaBoost models. As the number of basic learners increases, the accuracy of all three models generally improves. However, when the number of basic learners reaches 50 or more, there is no significant change in the accuracy of the models, indicating that the number of basic learners can enhance classification performance within a certain range.

5.3.3. Analysis of Classification Performance of Homogeneous and Heterogeneous Ensemble

As shown in

Table 8. Performance of homogeneous and heterogeneous ensemble models.

Model	Ensemble Method	Basic Learners	Precision	Recall	F1-Score	Accuracy Score
AdaBoost	Homogeneous	Decision tree	0.98010	0.97910	0.97951	0.97944
Random Forest	Homogeneous	Decision tree	0.97903	0.97887	0.97895	0.97889
Stacking	Homogeneous	Decision tree, LogisticRegression	0.98740	0.98718	0.98729	0.98722
AdaBoost	Heterogeneous	SVM, KNN, decision tree, LogisticRegression, RandomForest	0.91315	0.90944	0.90916	0.90944
Bagging	Heterogeneous	SVM, decision tree	0.96278	0.96278	0.96275	0.96278
Stacking	Heterogeneous	SVM, RandomForest, LogisticRegression	0.97778	0.97778	0.97778	0.98778

(bold fonts are the best indicators), there are differences in the performance of homogeneous and heterogeneous ensemble models. Homogeneous basic learners consistently demonstrate higher performance in the object detection task while heterogeneous basic learner ensembles perform less satisfactorily in some metrics, particularly in the case of AdaBoost, where performance is significantly lower than that of homogeneous basic learners. This discrepancy might be due to the inability of the heterogeneous ensemble models to effectively integrate the strengths of the basic learners or the unsuitable fusion method chosen for the ensemble. Additionally, the sequential training approach of AdaBoost, which aims to gradually improve model performance, may struggle to collaborate effectively with different types of basic learners. In contrast, the Bagging and Stacking methods perform better with heterogeneous basic learners, with the Stacking method achieving an Accuracy Score of 0.98778, indicating that heterogeneous ensembles can still yield advantages in certain ensemble methods. Therefore, selecting suitable basic learners, fusion strategies, and ensemble methods is crucial for enhancing model performance.

In this experiment, homogeneous ensembles have exhibited higher true positive rates (TPRs) and lower false positive rates (FPRs) across all methods as indicated by their ROC curves. Notably, in the Stacking method, the Area Under the Curve (AUC) approaches 1, suggesting an exceptionally high classification capability. In contrast, the AUC of the heterogeneous ensemble using AdaBoost shows a slight decline, as illustrated in Figure 16.

The confusion matrix for the heterogeneous ensemble model is shown in Figure 17. The confusion matrix for AdaBoost indicates a higher misclassification rate for the Sentinel-6 type of satellite, leading to a decrease in the overall performance of the model.

6. Discussion

6.1. Challenges and Outlook

Space target recognition, as a key component of SSA, plays a critical role in maintaining space security. Despite the significant progress made in recent years by target recognition methods based on ensemble learning, they still encounter challenges in the practical application of space target recognition.

• Imbalance in datasets: In this paper, experiments were conducted using balanced datasets, where the Boosting algorithm exhibited relatively poor performance. However, given its common application for handling unbalanced datasets, Boosting may demonstrate superior performance in practical scenarios. Moving forward, the development of effective data resampling techniques and cost-sensitive learning methods represents an important direction for future research.
• Selection of basic learners: The diversity of basic learners is challenging to precisely define and measure, and its correlation with ensemble learning performance remains elusive. Future research should investigate automated selection mechanisms for basic learners to dynamically identify the optimal combinations.
• Computational resources and efficiency: The application of ensemble learning-based space target recognition methods in space is constrained by the limited computational, storage, and energy resources of satellite on-board computers. Additionally, the high latency and narrow bandwidth of satellite communication pose challenges for the real-time data transmission required by these methods. The adaptability and robustness of models may also be compromised in complex and dynamic space environments. To address these limitations, strategies such as model lightweighting, energy optimization, distributed computing, communication optimization, and adaptive learning should be considered.
• Establishment of standardized space target recognition datasets: Currently, the field of space target recognition lacks large-scale annotated datasets. A comprehensive and diverse standardized database would provide a robust foundation for the research and validation of integrated learning algorithms. Additionally, it would establish unified annotation and evaluation benchmarks, thereby advancing the standardization of space target recognition technologies.
• Single-scene detection: The ensemble learning-based space target recognition method proposed in this paper has been evaluated only for single-target scenes and does not account for complex scenarios involving multiple satellites or colliding satellite fragments within an image. In real space environments, multiple floating objects may appear tightly clustered or as a single entity, or a satellite may fragment into multiple parts due to collisions, complicating accurate target identification. Future research should explore multi-target perception and separation techniques such as instance segmentation methods (e.g., Mask R-CNN) or graph neural networks (GNNs) to model spatial relationships between targets. Addressing the multi-target perception challenge will bring the method closer to practical application requirements and provide more robust support for space target monitoring and collision warning.

6.2. Conclusions

This paper has provided a comprehensive review in three areas: ensemble learning, space target recognition, and the application of ensemble learning for object recognition. The analysis has been validated using a self-constructed space target simulation dataset. Ensemble learning, by integrating multiple basic learners, enhances the accuracy and robustness of object recognition, effectively addressing challenges such as complex space target data and class imbalance, making it highly valuable and promising for applications in space target recognition. In the future, the integration of ensemble learning with deep learning, the automated selection of basic learners, and efficient computational resource optimization will be key research areas. Additionally, establishing standardized space target databases, improving model interpretability, and clarifying the relationship between ensemble system diversity and accuracy will further advance the application of ensemble learning in space target recognition.