Self-training Model Research Articles

WC-SBERT: Zero-Shot Topic Classification Using SBERT and Light Self-Training on Wikipedia Categories

In natural language processing (NLP), zero-shot topic classification requires machines to understand the contextual meanings of texts in a downstream task without using the corresponding labeled texts for training, which is highly desirable for various applications. In this article, we propose a novel approach to construct a zero-shot task-specific model called WC-SBERT with satisfactory performance. The proposed approach is highly efficient since it uses light self-training requiring target labels (target class names of downstream tasks) only, which is distinct from other research that uses both the target labels and the unlabeled texts for training. In particular, during the pre-training stage, WC-SBERT uses contrastive learning with multiple negative ranking losses to construct the pre-trained model based on the similarity between Wiki categories. For the self-training stage, online contrastive loss is utilized to reduce the distance between a target label and Wiki categories of similar Wiki pages to the label. Experimental results indicate that compared to existing self-training models, WC-SBERT achieves rapid inference on approximately 6.45 million Wiki text entries by utilizing pre-stored Wikipedia text embeddings, significantly reducing inference time per sample by a factor of 2,746 to 16,746. During the fine-tuning step, the time required for each sample is reduced by a factor of 23–67. Overall, the total training time shows a maximum reduction of 27.5 times across different datasets. Most importantly, our model has achieved state-of-the-art (SOTA) accuracy on two of the three commonly used datasets for evaluating zero-shot classification, namely the AG News (0.84) and Yahoo! Answers (0.64) datasets. The code for WC-SBERT is publicly available on GitHub, 1 and the dataset can also be accessed on Hugging Face. 2

Indoor occupancy monitoring using environmental feature fusion and semi-supervised machine learning models

Smart buildings optimize energy consumption and occupant comfort through Heating, Ventilation and Air Conditioning, and lighting management. Nevertheless, large venues require data fusion techniques to improve analysis and forecasting. This study aims to evaluate the effectiveness of using different feature fusion techniques, environmental sensors, and semi-supervised learning to estimate indoor occupancy in a 230 m2 office. Using five Internet of Things devices measuring air temperature, relative humidity, and barometric pressure, data was collected for 99 days with 6800 entries (on average) and only 14% labeled. Eight feature selection methods were evaluated along with three supervised and two semi-supervised classification methods. Results indicate that the Chi-squared-based approach for feature fusion outperformed others. Similarly, the semi-supervised Self-Training model achieved better performance than the supervised methods. This research shows that combining semi-supervised learning and data fusion allows for estimating the occupancy level in large indoor spaces with high accuracy and low labeling costs. Highlights This study pioneers in exploring semi-supervised learning and distinct feature fusion methods for estimating indoor occupancy levels in a 230 m 2 open office using only Internet of Things (IoT) environmental sensors (air temperature, relative humidity, and barometric pressure). A comprehensive comparison of statistical methods, feature selection, and dimensionality reduction techniques are conducted to determine their ability to generate robust feature fusion sets. The feature fusion selected through the Chi-squared test stood out with a high accuracy F1-score (average of 0.95) and an average accuracy of 0.99. The Self-Training model reached the best performance from semi-supervised learning, with an average F1-Score of 0.90 and an average accuracy of 0.97, based on a dataset with a large proportion of unlabelled data (16,847 entries) and only 9367 labels. For supervised learning, Random Forest achieved a high accuracy (average of 0.98) and F1-score (average of 0.93) across various feature sets.

Publics’ perceptions of legitimacy in corporate social advocacy: A computational analysis of the role of ideological congruence

For corporations engaged in corporate social advocacy (CSA), establishing legitimacy among publics is challenging when they take stands along clear ideological lines on controversial issues. This study examined two questions: (1) how would the congruence between individuals’ political ideologies and corporations’ CSA stances influence perceived CSA legitimacy; and (2) how would individuals’ political ideology strength moderate this relationship? A computational analysis was conducted to examine individual-level ideological congruence and perceived CSA legitimacy in six real-world CSA events. Pairing sentiment analysis with a self-trained model for individual ideology detection, this study cross-analyzed naturalistic data generated by 5181 ordinary users involved in the CSA events on Twitter. It was found that individuals’ perceptions of corporations involved in CSA were highly dependent on the extent to which their own political ideologies were congruent with the corporations’ CSA stances. Also, the influence of ideological congruence was amplified among individuals with stronger political ideologies. However, CSA legitimacy perceptions can be highly situational, with individual events potentially yielding varied results depending on factors such as the issue’s socio-historical context. This study overall demonstrates the important role that individual-level political factors can play in influencing publics’ CSA legitimacy perceptions. This study also highlights the feasibility of using cutting-edge computational methods in assessing publics’ voices on social media as a source of CSA legitimacy.

