Initial Training Samples Research Articles

From local to global: A transfer learning-based approach for mapping poplar plantations at national scale using Sentinel-2

Reliable estimates of poplar plantations area are not available at the French national scale due to the unsuitability and low update rate of existing forest databases for this short-rotation species. While supervised classification methods have been shown to be highly accurate in mapping forest cover from remotely sensed images, their performance depends to a great extent on the labelled samples used to build the models. In addition to their high acquisition cost, such samples are often scarce and not fully representative of the variability in class distributions. Consequently, when classification models are applied to large areas with high intra-class variance, they generally yield poor accuracies because of data shift issues. In this paper, we propose the use of active learning to efficiently adapt a classifier trained on a source image to spatially distinct target images with minimal labelling effort and without sacrificing the classification performance. The adaptation consists in actively adding to the initial local model new relevant training samples from other areas in a cascade that iteratively improves the generalisation capabilities of the classifier leading to a global model tailored to these different areas. This active selection relies on uncertainty sampling to directly focus on the most informative pixels for which the algorithm is the least certain of their class labels. Experiments conducted on Sentinel-2 time series revealed their high capacity to identify poplar plantations at a local scale with an average F-score ranging from 89.5% to 99.3%. For large area adaptation, the results showed that when the same number of training samples was used, active learning outperformed random sampling by up to 5% of the overall accuracy and up to 12% of the class F-score. Additionally, and depending on the class considered, the random sampling model required up to 50% more samples to achieve the same performance of an active learning-based model. Moreover, the results demonstrate the suitability of the derived global model to accurately map poplar plantations among other tree species with overall accuracy values up to 14% higher than those obtained with local models. The proposed approach paves the way for a national scale mapping in an operational context.

A Low Latency On-Body Typing System through Single Vibration Sensor

Nowadays, smart wristbands have become one of the most prevailing wearable devices, as they are small and portable. However, due to the limited size of the touch screens, smart wristbands typically have poor interactive experience. There are a few works appropriating the human body as a surface to type on. Yet, by using multiple sensors at high sampling rates, they are not portable and are energy-consuming in practice. To break this stalemate, we proposed a portable, cost efficient text-entry system, termed ViType, which first leverages a single small form factor sensor to achieve a practical user input with much lower sampling rates. To enhance the input accuracy with less vibration information introduced by lower sampling rates, ViType designs a set of novel mechanisms, including a fine-grained feature extraction to process the vibration signals, and a runtime calibration and adaptation scheme to recover from the error due to temporal instability. Extensive experiments have been conducted on 30 human subjects. The results demonstrate that ViType is robust against various confounding factors. The average recognition accuracy is 95 percent with an initial training sample size of 20 for each key. The accuracy is 1.54 times higher than the state-of-the-art on-body typing system. Furthermore, when turning on the runtime calibration and adaptation system to update and enlarge the training sample size, the accuracy can reach around 98 percent on average during one month.

Developing a Soft Sensor for MTBE Process Based on a Small Sample

The paper discusses the development of a soft sensor (SS) for an MTBE plant in the cases when the training sample either is small or does not comprise the whole of quality range because of process plant non-stationarity as well as the difficulty and high cost of retrieving more information. In order to build a soft sensor that provides higher accuracy in estimating the quality of the output product, an algorithm for extending the initial training sample using a rigorous model of the distillation column of the methyl-tert-butyl ether production under condition of exactly unknown values of the Murphree trays efficiency has been proposed. The resulting SS enables 40% improvement of product quality prediction.

Investigation of shear strength correlations and reliability assessments of sandwich structures by kriging method

Steel-concrete-steel (SCS) sandwich composite structure with corrugated-strip connectors (CSC) is the promising structure which is applied in offshore and building structures. The behavior prediction of shear connections is of major importance in SCS structures. The present study evaluated the existing shear strength correlations of SCS sandwich structures exploiting experimental data and Finite Element Analysis (FEA). The considered system is a double steel skin sandwich structure with CSC (DSCS). Due to the limitation of the literature regarding CSC development, some new correlations were proposed and studied relying on several FEA results through the Genetic Programming method. The accuracy of the estimated shear strength predicted by the existing and proposed equations was evaluated using the FEA data and push-out test results. The FE models were verified through experimental data. Moreover, the correlations were investigated based on reliability assessment due to the high importance of the reliability analysis of such structures. Given that high accuracy in estimating the shear strength fails to necessarily lead to acceptable results in structural reliability analysis, the reliability of the existing and proposed equations was evaluated using the Kriging model by considering experimental data. This meta-model could predict accurate values with a limited number of initial training samples.

