Improve Classification Accuracy Research Articles

Adapting YOLO11 for Classification of Unlabelled Data : A Semi-Supervised Approach

Computer vision has advanced rapidly, improving object detection and classification. The YOLO (You Only Look Once) model family is widely used because of its real-time processing efficiency and object detection accuracy. However, most YOLO-based methods focus on object detection rather than categorization. Deep learning models sometimes require large amounts of labelled data, which can be expensive and time-consuming, especially in specialized sectors. Traditional classification methods require extensive manual dataset annotation and fully supervised learning. When labelled data is scarce, this reliance is a major issue. Semi-supervised learning (SSL) improves model efficacy by using labelled and unlabeled data. SSL methods to reduce labelling burdens have been studied, however SSL with YOLO-based designs has not. This project uses semi-supervised learning to adapt YOLO11, an advanced deep learning framework, for classification issues. This work aims to improve classification accuracy in environments with little tagged data, reducing manual tagging. This project will test self-training, pseudo-labeling, and consistency regularization in YOLO11-based classification models. This study uses semi-supervised learning to link object detection and categorization in the YOLO framework. The proposed method could be used in autonomous driving, surveillance, and medical imaging, where comprehensive labelling is often impossible. This work will explain YOLO11 classification optimization and validate SSL techniques in deep learning classification challenges. On the Adidas logo dataset, YOLOv11, the latest framework, was used to classify images according to the object box size into large, medium, and small groups, then chose the best group to use it for the next step, which is the object detection, these steps yielded encouraging results, with a MAP train of 0.913 and a 36-minute training time.

Few-shot traffic classification based on autoencoder and deep graph convolutional networks

Traffic classification is a crucial technique in network management that aims to identify and manage data packets to optimize network efficiency, ensure quality of service, enhance network security, and implement policy management. As graph convolutional networks (GCNs) take into account not only the features of the data itself, but also the relationships among sets of data during classification. Many researchers have proposed their own traffic classification methods based on GCN in recent years. However, most of the current approaches use two-layer GCN primarily due to the over-smoothing problem associated with deeper GCN. In scenarios with small samples, a two-layer GCN may not adequately capture relationships among traffic data, leading to limited classification performance. Additionally, during graph construction, traffic usually needs to be trimmed to a uniform length, and for traffic with insufficient length, zero-padding is typically applied to extension. This zero-padding strategy poses significant challenges in traffic classification with small samples. In this paper, we propose a method based on autoencoder (AE) and deep graph convolutional networks (ADGCN) for traffic classification for few-shot datasets. ADGCN first utilizes an AE to reconstruct the traffic. AE enables shorter traffic to learn abstract feature representations from longer traffic of the same class to replace zeros, mitigating the adverse effects of zero-padding. The reconstructed traffic is then classified using GCNII, a deep GCN model that addresses the challenge of insufficient data samples. ADGCN is an end-to-end traffic classification method applicable to various scenarios. According to experimental results, ADGCN can achieve a classification accuracy improvement of 3.5 to 24% compared to existing state-of-the-art methods. The code is available at https://github.com/han20011019/ADGCN.

From 1-D to 3-D: LIBS Pseudohyperspectral Data Cube Deep Learning Mechanism Used in Nuclear Metal Materials Classification.

In this paper, we propose a new spectral data mechanism called LIBS pseudohyperspectral data cube. This mechanism allows for the utilization of multidimensional information from laser-induced plasma, transforming 1-D LIBS spectra into 3-D data cube. Specifically, two additional dimensions are introduced to capture spectral variations information, allowing more features to be learned during pretraining. Proposed mechanism can make the LIBS system more robust when handling unstable spectra acquired onsite, and can also allow LIBS take full advantage of deep learning algorithms. In the context of nuclear power plants, traditional LIBS classification faces significant challenges due to unstable spectra, which reduce the accuracy of classifying similar or tiny extreme condition materials. By combining deep learning algorithms with LIBS pseudohyperspectral data cube, we can capture spectral and other dimensional features to enhance classification accuracy. Experimental results show that, compared to traditional 1-D data processing, the new method significantly improves the classification accuracy of unstable spectra. Moreover, by incorporating an attention mechanism, the model can adaptively adjust the weights of different features, further improving classification accuracy to over 99%. Visualizing the attention mechanism's weight matrix allows us to identify the importance of different features in classification. Additionally, t-SNE visualizations demonstrate the clustering of different categories in the feature space, further validating the performance of the new method. We believe this data cube mechanism offers an effective new approach for applying deep learning algorithms and enhancing data dimensionality in the LIBS field.