Development of a Project Based Arduino Uno R3- Assisted Robotic Training Design

This type of research and development aims to create designs and products for robotics training, as well as robotic training design feasibility through expert validation and eval-uation by respondents. Based on a survey of 49 people surveyed, 28.3% said they did not master the microcontroller material, 10.8% said they had learned it very well, and 17.4% said they didn't know. This suggests that a lower level of competence compared to the SKKNI makes robotics training necessary. A self-training model is used for this training. The study uses a branch version of the ADDIE development model, which is carried out in five stages: 1. analysis, 2. design, 3. development, 4. implementation, and 5. evaluation. This media produces starter kits and video learning, The ADDIE method is a research method that runs from the stage of analysis to its evaluation, the analysis researchers must know the gap of competence from students to industry, search for in-tuitive needs, make the need and the gap into the design of training until the last evalua-tion process, this is finally the reason why Addie is used in this research. Students of SMKN 1 Cilegon and students of Professional Education Electrical Engineering Universi-tas Sultan Ageng Tirtayasa were used in this research. The analytical techniques used are qualitative and quantitative descriptive analysis. Evaluation results from starter kits, learning videos, and learning materials are used to determine the level of eligibility for robotic training designs. The results of the starter kit evaluation showed that the media starter kit was in the category "Very Fit" with an average score of 40.7 with a range x33. In the respondent assessment, it was known that the average rating of the nine-teen respondents was 60.3 with the range x 51, suggesting that the Arduino Uno R3 project-based robotic training design containing a starter kit, video learning modules, and learning materials belonged to the category of "Very Feasible." In other words, this robotics training deserves to be done.

Civil and Commercial Jurisprudence Observations on Intelligent Social Governance

Abstract The purpose of civil and commercial jurisprudence observation is to realize the intelligent and innovative development of social governance and to improve the efficiency of related cases. This paper develops a civil and commercial jurisprudence case information analysis system that utilizes MVC architecture and intelligent technology. The BERT pre-training model and self-training model are combined for the construction of a civil and commercial jurisprudence legal information extraction model, which helps judicial personnel understand the scope of civil and commercial jurisprudence cases in the process of social governance. Then, the ILJP is used as the basis to combine the fact encoder, the element predictor and the legal cause predictor to carry out the construction of the interpretable legal cause intelligent research and judgment model, which provides an intelligent research and judgment program for the civil and commercial jurisprudence social governance cases. The effectiveness of the civil and commercial law case analysis system is an important guarantee to enhance intelligent social governance, so this paper carries out a test and analysis of its intelligent algorithm and system performance. The results show that the legal information extraction model tends to stabilize after the first 15 iterations of training, and its validation loss value is around 0.173, and the training loss value of the model is around 0.008 when the training of the intelligent research and judgment model reaches 200 rounds. When the information analysis system of civil and commercial law cases reaches 1,000 concurrent users, the average response time of each logical operation does not exceed 3s, and the occupancy rate of CPU and memory resources is over 30%. The information age needs to fully combine technology and social governance in order to deeply mine the information of civil and commercial jurisprudence related cases, improve the efficiency of social governance, and innovate the intelligent development of social governance.

Customization of a deep neural network using local data for seismic phase picking

Deep-learning (DL) pickers have demonstrated superior performance in seismic phase picking compared to traditional pickers. DL pickers are extremely effective in processing large amounts of seismic data. Nevertheless, they encounter challenges when handling seismograms from different tectonic environments or source types, and even a slight change in the input waveform can considerably affect their consistency. Here, we fine-tuned a self-trained deep neural network picker using a small amount of local seismic data (26,875 three-component seismograms) recorded by regional seismic networks in South Korea. The self-trained model was developed using publicly available waveform datasets, comprising over two million three-component seismograms. The results revealed that the Korean-fine-tuned phase picker (KFpicker) effectively enhanced picking quality, even when applied to data that were not used during the fine-tuning process. When compared to the performance of the pre-trained model, this improvement was consistently observed regardless of variations in the positions of seismic phases in the input waveform, Furthermore, when the KFpicker predicted the phases for overlapping input windows and used the median value of probabilities as a threshold for phase detection, a considerable decrease was observed in the number of false picks. These findings indicate that fine-tuning a deep neural network using a small amount of local data can improve earthquake detection in the region of interest, while careful data augmentation can enhance the robustness of DL pickers against variations in the input window. The application of KFpicker to the 2016 Gyeongju earthquake sequence yielded approximately twice as many earthquakes compared to previous studies. Consequently, detailed and instantaneous statistical parameters of seismicity can be evaluated, making it possible to assess seismic hazard during an earthquake sequence.