An Iterative Training Sample Updating Approach for Domain Adaptation in Hyperspectral Image Classification

Acquiring training samples in remote sensing images is always expensive and time-consuming. As a consequence, it would be preferable if one domain without training samples (the target domain) could be classified given <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">a priori</i> knowledge from another domain (the source domain). In this letter, an iterative training sample updating (ITSU) approach is proposed based on <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">a posteriori</i> spatial feature extraction. First, the classifier is trained with initial training samples from the source domain and applied to the target domain, producing a preclassification map. Then, as an invariant feature, the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">a posteriori</i> spatial features are extracted with a guided filter. Based on the spectral features and the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">a posteriori</i> spatial features, a criterion measuring the similarity of the cross-domain samples is defined. New training samples from the target domain are assigned with pseudo-labels, and the original samples in the source domain are removed. Furthermore, the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">a posteriori</i> spatial feature maps are fed back to the input images, and new classifiers are trained with an updated training sample set in the updated feature space. This procedure is repeated until the stopping rule is satisfied. Finally, the adapted classifier is obtained based on the updated training samples. The experimental results on three hyperspectral data sets indicated that ITSU achieved the best performance compared with the other two state-of-the-art methods.

Iterative Training Sample Expansion to Increase and Balance the Accuracy of Land Classification From VHR Imagery

Imbalanced training sets are known to produce suboptimal maps for supervised classification. Therefore, one challenge in mapping land cover is acquiring training data that will allow classification with high overall accuracy (OA) in which each class is also mapped onto similar user's accuracy. To solve this problem, we integrated local adaptive region and box-and-whisker plot (BP) techniques into an iterative algorithm to expand the size of the training sample for selected classes in this article. The major steps of the proposed algorithm are as follows. First, a very small initial training sample (ITS) for each class set is labeled manually. Second, potential new training samples are found within an adaptive region by conducting local spectral variation analysis. Lastly, three new training samples are acquired to capture information regarding intraclass variation; these samples lie in the lower, median, and upper quartiles of BP. After adding these new training samples to the ITS, classification is retrained and the process is continued iteratively until termination. The proposed approach was applied to three very high-resolution (VHR) remote-sensing images and compared with a set of cognate methods. The comparison demonstrated that the proposed approach produced the best result in terms of OA and exhibited superiority in balancing user's accuracy. For example, the proposed approach was typically 2%-10% more accurate than the compared methods in terms of OA and it generally yielded the most balanced classification.

Initial training samples selection using clustering algorithm for audio recognition

The quantity and quality of training samples are one of the most critical factors for the performance of the classifier, whatever traditional classification method or deep learning algorithm. To improve the quality of the training samples with less manual tagging effort, this paper proposed a representative samples selection method based on clustering algorithm to choose more representative samples supplied to human to annotate. These representative ones can describe class distribution better than samples annotated at random. On the one hand, these selected samples can train a more generalized classifier. On the other hand, this samples selection method can save expensive annotation work which is thought a lot by people now. The results of experiments in melodrama Friends corpus show better classification performance compared to traditional annotation methods.

Kinematic parameter optimization of a flapping ellipsoid wing based on the data-informed self-adaptive quasi-steady model

This paper constructs an optimization framework based on the data-informed self-adaptive quasi-steady model. The framework aims at achieving a specific aerodynamic force coefficient by optimizing the kinematic parameters of the flapping motion of an ellipsoid wing. All the model coefficients of this quasi-steady model are calibrated empirically by the data-informed training. At each optimization iteration, the data-informed training is implemented by the local ridge regression, where the initial training samples are extracted from simulation examples, and the weight coefficients are calculated by the compactly supported radial basis function with the previous optimal solution as the center point. Furthermore, a numerical simulation is conducted to evaluate the accurate aerodynamic force coefficient corresponding to the current optimal solution. The relative error between the accurate simulation result and optimization objective is calculated as the convergence criteria of the optimization. Then, the effects of the kinematic parameters on the time-averaged lift coefficient are first investigated, which indicate that the in-phase flapping with high flapping angle amplitude and medium geometric angle of attack amplitude is beneficial to the lift coefficient. Moreover, the kinematic optimization is conducted for a three-dimensional flapping ellipsoid wing in the hovering mode. The results demonstrate that the leading-edge vortex is crucial for the force generation. Moreover, in one flapping period, the asymmetrical wake and two unequal lift coefficient peaks emerge under the figure-O motion pattern while the vortex structures are highly symmetrical under the figure-8 motion pattern.