Ensemble Classifier for Web Data Scraping with Lexicon Support

An ensemble classifier for web data is used for selective web scraping with lexical support in an innovative way to improving the accuracy and efficiency of data classification from web sources. Web scraping, is a method for obtaining information from webpages, frequently produces massive, unstructured datasets and high risk in data reliability which leads to misuse in communication that are difficult to manage. To overcome this, selective online scraping is used to target certain information important to the classification task, resulting in less noise and higher data quality. The ensemble classifier integrates numerous machine learning models to maximize their strengths, resulting in better overall performance. In this approach, separate classifiers are trained on distinct subsets of scraped data that are chosen based on predetermined criteria utilizing lexicons, which are collections of domain-specific words and phrases. These lexicons guide the selective scraping process, ensuring that only the most relevant data is captured, hence improving classifier accuracy. After scraping and pre-processing the data, the ensemble method aggregates predictions from each classifier, generally using techniques like majority voting, stacking, or weighted average, to get a final classification result. This strategy not only promotes robustness by reducing the risk of overfitting, but it also improves flexibility across other domains by incorporating lexical assistance tailored to specific themes or sectors. The combination of selective web scraping and lexical assistance enables more targeted and resource-efficient data collecting, while the use of an ensemble classifier assures excellent accuracy and reliability in classification tasks. This methodology is especially useful in circumstances where the online data is large, dynamic, and contains a lot of unnecessary or noisy information. The resulting system provides a scalable and effective solution for real-time web data classification, with applications in sentiment analysis, content categorization, and market intelligence.

Multiple perception contrastive learning for automated ovarian tumor classification in CT images.

Ovarian cancer is among the most common malignant tumours in women worldwide, and early identification is essential for enhancing patient survival chances. The development of automated and trustworthy diagnostic techniques is necessary because traditional CT picture processing mostly depends on the subjective assessment of radiologists, which can result in variability. Deep learning approaches in medical image analysis have advanced significantly, particularly showing considerable promise in the automatic categorisation of ovarian tumours. This research presents an automated diagnostic approach for ovarian tumour CT images utilising supervised contrastive learning and a Multiple Perception Encoder (MP Encoder). The approach incorporates T-Pro technology to augment data diversity and simulates semantic perturbations to increase the model's generalisation capability. The incorporation of Multi-Scale Perception Module (MSP Module) and Multi-Attention Module (MA Module) enhances the model's sensitivity to the intricate morphology and subtle characteristics of ovarian tumours, resulting in improved classification accuracy and robustness, ultimately achieving an average classification accuracy of 98.43%. Experimental results indicate the method's exceptional efficacy in ovarian tumour classification, particularly in cases involving tumours with intricate morphology or worse picture quality, thereby markedly enhancing classification accuracy. This advanced deep learning framework proficiently tackles the complexities of ovarian tumour CT image interpretation, offering clinicians enhanced diagnostic support and aiding in the optimisation of early detection and treatment strategies for ovarian cancer.

Enhancing Hierarchical Classification in Tree-Based Models Using Level-Wise Entropy Adjustment

Hierarchical classification, which organizes items into structured categories and subcategories, has emerged as a powerful solution for handling large and complex datasets. However, traditional flat classification approaches often overlook the hierarchical dependencies between classes, leading to suboptimal predictions and limited interpretability. This paper addresses these challenges by proposing a novel integration of tree-based models with hierarchical-aware split criteria through adjusted entropy calculations. The proposed method calculates entropy at multiple hierarchical levels, ensuring that the model respects the taxonomic structure during training. This approach aligns statistical optimization with class semantic relationships, enabling more accurate and coherent predictions. Experiments conducted on real-world datasets structured according to the GS1 Global Product Classification (GPC) system demonstrate the effectiveness of our method. The proposed model was applied using tree-based ensemble methods combined with the newly developed hierarchy-aware metric Penalized Information Gain (PIG). PIG was implemented with level-wise entropy adjustments, assigning greater weight to higher hierarchical levels to maintain the taxonomic structure. The model was trained and evaluated on two real-world datasets based on the GS1 Global Product Classification (GPC) system. The final dataset included approximately 30,000 product descriptions spanning four hierarchical levels. An 80-20 train–test split was used, with model hyperparameters optimized through 5-fold cross-validation and Bayesian search. The experimental results showed a 12.7% improvement in classification accuracy at the lowest hierarchy level compared to traditional flat classification methods, with significant gains in datasets featuring highly imbalanced class distributions and deep hierarchies. The proposed approach also increased the F1 score by 12.6%. Despite these promising results, challenges remain in scaling the model for very large datasets and handling classes with limited training samples. Future research will focus on integrating neural networks with hierarchy-aware metrics, enhancing data augmentation to address class imbalance, and developing real-time classification systems for practical use in industries such as retail, logistics, and healthcare.