Variety-aware GAN and online learning augmented self-training model for knowledge graph entity alignment

Recently, self-training strategies are adopted in some entity alignment methods, which address the scarcity of training data by selecting newly-aligned pairs from the predicted alignment of unlabeled data. However, the boundary between positive and negative pairs in predicted alignment is difficult to determine, which may lead to inappropriate alignment in newly-aligned pairs. Besides, some pre-aligned pairs have been fitted by the method during training iteration, and combining all pre-aligned pairs with newly-aligned pairs to retrain the method may result in overfitting problems. To address these problems, a Self-training Entity Alignment Framework based on Variety-aware GAN and Online Learning Algorithm named SEAGAN is proposed in this paper. To select reliable newly-aligned pairs from the predicted alignment, a variety-aware GAN with a metric of match variety that eliminates negative pairs differing significantly from positive pairs is designed. It leverages the distribution of entity pairs to determine the boundary between different types of pairs. Moreover, SEAGAN designs an online learning algorithm that combines newly-aligned pairs with their one-hop and two-hop neighbor entities in pre-aligned pairs to update model parameters, which alleviates overfitting problems. We conduct extensive experiments on four real-world datasets to compare SEAGAN with fifteen state-of-the-art methods. Experiment results show that SEAGAN has better performance than state-of-the-art methods on metrics of Prec@1, Prec@5, and MRR.

Pseudo-label self-training model for transfer learning algorithm

When aligning joint distributions between domains, the existing transfer learning algorithms usually assign pseudo labels due to the lack of labels in target domain. However, the noise in pseudo labels will affect the performance of transfer learning. Pseudo-label self-training for transfer learning (PST-TL) model is proposed to generate reliable pseudo labels for target domain and have a wide range of applications in existing algorithms. Pseudo labels are predicted by an ensemble classifier using absolute majority vote, and labels predicted successfully are considered to be high confidence. The training of ensemble classifier applies the self-training of joint pseudo labels strategy, adding strongly stable data to training set of the classifier. The semi-supervised and unsupervised transfer learning tasks in experiment show that the existing transfer learning algorithm can significantly improve the transfer performance after embedded by PST-TL model.

Animal Species Recognition with Deep Convolutional Neural Networks from Ecological Camera Trap Images.

Accurate identification of animal species is necessary to understand biodiversity richness, monitor endangered species, and study the impact of climate change on species distribution within a specific region. Camera traps represent a passive monitoring technique that generates millions of ecological images. The vast numbers of images drive automated ecological analysis as essential, given that manual assessment of large datasets is laborious, time-consuming, and expensive. Deep learning networks have been advanced in the last few years to solve object and species identification tasks in the computer vision domain, providing state-of-the-art results. In our work, we trained and tested machine learning models to classify three animal groups (snakes, lizards, and toads) from camera trap images. We experimented with two pretrained models, VGG16 and ResNet50, and a self-trained convolutional neural network (CNN-1) with varying CNN layers and augmentation parameters. For multiclassification, CNN-1 achieved 72% accuracy, whereas VGG16 reached 87%, and ResNet50 attained 86% accuracy. These results demonstrate that the transfer learning approach outperforms the self-trained model performance. The models showed promising results in identifying species, especially those with challenging body sizes and vegetation.

Spanish Corpora of tweets about COVID-19 vaccination for automatic stance detection

The paper presents new annotated corpora for performing stance detection on Spanish Twitter data, most notably Health-related tweets. The objectives of this research are threefold: (1) to develop a manually annotated benchmark corpus for emotion recognition taking into account different variants of Spanish in social posts; (2) to evaluate the efficiency of semi-supervised models for extending such corpus with unlabelled posts; and (3) to describe such short text corpora via specialised topic modelling.A corpus of 2,801 tweets about COVID-19 vaccination was annotated by three native speakers to be in favour (904), against (674) or neither (1,223) with a 0.725 Fleiss’ kappa score. Results show that the self-training method with SVM base estimator can alleviate annotation work while ensuring high model performance. The self-training model outperformed the other approaches and produced a corpus of 11,204 tweets with a macro averaged f1 score of 0.94. The combination of sentence-level deep learning embeddings and density-based clustering was applied to explore the contents of both corpora. Topic quality was measured in terms of the trustworthiness and the validation index.