OIC-MCE: A Practical Land Cover Mapping Approach for Limited Samples Based on Multiple Classifier Ensemble and Iterative Classification

Land cover samples are usually the foundation for supervised classification. Unfortunately, for land cover mapping in large areas, only limited samples can be used due to the time-consuming and labor-intensive sample collection. A novel and practical Object-oriented Iterative Classification method based on Multiple Classifiers Ensemble (OIC-MCE) was proposed in this paper. It systematically integrated object-oriented segmentation, Multiple Classifier Ensemble (MCE), and Iterative Classification (IC). In this method, the initial training samples were updated self-adaptively during the iterative processes. Based on these updated training samples, the inconsistent regions (ICR) in the classification results of the MCE method were reclassified to reduce their uncertainty. Three typical case studies in the China-Pakistan Economic Corridor (CPEC) indicate that the overall accuracy of the OIC-MCE method is significantly higher than that of the single classifier. After five iterations, the overall accuracy of the OIC-MCE approach increased by 5.58%–8.38% compared to the accuracy of the traditional MCE method. The spatial distribution of newly added training samples generated by the OIC-MCE approach was relatively uniform. It was confirmed by ten repeated experiments that the OIC-MCE approach has good stability. More importantly, even if the initial sample size reduced by 65%, the quality of the final classification result based on the proposed OIC-MCE approach would not be greatly affected. Therefore, the proposed OIC-MCE approach provides a new solution for land cover mapping with limited samples. Certainly, it is also well suited for land cover mapping with abundant samples.

Using Training Samples Retrieved from a Topographic Map and Unsupervised Segmentation for the Classification of Airborne Laser Scanning Data

The labeling of point clouds is the fundamental task in airborne laser scanning (ALS) point clouds processing. Many supervised methods have been proposed for the point clouds classification work. Training samples play an important role in the supervised classification. Most of the training samples are generated by manual labeling, which is time-consuming. To reduce the cost of manual annotating for ALS data, we propose a framework that automatically generates training samples using a two-dimensional (2D) topographic map and an unsupervised segmentation step. In this approach, input point clouds, at first, are separated into the ground part and the non-ground part by a DEM filter. Then, a point-in-polygon operation using polygon maps derived from a 2D topographic map is used to generate initial training samples. The unsupervised segmentation method is applied to reduce the noise and improve the accuracy of the point-in-polygon training samples. Finally, the super point graph is used for the training and testing procedure. A comparison with the point-based deep neural network Pointnet++ (average F1 score 59.4%) shows that the segmentation based strategy improves the performance of our initial training samples (average F1 score 65.6%). After adding the intensity value in unsupervised segmentation, our automatically generated training samples have competitive results with an average F1 score of 74.8% for ALS data classification while using the ground truth training samples the average F1 score is 75.1%. The result shows that our framework is feasible to automatically generate and improve the training samples with low time and labour costs.

Automated detection of impervious surfaces using night-time light and Landsat images based on an iterative classification framework

ABSTRACTA novel workflow for automated detecting of impervious surface by using night-time light and Landsat images at the individual city scale is proposed. This approach is composed by of three steps. In the beginning, urban, peri-urban and rural regions are detected from the night-time light image by a contour line algorithm. Then, using Landsat TM image, region-specific spectral index analysis is employed to generate initial training samples of urban land covers. Finally, an iterative classification framework is applied to select new training samples by integrating spectral and spatial information and to obtain the final mapping result. Experimental results of two cities show that the proposed method produces higher classification accuracy than the ones using the manual-sampling methods. Moreover, further validations suggested that the spatial information is able to effectively increase the producer’s accuracy of impervious surface. This automated approach is potentially important for large-scale regional impervious surface mapping and application.

Partial Adversarial Training for Neural Network-Based Uncertainty Quantification

Currently available uncertainty quantification (UQ) neural networks (NNs) are trained through the statistical error minimization. Therefore, NNs perform poorly for critical input patterns. Some input patterns have lower coverage probabilities than others. Such input dependent performance is evident in electricity price prediction where different input features are coming from heterogeneous monitoring sources. This paper proposes a prediction interval (PI) based UQ of the electricity price with the proposed partial adversarial training to achieve the input-domain independent performance. The proposed training consists of initial training, adversarial sample generations from critical samples and a final training with the combined datasets of critical samples and initial training samples. Critical situations are situations where prediction systems struggle to predict and make higher statistical errors. Multiple NNs are first trained with different initializations. Each uncovered sample to the NN pair in the training set generates an adversarial sample. The adversarial dataset is concatenated with the initial samples. The final NN training is performed with the combined dataset. The technique is visualized with rough sketches in both time and input domain. The feasibility and performance are examined on experimental electricity market price datasets.