Efficient remote sensing image classification using the novel STConvNeXt convolutional network

Remote sensing images present formidable classification challenges due to their complex spatial organization, high inter-class similarity, and significant intra-class variability. To address the balance between computational efficiency and feature extraction capability in existing methods, this paper innovatively proposes a lightweight convolutional network, STConvNeXt. In its architectural design, the model incorporates a split-based mobile convolution module with a hierarchical tree structure. It employs parameterized depthwise separable convolutions to reduce computational complexity and constructs a multi-level feature tree to facilitate cross-scale feature fusion. For feature enhancement, a fast pyramid pooling module replaces the traditional spatial pyramid structure, effectively reducing the number of parameters while preserving large-scale contextual awareness. In terms of training strategy, a dynamic threshold loss function is introduced, utilizing a learnable inter-class margin to improve the model’s ability to distinguish difficult-to-classify samples. Systematic experiments on the UCMerced, AID, and NWPU-RESISC45 benchmark datasets validate the effectiveness of the proposed approach: compared with the ConvNeXt baseline, STConvNeXt reduces both parameter count (by 56.49%) and FLOPs (by 49.89%), while improving classification accuracy by 1.2–2.7%. Furthermore, compared with the current state-of-the-art remote sensing scene classification models, our method still exhibits significant advantages. Ablation studies further confirm the effectiveness of each module design, particularly demonstrating that the model maintains excellent classification accuracy despite a substantial reduction in parameters.

An innovative approach to classify meniscus tears by reducing vision transformers features with elasticnet approach

Meniscal tears, a prevalent orthopedic condition caused by abrupt knee movements, excessive load, or injury, require an accurate diagnosis for effective treatment. This study investigates the vision transformer (ViT) models' efficacy in automated classification of meniscus pathologies. It also explores how feature reduction using the ElasticNet method can improve classification accuracy and computational efficiency. The study utilized MRI scans from a dataset comprising 5000 images collected from clinical cases. Initially, classification was performed using EfficientNet and SqueezeNet architectures. Subsequently, feature extraction was conducted using ViT models, generating a feature set of 1000 dimensions. ElasticNet was employed to reduce features before reclassification using support vector machines (SVM). Model performance was evaluated based on accuracy, precision, sensitivity, and specificity. The ViT_base_32 model achieved a classification accuracy of 99.9% with a processing time of 1.2 s. Feature reduction via ElasticNet significantly enhanced classification performance while maintaining high precision, sensitivity, and specificity. These improvements demonstrate the effectiveness of combining ViT models with ElasticNet to diagnose meniscal tears. The findings highlight the potential of vision transformer models, in conjunction with ElasticNet, to provide rapid and highly accurate diagnostic assistance for meniscal injuries. This methodology shows promise for application to other medical diagnostic domains, offering valuable advancements in healthcare technology.

Temporal Dynamics in Short Text Classification: Enhancing Semantic Understanding Through Time-Aware Model

Traditional text classification models predominantly rely on static text representations, failing to capture temporal variations in language usage and evolving semantic meanings. This limitation reduces their ability to accurately classify time-sensitive texts, where understanding context, detecting trends, and addressing semantic shifts over time are critical. This paper introduces a novel time-aware short text classification model incorporating temporal information, enabling tracking of and adaptation to evolving language semantics. The proposed model enhances contextual understanding by leveraging timestamps and significantly improves classification accuracy, particularly for time-sensitive applications such as News topic classification. The model employs a hybrid architecture combining Convolutional Neural Networks (CNNs) and Bidirectional Long Short-Term Memory (BiLSTM) networks, enriched with attention mechanisms to capture both local and global dependencies. To further refine semantic representation and mitigate the effects of semantic drift, the model fine-tunes GloVe embeddings and employs synonym-based data augmentation. The proposed approach is evaluated on three benchmark dynamic datasets, achieving superior performance with classification accuracy reaching 92% for the first two datasets and 85% for the third dataset. Furthermore, the model is applied to a different-fields categorization and trend analysis task, demonstrating its capability to capture temporal patterns and perform detailed trend analysis of domain-agnostic textual content. These results underscore the potential of the proposed framework to provide deeper insights into the evolving nature of language and its impact on short-text classification. This work advances natural language processing by offering a comprehensive time-aware classification framework, addressing the challenges of temporal dynamics in language semantics.