Automatic Evaluation of Neural Network Training Results

This article is dedicated to solving the problem of an insufficient degree of automation of artificial neural network training. Despite the availability of a large number of libraries for training neural networks, machine learning engineers often have to manually control the training process to detect overfitting or underfitting. This article considers the task of automatically estimating neural network training results through an analysis of learning curves. Such analysis allows one to determine one of three possible states of the training process: overfitting, underfitting, and optimal training. We propose several algorithms for extracting feature descriptions from learning curves using mathematical statistics. Further state classification is performed using classical machine learning models. The proposed automatic estimation model serves to improve the degree of automation of neural network training and interpretation of its results, while also taking a step toward constructing self-training models. In most cases when the training process of neural networks leads to overfitting, the developed model determines its onset ahead of the early stopping method by 3–5 epochs.

Constructing a high-performance self-training model based on support vector classifiers to detect gold mineralization-related geochemical anomalies for gold exploration targeting

In an area where mineral deposits have been discovered, both anomaly detection algorithms and supervised classification algorithms can be adopted to detect mineralization-related geochemical anomalies for mineral exploration targeting. However, anomaly detection algorithms themselves cannot use known mineral deposits as the supervisors to improve the performance of geochemical anomaly detection models. Supervised classification algorithms can not properly deal with the extreme class-imbalance of geochemical exploration data in the establishment of classification models. Therefore, neither anomaly detection models nor supervised classification models perform well in the detection of mineralization-related geochemical anomalies. In order to obtain a high-performance mineral exploration targeting model, the self-training algorithm was adopted to construct a self-training model based on support vector classifiers to detect gold mineralization-related geochemical anomalies in the Chengde area in Hebei Province (China). The self-training model was compared with the support vector classification model, logistic regression model and five dictionary learning models in terms of area under the curve (AUC) and lift index. The AUC value of the self-training model (0.89) is much higher than those of the two supervised classification models (0.81–0.82) and five anomaly detection models (0.79–0.81). The lift index of the self-training model (11.87) is also much higher than those of the two supervised classification models (2.77–4.04) and five anomaly detection models (2.64–3.48). Therefore, the self-training model performs much better than the two supervised classification models and five anomaly detection models in the detection of gold mineralization-related geochemical anomalies. The gold mineralization-related geochemical anomalies detected by the self-training model highly conform to the metallogenic and geological features in the study area. Therefore, it is a prospective method to establish a self-training model to detect mineralization-related geochemical anomalies in an area where mineral deposits have been discovered.

A Semi-Supervised Machine Learning Approach in Predicting High-Risk Pregnancies in the Philippines.

Early risk tagging is crucial in maternal health, especially because it threatens both the mother and the long-term development of the baby. By tagging high-risk pregnancies, mothers would be given extra care before, during, and after pregnancies, thus reducing the risk of complications. In the Philippines, where the fertility rate is high, especially among the youth, awareness of risks can significantly contribute to the overall outcome of the pregnancy and, to an extent, the Maternal mortality rate. Although supervised machine learning models have ubiquity as predictors, there is a gap when data are weak or scarce. Using limited collected data from the municipality of Daraga in Albay, the study first compared multiple supervised machine learning algorithms to analyze and accurately predict high-risk pregnancies. Through hyperparameter tuning, supervised learning algorithms such as Decision Tree, Random Forest, Support Vector Machine, K-Nearest Neighbors, Naïve Bayes, and Multilayer Perceptron were evaluated by using 10-fold cross validation to obtain the best parameters with the best scores. The results show that Decision Tree bested other algorithms and attained a test score of 93.70%. To address the gap, a semi-supervised approach using a Self-Training model was applied to the modified Decision Tree, which was then used as the base estimator with a 30% unlabeled dataset and achieved a 97.01% accuracy rate which outweighs similar studies.