Estimation of the complexity of constructing a logical classification tree for an arbitrary case in conditions of strong class separation of the initial training sample

The paper offers an estimation of the complexity of the constructed logical tree structure for classifying an arbitrary case in the conditions of a strong class division of the initial training sample. The principal solution to this question is of a defining nature, regarding the assessment of the structural complexity of classification models (in the form of tree-like structures of LCT/ACT) of discrete objects for a wide range of applied classification and recognition problems in terms of developing promising schemes and methods for their final optimization (minimization) of post-pruning structure. The presented research is relevant not only for constructions (structures) of logical classification trees, but also allows us to extend the scheme of complexity estimation to the General case of algorithmic structures (ACT models) of classification trees (the concept of algorithm trees and trees of generalized features - TGF). Is investigated the actual question of the concept of decision trees (tree recognition) – evaluation of the maximum complexity of the General scheme of constructing a logical tree based classification procedure of stepwise selection of sets of elementary features (they can be diverse sets and combinations) that for given initial training sample (array of discrete information) builds a tree structure (classification model), from a set of elementary features (basic attributes) are estimated at each stage of the scheme of the model in this sample for the case of strong separation of classes. Modern information systems and technologies based on mathematical approaches (models) of pattern recognition (structures of logical and algorithmic classification trees) are widely used in socio-economic, environmental and other systems of primary analysis and processing of large amounts of information, and this is due to the fact that this approach allows you to eliminate a set of existing disadvantages of well-known classical methods, schemes and achieve a fundamentally new result. The research is devoted to the problems of classification tree models (decision trees), and offers an assessment of the complexity of logical tree structures (classification tree models), which consist of selected and ranked sets of elementary features (individual features and their combinations) built on the basis of the General concept of branched feature selection. This method, when forming the current vertex of the logical tree (node), provides the selection of the most informative (qualitative) elementary features from the source set. This approach allows you to significantly reduce the size and complexity of the tree (the total number of branches and tiers of the structure) and improve the quality of its subsequent instrumental analysis (the final decomposition of the model).

Estimation of general complexity of the procedure for constructing a binary logical classification tree for an arbitrary case

We propose an upper estimate of the complexity of the binary logical tree synthesis procedure for classifying an arbitrary case (for conditions of weak and strong separation of classes in the training sample). The solution to this question is of a fundamental nature, regarding the assessment of the structural complexity of classification models (in the form of tree structures) of discrete objects for a wide range of applied classification and recognition problems in terms of developing promising schemes and methods for their final optimization (minimization) of the structure. This research is relevant not only for the constructions of logical classification trees, but also allows us to extend the complexity estimation scheme itself to the general case of algorithmic structures of classification trees (concepts of algorithm trees and generalized feature trees). The current issue of complexity of the general procedure for constructing a logical classification tree based on the concept of step-by-step selection of sets of elementary features (their possible heterogeneous sets and combinations), which for a given initial training sample (an array of discrete information) builds a tree structure (classification model), from a set of elementary features (basic attributes) evaluated at each stage of the model construction scheme for this sample. Thus, modern information technologies based on mathematical models of pattern recognition (logical and algorithmic classification trees) are widely used in socio-economic, environmental and other systems of primary analysis and processing of large amounts of information. This is due to the fact that this approach allows you to eliminate a set of existing disadvantages of well-known classical methods and schemes and achieve a fundamentally new result. The work is devoted to the problems of classification tree models (decision trees), and offers an assessment of the complexity of logical tree structures (classification tree models), which consist of selected and ranked sets of elementary features built on the basis of the General concept of branched feature selection. This method, when forming the current vertex of the logical tree (node), provides the selection of the most informative (qualitative) elementary features from the source set. This approach allows you to significantly reduce the size and complexity of the tree (the total number of branches and tiers of the structure) and improve the quality of its subsequent analysis.