Better Network Optimization Through Batch Normalization in Left Ventricle Chamber Classification

Convolutional neural networks (CNNs) have emerged as a prominent deep learning technique for medical image classification. This study investigated the impact of batch normalization layer placement on the performance of the CNNs model in classifying the left ventricle segment in Delayed-enhancement cardiac magnetic resonance (De-CMR) image slices. Three batch normalization arrangements, including one without a batch normalization layer, were examined to assess their impact. Additionally, the influence of three learning rates (0.0001, 0.001, 0.01) from two different types of optimizers, namely Adam and Sgdm, was explored to identify the optimal configuration for our proposed CNN model. A model without batch normalization was used as a baseline for comparison. The results show that placing batch normalization after the convolutional layers, combined with the Adam optimizer and a learning rate of 0.0001, yielded the best performance, improving classification accuracy from 83.1% to 88.4%. These results highlight the significance of batch normalization layers with optimal configuration in enhancing the performance in the classification of the left ventricle and non-LV chambers in De-CMR images, thereby facilitating improvements in the streamlined workflow for automated myocardial infarction diagnosis.

Formulation of K-nearest neighbor model by varying the distance metrics of mahalanobis, correlation and cosine in discriminating different grades of Aquilaria oil

This study formulates a K-Nearest Neighbor (K-NN) model for the classification of Aquilaria (agarwood) oil grades by varying distance metrics: Mahalanobis, Correlation and Cosine. A total of 96 agarwood oil samples were analyzed using Gas chromatography-mass spectrometry (GC-MS), identifying key chemical compounds contributing to quality differentiation. The proposed model achieved 100% classification accuracy at k = 1 across all distance metrics with Mahalanobis distance consistently outperforming others as k increased. Cross-validation results demonstrated the lowest error rates for Mahalanobis followed by Correlation and Cosine, confirming its robustness in high-dimensional chemical datasets. These findings highlight the effectiveness of machine learning in agarwood oil grading and underscore the importance of selecting optimal distance metrics for improved classification accuracy. Future work could incorporate ensemble learning and advanced feature selection to further refine performance.

Self-supervised learning reduces label noise in sharp wave ripple classification.

In the field of electrophysiological signal analysis, the classification of time-series datasets is essential. However, these datasets are often compromised by the prevalent issue of incorrect attribution of labels, known as label noise, which may arise due to insufficient information, inappropriate assumptions, specialists' mistakes, and subjectivity, among others. This critically impairs the accuracy and reliability of data classification, presenting significant barriers to extracting meaningful insights. Addressing this challenge, our study innovatively applies self-supervised learning (SSL) for the classification of sharp wave ripples (SWRs), high-frequency oscillations involved in memory processing that were generated before or after the encoding of spatial information. This novel SSL methodology diverges from traditional label correction techniques. By utilizing SSL, we effectively relabel SWR data, leveraging the inherent structural patterns within time-series data to improve label quality without relying on external labeling. The application of SSL to SWR datasets has yielded a 10% increase in classification accuracy. While this improved classification accuracy does not directly enhance our understanding of SWRs, it opens up new pathways for research. The study's findings suggest the transformative capability of SSL in improving data quality across various domains reliant on precise time-series data classification.

Operationalization of the construct “Business model of a Bank”: clustering analyses with deep neural networks