REAL TIME YOGA POSE DETECTION USING DEEPLEARNING: A REVIEW

With the increase in the number of yoga practitioners every year, the risk of injuries as a result of incorrect yoga postures has also increased. A selftraining model that can evaluate the posture of individuals is the optimal solution for this issue. This objective can be attained with the aid of computer vision and deep learning. A model that can detect theyoga pose performed by an individual, evaluate it in comparison to the pose performed by an expert, and provide the individual with instructive feedback would be an effective solution to this problem. Recently, numerous researchers have conducted experiments on the detection and performance of yoga poses in real time. This paper discusses the methods undertaken in brief and compares the tools and algorithms they used for conducting pose estimation, pose detection as well as pose assessment. Itdiscusses the accuracy, precision, and similarity of pose classification obtained by the researchers and the future scope of the research.

Based on Quantum Topological Stabilizer Color Code Morphism Neural Network Decoder

Solving for quantum error correction remains one of the key challenges of quantum computing. Traditional decoding methods are limited by computing power and data scale, which restrict the decoding efficiency of color codes. There are many decoding methods that have been suggested to solve this problem. Machine learning is considered one of the most suitable solutions for decoding task of color code. We project the color code onto the surface code, use the deep Q network to iteratively train the decoding process of the color code and obtain the relationship between the inversion error rate and the logical error rate of the trained model and the performance of error correction. Our results show that through unsupervised learning, when iterative training is at least 300 times, a self-trained model can improve the error correction accuracy to 96.5%, and the error correction speed is about 13.8% higher than that of the traditional algorithm. We numerically show that our decoding method can achieve a fast prediction speed after training and a better error correction threshold.

Spatial Disaggregation of Population Subgroups Leveraging Self-Trained Multi-Output Gradient Boosting Regression Trees

Abstract. Accurate and consistent estimations on the present and future population distribution, at fine spatial resolution, are fundamental to support a variety of activities. However, the sampling regime, sample size, and methods used to collect census data are heterogeneous across temporal periods and/or geographic regions. Moreover, the data is usually only made available in aggregated form, to ensure privacy. In an attempt to address these issues, several previous initiatives have addressed the use of spatial disaggregation methods to produce high-resolution gridded datasets describing the human population distribution, although these projects have usually not addressed specific population subgroups. This paper describes a spatial disaggregation method based on self-training regression models, innovating over previous studies in the simultaneous prediction of disaggregated counts for multiple inter-related variables, by leveraging multi-output models based on gradient tree boosting. We report on experiments for two case studies, using high-resolution data (i.e., counts for different subgroups available at a resolution of 100 meters) for the municipality of Amsterdam and the region of Greater Copenhagen. Results show that the proposed approach can capture spatial heterogeneity and the dependency on local factors, outperforming alternatives (e.g., seminal disaggregation algorithms, or approaches leveraging individual regression models for each variable) in terms of averaged error metrics, and also upon visual inspection of spatial variation in the resulting maps.

An Improved Self-Training Model with Fine-Tuning Teacher/Student Exchange Strategy to Detect Computer-Generated Images

Computer-generated (CG) images have become indistinguishable from natural images due to powerful image rendering technology. Fake CG images have brought huge troubles to news media, judicial forensics, and other fields. How to detect CG image has become a key point to solve the problems mentioned above. The image classification method based on deep learning, due to its strong self-learning ability, can automatically determine the differences in the image features between CG images and natural images and can be used to detect CG images. However, deep learning often requires a large amount of labeled data, which is usually a tedious and complex task. This paper proposes an improved self-training strategy with fine-tuning teacher/student exchange (FTTSE) to solve the problem of missing labeled datasets. Our method is actually a strategy based on semisupervised learning to train the teacher model through labeled data and to predict the unlabeled data by the teacher model to generate pseudo labels. The student model is obtained by continuous training on the mixed dataset composed of labeled and pseudo-labeled data. A teacher/student exchange strategy is designed for iterative training; i.e., the identities of the teacher model and the student model are exchanged at the beginning of each round of iteration. And then the new teacher model is used to predict pseudo labels, and the new student model exchanged from teacher model in the previous round of iteration is fine-tuned and retrained by the mixed dataset with new pseudo labels. Furthermore, we introduced malicious image attacks to perturb the mixed dataset to improve the robustness of the student model. The experimental results show that the improved self-training model we proposed can stably maintain the image classification ability even if the testing images are maliciously attacked. After iterative training, the CG image detection accuracy of the final model increases by 5.18%. The robustness against 100% malicious attacks is also improved, where the final trained model has an accuracy improvement of 7.63% higher than the initial model. The self-training model with FTTSE strategy proposed in this paper can effectively enhance the detection ability of the existing model and can greatly improve the robustness of the model with iterative training.