Fake Review Detection Based on Multiple Feature Fusion and Rolling Collaborative Training

Fake reviews may mislead consumers. A large number of fake reviews will even cause huge property losses and public opinion crises. Therefore, it is necessary to detect and filter fake reviews. However, most existing methods have lower accuracy in detecting fake reviews due to they just use single features and lack of labeled experimental data. To solve this problem, we propose a novelty method to detect fake reviews based on multiple feature fusion and rolling collaborative training. First, the method requires an initial index system with multiple features such as text features, sentiment features of reviews and behavior features of reviewers. Second, the method needs an initial training sample set. Thus, we designed related algorithms to extract all the features of a review. Then the classification of the review is labeled manually. Finally, the method uses the initial sample set to train 7 classifiers, and the most accurate classifier will be selected to classify new reviews. The novelty of the method lies in that the features and the classification labels of the new reviews will be added into the initial sample set as new samples. So the size of the sample set will increase automatically. The experimental results in the reviews of yelp shopping website show that the accuracy of the proposed method for detecting fake reviews is 84.45%, which is 3.5% higher than the baseline methods. And compared with the latest deep learning model, its baseline precision has increased by 5.3%. According to the Friedman test, the support vector machine (SVM) classifier and random forest (RF) classifier has been proven to be the best one by statistical means. It means our method which uses multiple features has higher accuracy than the baseline models. Meanwhile, it also resolves the problem of lacking labeled training samples in fake reviews detection.

ОСОБЛИВОСТІ ПРОГРАМНИХ РІШЕНЬМОДЕЛЕЙ ЛОГІЧНИХ ДЕРЕВ КЛАСИФІКАЦІЇ НА ОСНОВІ СЕЛЕКЦІЇ НАБОРІВ ЕЛЕМЕНТАРНИХ ОЗНАК

Urgency of the research.Currently there are several independent approaches (concepts) to solve the classification problem in the general setting, and the development of various concepts, approaches, methods, and models that cover the general issues of the theory of artificial intelligence and information systems, all of these approaches in a recognition theory have their advantages and disadvantages and form a single tool to solve applied problems of the theory of artificial intelligence. This study will focus on the current concept of decision trees (classification trees). The general problem of software (algorithmic) construction of logical recognition trees (classification) is considered. The object of this research is logical classification trees (LСT structures). The subject of the research is actual methods and algorithmic schemes for constructing logical classification trees. Target setting.The main existing methods and algorithms for working with arrays of discrete information in the construc-tion of recognition functions (classifiers) do not allow you to achieve a predetermined level of accuracy (efficiency) of the classification system and regulate their complexity in the construction process. However, this disadvantage is absent in meth-ods and schemes for building recognition systems based on the concept of logical classification trees (decision trees). That is, the coverage of the training sample the set of elementary signs in the case of LCT generates a fixed tree data structure (model LCT), which provides compression and conversion initial data TS, and therefore allows significant optimization and savings of hardware resources of the system, and is based on a single methodology – the optimal approximation test sample set of elementary features (attributes) that are included in some schema (operator) constructed in the learning process.Actual scientific researches and issues analysis. The possibility of an effective and economical software (algorithmic) scheme for constructing a logical classification tree (LCT structuremodel) based on the source arrays of training samples (arrays of discrete information) of a large sample.The research objective. Development of a simple and high-quality software method (algorithm and software system) for building models (structures) LCTfor large arrays of initial samples by synthesizing minimal forms of classification and recog-nition trees that provide an effective approximation of educational information with a set of ranked elementary features (at-tributes) is created on the basis of ascheme for branched feature selection in a wide range of applied problems.The statement of basic materials. We propose a general program scheme for constructing structures of logical classifi-cation trees, which for a given initial training sample builds a tree structure (classification model), which consists of a set of elementary features evaluated at each step of building the model for this sample. A method and ready-made software system build logic trees the main idea is to approximate the initial random sampling of the volume set of elementary features. This method provides the selection of the most informative (qualitative) elementary features from the source set when forming the current vertex of the logical tree (node). This approach allows to significantly reduce the size and complexity of the tree (the total number of branches and tiers of the structure) and improve the quality of its subsequent analysis.Conclusions. The developed and proposed mathematical support for constructing LCT structures (classification tree mod-els) allows it to be used for solving a wide range of practical problems of recognition and classification, and the prospectsfor further research may consist in creating a limited method of logical classification tree (LCT structures), which consists in maintaining the criterion for stopping the procedure for constructing a logical tree by the depth of the structure, optimizing its software implementations, as well as experimental studies of this method for a wider range of practicalproblems.