Abstract This paper presents a framework to operationalize the multidimensional construct of a bank’s business model (BBM). We conceptualize the construct from a structural perspective, defining it as its balance sheet’s strategic composition and structure, encompassing asset allocation and funding sources. In contrast to prior research, our study describes the strategic decisions made by bank management as the starting point for the analysis, excluding the results of entrepreneurial activity from the construct’s conceptualization. We analyze balance sheet data for 163 European SSM banks and their subsidiaries—which we call SSM institutions—from 2014 to 2023, sourced from the S&P MI platform’s SNL Financial Institutions database. The study focuses on six balance sheet positions—three from the asset and three from the liability side—expressed as ratios to total assets. We apply a deep autoencoder-based clustering (DAC) model to operationalize the construct and compare the results with the k-means and k-medoids approach. Our empirical analyses identify four BBMs: diversified retail, non-diversified retail, wholesale, and investment-oriented banking. The DAC model leverages the nonlinear capabilities of deep learning, outperforming traditional clustering methods. This paper contributes to the literature on BBMs on a theoretical and technical level. Theoretically, a methodology for operationalizing the construct of a BBM is presented, which can be used to conduct cause-effect analyses. Technically, advanced clustering techniques, including deep learning models, are used to improve classification accuracy and provide new insights into the diversity of banks. The approach presented in this study offers valuable applications for both academics and practitioners in analyzing the impact of BBMs on other constructs, such as performance. Policymakers can leverage this framework to evaluate and guide the development of resilient business models.

Study on lithology identification using a multi-objective optimization strategy to improve integrated learning models: a case study of the Permian Lucaogou Formation in the Jimusaer Depression

Lithology identification is a critical task in logging interpretation and reservoir evaluation, with significant implications for recognizing oil and gas reservoirs. The challenge in shale reservoirs lies in the similar logging response characteristics of different lithologies and the imbalanced data scale, leading to fuzzy lithology classification boundaries and increased difficulty in identification. This study focuses on the shale reservoir of the Permian Lucaogou Formation in the Jimusaer Depression for lithology identification. Initially, a comprehensive sampling model—Smote-Tomek (ST) is used to introduce new feature information into the dataset while removing redundant features, effectively addressing the issue of data imbalance. Then, by combining the multi-objective optimization strategy Artificial Rabbit Optimization (ARO) with the Light Gradient Boosting Machine (LightGBM) model, a new intelligent lithology identification model (ST-ARO-LightGBM) is proposed, aimed at solving the problem of non-optimal hyperparameter settings in the model. Finally, the proposed new intelligent lithology identification model is compared and analyzed with six models: K-Nearest Neighbors (KNN), Decision Tree (DT), Gradient Boosting Decision Tree (GBDT), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and LightGBM, all after comprehensive sampling. The experimental results show that the ST-ARO-LightGBM model outperforms other classification models in terms of classification evaluation metrics for different lithologies, with an overall classification accuracy improvement of 9.13%. The method proposed in this paper can solve the problem of non-equilibrium in rock samples, and can further improve the classification performance of traditional machine learning, and provide a method reference for the lithology classification of shale reservoirs.

Predictive modeling and optimization in dermatology: Machine learning for skin disease classification.

Predictive modeling and optimization in dermatology: Machine learning for skin disease classification.

Application of the Random Forest Algorithm for Accurate Bipolar Disorder Classification

Bipolar disorder (BD) is a complex psychiatric condition characterized by alternating episodes of mania and depression, posing significant challenges for accurate and timely diagnosis. This study explores the use of the Random Forest (RF) algorithm as a machine learning approach to classify patients with BD and healthy controls based on electroencephalogram (EEG) data. A total of 330 participants, including euthymic BD patients and healthy controls, were analyzed. EEG recordings were processed to extract key features, including power in frequency bands and complexity metrics such as the Hurst Exponent, which measures the persistence or randomness of a time series, and the Higuchi’s Fractal Dimension, which is used to quantify the irregularity of brain signals. The RF model demonstrated robust performance, achieving an average accuracy of 93.41%, with recall and specificity exceeding 93%. These results highlight the algorithm’s capacity to handle complex, noisy datasets while identifying key features relevant for classification. Importantly, the model provided interpretable insights into the physiological markers associated with BD, reinforcing the clinical value of EEG as a diagnostic tool. The findings suggest that RF is a reliable and accessible method for supporting the diagnosis of BD, complementing traditional clinical practices. Its ability to reduce diagnostic delays, improve classification accuracy, and optimize resource allocation make it a promising tool for integrating artificial intelligence into psychiatric care. This study represents a significant step toward precision psychiatry, leveraging technology to improve the understanding and management of complex mental health disorders.