A Self-Trained Model for Cloud, Shadow and Snow Detection in Sentinel-2 Images of Snow- and Ice-Covered Regions

Screening clouds, shadows, and snow is a critical pre-processing step in many remote-sensing data processing pipelines that operate on satellite image data from polar and high mountain regions. We observe that the results of the state-of-the-art Fmask algorithm are not very accurate in polar and high mountain regions. Given the unavailability of large, labeled Sentinel-2 training datasets, we present a multi-stage self-training approach that trains a model to perform semantic segmentation on Sentinel-2 L1C images using the noisy Fmask labels for training and a small human-labeled dataset for validation. At each stage of the proposed iterative framework, we use a larger network architecture in comparison to the previous stage and train a new model. The trained model at each stage is then used to generate new training labels for a bigger dataset, which are used for training the model in the next stage. We select the best model during training in each stage by evaluating the multi-class segmentation metric, mean Intersection over Union (mIoU), on the small human-labeled validation dataset. This effectively helps to correct the noisy labels. Our model achieved an overall accuracy of 93% compared to the Fmask 4 and Sen2Cor 2.8, which achieved 75% and 76%, respectively. We believe our approach can also be adapted for other remote-sensing applications for training deep-learning models with imprecise labels.

A Comparative Analysis of Semi-Supervised Learning in Detecting Burst Header Packet Flooding Attack in Optical Burst Switching Network

This paper presents a comparative analysis of four semi-supervised machine learning (SSML) algorithms for detecting malicious nodes in an optical burst switching (OBS) network. The SSML approaches include a modified version of K-means clustering, a Gaussian mixture model (GMM), a classical self-training (ST) model, and a modified version of self-training (MST) model. All the four approaches work in semi-supervised fashion, while the MST uses an ensemble of classifiers for the final decision making. SSML approaches are particularly useful when a limited number of labeled data is available for training and validation of the classification model. Manual labeling of a large dataset is complex and time consuming. It is even worse for the OBS network data. SSML can be used to leverage the unlabeled data for making a better prediction than using a smaller set of labelled data. We evaluated the performance of four SSML approaches for two (Behaving, Not-behaving), three (Behaving, Not-behaving, and Potentially Not-behaving), and four (No-Block, Block, NB- wait and NB-No-Block) class classifications using precision, recall, and F1 score. In case of the two-class classification, the K-means and GMM-based approaches performed better than the others. In case of the three-class classification, the K-means and the classical ST approaches performed better than the others. In case of the four-class classification, the MST showed the best performance. Finally, the SSML approaches were compared with two supervised learning (SL) based approaches. The comparison results showed that the SSML based approaches outperform when a smaller sized labeled data is available to train the classification models.

A semi-supervised learning method for hyperspectral imagery based on self-training and local-based affinity propagation

ABSTRACT In hyperspectral remote sensing, the classification of hyperspectral imagery is an important issue of concern. However, obtaining sufficient labelled samples for the classification is hard work and consumes enormous labour and material resources. To solve this problem, many scholars make their efforts to the semi-supervised learning method. The self-training model is a popular semi-supervised learning method, which trains a classifier using a few labelled samples and many unlabelled samples in an iterative manner. Nevertheless, the self-training model may provide misclassification as adding incorrect unlabelled samples. This paper proposes a novel semi-supervised learning method, which attempts to introduce the affinity propagation and spatial constraint to self-training model. In the proposed algorithm, a self-training model is employed to expand the labelled sample set. The affinity propagation is applied to select a group of reliable unlabelled samples. To further protect the reliability of unlabelled samples, the selection process of self-training model is constrained in the spatial local area of the labelled samples. In the iteration, the reliable unlabelled samples are selected and added to the labelled sample set. After that, the classifier is trained by the updated labelled samples. The algorithm finishes as the iterations exceed the predetermined value or the labelled sample set is unchanged. Different from the traditional self-training-based semi-supervised learning method, there are mainly three improvements. First, the selection of reliable unlabelled samples is conducted by the affinity propagation, which can effectively improve the reliability of the unlabelled samples. Second, the spatial constraint is introduced to restrict the selection process of self-training model, which can protect the reliability of unlabelled samples and alleviate the computational complexity. Finally, the spectral correlation angle is applied to construct the similarity matrix of affinity propagation, so that the algorithm is more suitable for hyperspectral imagery. A comparative analysis was conducted with classical self-training semi-supervised algorithms to test the effectiveness of the method. It was found that the proposed method provided higher overall accuracy, average accuracy and kappa coefficient (κ) than the competitors.