Antenna Optimization Based on Co-Training Algorithm of Gaussian Process and Support Vector Machine

For the optimal design of electromagnetic components, surrogate model methods can usually be used, but obtaining labeled training samples from full-wave electromagnetic simulation software is most time-consuming. How to use relatively few labeled samples to obtain a relatively high-precision surrogate model is the current electromagnetic research hotspot. This article proposes a semi-supervised co-training algorithm based on Gaussian process (GP) and support vector machine (SVM). By using a small number of initial training samples, the initial GP model and initial SVM model can be trained by some basic parameter settings. Moreover, the accuracy of these two models can be improved by using the differences between these two models and combining with unlabeled samples for jointly training. In the co-training process, to ensure the performance of the proposed algorithm, a stop criterion set in advance to control the number of unlabeled samples introduced. Therefore, the accuracy of the model can be prevented from being reduced by introducing too much unlabeled samples, which can find the best solution in the limited time. The proposed co-training algorithm is evaluated by benchmark functions, optimal design of Yagi microstrip antenna (MSA) and GPS Beidou dual-mode MSA. The results show that the proposed algorithm fits the benchmark functions well. For the problem of resonant frequency modeling of the above two different MSAs, under the condition of using the same labeled samples, the predictive ability of the proposed algorithm is improved compared with the traditional supervised learning method. Moreover, for the groups of antenna sizes that meet the design requirements, the fitting effects of their return loss curve (S11) are well. The effectiveness of the proposed co-training algorithm has been well verified, which can be used to replace the time-consuming electromagnetic simulation software for prediction.

On semi-supervised learning

Major efforts have been made, mostly in the machine learning literature, to construct good predictors combining unlabelled and labelled data. These methods are known as semi-supervised. They deal with the problem of how to take advantage, if possible, of a huge amount of unlabelled data to perform classification in situations where there are few labelled data. This is not always feasible: it depends on the possibility to infer the labels from the unlabelled data distribution. Nevertheless, several algorithms have been proposed recently. In this work, we present a new method that, under almost necessary conditions, attains asymptotically the performance of the best theoretical rule when the size of the unlabelled sample goes to infinity, even if the size of the labelled sample remains fixed. Its performance and computational time are assessed through simulations and in the well- known “Isolet” real data of phonemes, where a strong dependence on the choice of the initial training sample is shown. The main focus of this work is to elucidate when and why semi-supervised learning works in the asymptotic regime described above. The set of necessary assumptions, although reasonable, show that semi-parametric methods only attain consistency for very well-conditioned problems.

Hyperspectral Image Classification With Small Training Sample Size Using Superpixel-Guided Training Sample Enlargement

Hyperspectral image (HSI) classification (HIC) has attracted much attention in the last decade. Spectral–spatial HIC methods have been the state-of-the-art methods in recent years. Small labeled training sample size (SLTSS) problem is still an important issue in HIC. This paper presents a spectral–spatial HIC method that is based on superpixel (SP) segmentation and distance-weighted linear regression classifier to tackle the SLTSS problem. First, SP segmentation is applied to the original HSI. Then, those SPs that contain training samples belonging to only one class are first searched out, and all the pixels of each of these SPs are assigned to the class of the training samples it contains. Next, with the identified labels, all these classified pixels are added to the initial training sample set for training sample set enlargement. Later, using this enlarged training sample set, the distance-weighted linear regression classifier is applied to classify each mean vector of each SP. Finally, the last classification map is obtained by assigning each SP with the same label as its mean vector. Experimental results on three HSI data sets demonstrate that the proposed approach can solve SLTSS problem very well and outperforms several state-of-the-art algorithms in classification accuracy under different training samples sizes.

Training sample refining method using an adaptive neighbor to improve the classification performance of very high-spatial resolution remote sensing images

A training sample refining method is proposed to improve the classification performance of very high-spatial resolution (VHR) remote sensing images. The proposed approach involves three major steps. First, for a given image, an initial sample set with a limited number for each class is prepared manually. Second, neighboring pixels around each available labeled pixel are gradually distinguished by an adaptive extension algorithm. When an iterative extension around the available pixel is terminated, the neighboring pixels that are within the extended region are taken into account as candidate training samples. The candidate training sample is then used to refine the signature of each initial sample. Third, when the whole available labeled pixels are scanned and processed pixel-by-pixel in the above manner, the revised training sample set is trained specially for a supervised classifier for classification. Three VHR remote sensing images with limited initial samples are used for evaluating different classifiers and advanced methods based on spatial–spectral features to investigate the feasibility and performance of the proposed approach. Higher classification performance and accuracies are obtained by our proposed approach with respect to the classification maps based on the initial training sample set and an existing method that improves the initial training set by a regular window.