Improvement of separation performance of the hydrocyclone with dual enlarged outlets and an arc overflow tube

The application of hydrocyclones can be traced back more than 100 years. Due to their unique advantages, they have been widely used in various separation industries. However, the classification efficiency of hydrocyclones is relatively low due to structural limitations. Therefore, improving the classification performance of hydrocyclones has become a research hotspot and challenge. As critical components, the vortex finder and underflow outlet significantly affect the classification performance of hydrocyclones. This paper designs a series of vortex finder and underflow outlet structures to improve classification accuracy. Numerical analysis and experimental verification methods are used, with pressure field, velocity field, turbulence field, and classification efficiency as evaluation criteria, to investigate the mechanisms underlying the classification performance of hydrocyclones with different structures. Results show that the pressure drop of the Type D (double outlet expansion) hydrocyclone is the smallest, indicating lower energy consumption under identical operating conditions, thereby reducing operating costs. The Type D with an arc vortex finder shifts the LZVV inward, enlarging the separation space for coarse particles while reducing it for fine particles. The circulating flow of the Type D hydrocyclone decreases by 2.26 percentage points compared to the Type A (Base) hydrocyclone, and the short-circuit flow decreases by 4.43 percentage points, resulting in a more stable internal flow field. Type D has a smaller cut size and a higher steepness index, demonstrating stronger cutting ability and higher classification accuracy. Experimental verification shows that the −6 μm underflow content of the Type D hydrocyclone is 13.93 percentage points lower than that of the Type A hydrocyclone, while the −75 μm overflow content decreases by 3.97 percentage points. The issues of fine particles in the underflow and coarse particles in the overflow are effectively mitigated. The data obtained in this study provide theoretical and empirical support for the development of novel hydrocyclone structures.

A Lightweight Network Based on Dynamic Split Pointwise Convolution Strategy for Hyperspectral Remote Sensing Images Classification

For reducing the parameters and computational complexity of networks while improving the classification accuracy of hyperspectral remote sensing images (HRSIs), a dynamic split pointwise convolution (DSPC) strategy is presented, and a lightweight convolutional neural network (CNN), i.e., CSM-DSPCss-Ghost, is proposed based on DSPC. A channel switching module (CSM) and a dynamic split pointwise convolution Ghost (DSPC-Ghost) module are presented by combining the presented DSPC with channel shuffling and the Ghost strategy, respectively. CSM replaces the first expansion pointwise convolution in the MobileNetV2 bottleneck module to reduce the parameter number and relieve the increasing channel correlation caused by the original channel expansion pointwise convolution. DSPC-Ghost replaces the second pointwise convolution in the MobileNetV2 bottleneck module, which can further reduce the number of parameters based on DSPC and extract the depth spectral and spatial features of HRSIs successively. Finally, the CSM-DSPCss-Ghost bottleneck module is presented by introducing a squeeze excitation module and a spatial attention module after the CSM and the depthwise convolution, respectively. The presented CSM-DSPCss-Ghost network consists of seven successive CSM-DSPCss-Ghost bottleneck modules. Experiments on four measured HRSIs show that, compared with 2D CNN, 3D CNN, MobileNetV2, ShuffleNet, GhostNet, and Xception, CSM-DSPCss-Ghost can significantly improve classification accuracy and running speed while reducing the number of parameters.

Glo-net: A dual task branch based neural network for multi-class glomeruli segmentation.

Glo-net: A dual task branch based neural network for multi-class glomeruli segmentation.

Two-stage cascade learning for accurate vegetation coverage estimation in Qilian County using Sentinel-2 satellite data

ABSTRACT Fractional Vegetation Cover (FVC) is an important parameter reflecting vegetation status, accurately grasping the change in vegetation cover can help in evaluating ecosystem quality and maintaining its functions. Based on Sentinel-2 multispectral remote sensing data, this paper firstly proves that the ensemble learning method XGBoost can improve the estimation accuracy (R2 = 0.9179, RMSE = 0.1399, MAE = 0.0609). The factors influencing the estimation of vegetation coverage were discussed in this process. The results show that the FVC estimation model is the best when considering VIRE1, VIRE2, VIRE3, SWIR1, SWIR2, NDVI, SAVI and EVI attributes. On this basis, the two-stage cascade learning for vegetation coverage estimation is put forward to construct the prediction model by combining clustering and ensemble learning. The strategy firstly uses an unsupervised learning algorithm to process the data initially and output the coarse-grained segmentation results, thus reducing the data processing complexity. Then, the supervised learning algorithm is used to fine-tune these rough results and output finer segmentation results. This method combines the advantages of unsupervised and supervised learning to improve classification accuracy, model reliability, and overall performance. Finally, the R2 of the cascade learning model composed of K-means and XGBoost reaches 0.9796, the RMSE drops to 0.0423, and the MAE is 0.0294. By comparing the estimation results of different cascade models, the superiority and applicability of the proposed method were verified.