Individual Classification Results Research Articles

Neural network for determining an asteroid mineral composition from reflectance spectra

Context.Chemical and mineral compositions of asteroids reflect the formation and history of our Solar System. This knowledge is also important for planetary defence and in-space resource utilisation. In the next years, space missions will generate extensive spectral datasets from asteroids or planets with spectra that will need to be processed in real time.Aims.We aim to develop a fast and robust neural-network-based method for deriving the mineral modal and chemical compositions of silicate materials from their visible and near-infrared spectra. The method should be able to process raw spectra without significant pre-processing.Methods.We designed a convolutional neural network with two hidden layers for the analysis of the spectra, and trained it using labelled reflectance spectra. For the training, we used a dataset that consisted of reflectance spectra of real silicate samples stored in the RELAB and C-Tape databases, namely olivine, orthopyroxene, clinopyroxene, their mixtures, and olivine-pyroxene-rich meteorites.Results.We used the model on two datasets. First, we evaluated the model reliability on a test dataset where we compared the model classification with known compositional reference values. The individual classification results are mostly within 10 percentage-point intervals around the correct values. Second, we classified the reflectance spectra of S-complex (Q-type and V-type, also including A-type) asteroids with known Bus–DeMeo taxonomy classes. The predicted mineral chemical composition of S-type and Q-type asteroids agree with the chemical composition of ordinary chondrites. The modal abundances of V-type and A-type asteroids show a dominant contribution of orthopyroxene and olivine, respectively. Additionally, our predictions of the mineral modal composition of S-type and Q-type asteroids show an apparent depletion of olivine related to the attenuation of its diagnostic absorptions with space weathering. This trend is consistent with previous results of the slower pyroxene response to space weathering relative to olivine.Conclusions.The neural network trained with real silicate samples and their mixtures is applicable for a quantitative mineral evaluation of spectra of asteroids that are rich in dry silicates. The modal abundances and mineral chemistry of common silicates (olivine and pyroxene) can be derived with an accuracy better than 10 percentage points. The classification is fast and has a relatively small computer-memory footprint. Therefore, our code is suitable for processing large spectral datasets in real time.

Prediction of Navigational Decisions in the Real-World: A Visual P300 Event-Related Potentials Brain-Computer Interface

ABSTRACT Despite the widespread availability of portable neuroimaging systems, current applications of brain-computer interfaces (BCI) have largely remained confined to laboratory and clinical settings. In order to develop BCI that will assist individuals across a wide range of everyday life behaviors, it is critical to test paradigms that have been successful in highly controlled setups within the frame of real-world environments. In the present study, a visual P300 paradigm was used to elicit neural responses reflective of decision making within the frame of real-world navigation. Participants (n = 8) were equipped with a tablet and a mobile EEG system while navigating through university corridors. Upon reaching an intersection, participants were presented with left and right arrows flashing on the tablet (and additional non-directional stimuli in the distractor condition). Neural responses elicited by the presentation of such stimuli (Event-Related Potentials) were recorded along with the participant’s navigational decisions. The participants completed two separate sessions to collect data for both the training and testing of an offline BCI system aimed at predicting their navigational decisions based on the extraction of event-related potentials features. Single-trial classification accuracy reached 59.6% (up to 72.3% when classifying group averages of six trials). Individual classification results were contrasted with single-trial analysis. Results are discussed in terms of their implications for the design of real-world BCI applications, and several recommendations to improve experimental protocols are proposed.

DTI based Alzheimer’s disease classification with rank modulated fusion of CNNs and random forest

Automated classification of Alzheimer’s disease (AD) plays a key role in the diagnosis of dementia. In this paper, we solve for the first time a direct four-class classification problem, namely, AD, Normal Control (CN), Early Mild Cognitive Impairment (EMCI) and Late Mild Cognitive Impairment (LMCI) by processing Diffusion Tensor Imaging (DTI) in 3D. DTI provides information on brain anatomy in form of Fractional Anisotropy (FA) and Mean Diffusivity (MD) along with Echo Planar Imaging (EPI) intensities. We separately train CNNs, more specifically, VoxCNNs on FA values, MD values, and EPI intensities on 3D DTI scan volumes. In addition, we feed average FA and MD values for each brain region, derived according to the Colin27 brain atlas, into a random forest classifier (RFC). These four (three separately trained VoxCNNs and one RFC) models are first applied in isolation for the above four-class classification problem. Individual classification results are then fused at the decision level using a modulated rank averaging strategy leading to a classification accuracy of 92.6%. Comprehensive experimentation on publicly available ADNI database clearly demonstrates the effectiveness of the proposed solution.

Classification of Apples Based on the Shelf Life Using ANN and Data Fusion

In this research, two groups of gala variety apples were stored in two conditions of storage for 10 weeks: the first group at 0 °C with 95% relative humidity and the second group at 20 °C with 40% relative humidity, and each week, 10 samples of both groups were selected for experiments. After weighing and measuring the dimensions of the apples, an artificial intelligence and data fusion model was used to classify the apples based on shelf life. The non-destructive acoustic test was performed with a pendulum impactor, which impact signals simultaneously detected by sound and vibration sensors. The acoustic and vibrational signals of the impact were converted from the time domain to the frequency domain using a fast Fourier transform. Then, the pattern recognition technique using the artificial neural network was used to classify the fruits based on the shelf life at 1, 2, and 3week intervals. The dominant acoustic and vibrational frequencies and masses of the samples as three features delivered to the artificial neural network and the shelf life of samples were estimated by individual, binary, and trinary set of features. The binary and trinary set of features are used as mid-level fusion by ANN. The individual features classification results were fused by the conventional and Yager’s modified Dempster-Shafer method as the high-level fusion. In the feature-level fusion, which was done using the artificial neural network, the classification accuracy of the first group increased by an average of 10.84% and the second group by 10.14%. In the decision fusion, the lowest and highest increases of accuracy by the Yager and Dempster-Shafer methods were 1.6 and 19.8% for the first group and 3.7 and 12.5% for the second group, respectively.

3D Iterative Spatiotemporal Filtering for Classification of Multitemporal Satellite Data Sets

The current practice in land cover/land use change analysis relies heavily on the individually classified maps of the multi-temporal data set. Due to varying acquisition conditions (e.g., illumination, sensors, seasonal differences), the classification maps yielded are often inconsistent through time for robust statistical analysis. 3D geometric features have been shown to be stable for assessing differences across the temporal data set. Therefore, in this article we investigate the use of a multi-temporal orthophoto and digital surface model derived from satellite data for spatiotemporal classification. Our approach consists of two major steps: generating per-class probability distribution maps using the random-forest classifier with limited training samples, and making spatiotemporal inferences using an iterative 3D spatiotemporal filter operating on per-class probability maps. Our experimental results demonstrate that the proposed methods can consistently improve the individual classification results by 2%–6% and thus can be an important postclassification refinement approach.

A Context-Supported Deep Learning Framework for Multimodal Brain Imaging Classification

Over the past decade, “content-based” multimedia systems have realized success. By comparison, brain imaging and classification systems demand more efforts for improvement with respect to accuracy, generalization, and interpretation. The relationship between electroencephalogram (EEG) signals and corresponding multimedia content needs to be further explored. In this paper, we integrate implicit and explicit learning modalities into a context-supported deep learning framework. We propose an improved solution for the task of brain imaging classification via EEG signals. In our proposed framework, we introduce a consistency test by exploiting the context of brain images and establishing a mapping between visual-level features and cognitive-level features inferred based on EEG signals. In this way, a multimodal approach can be developed to deliver an improved solution for brain imaging and its classification based on explicit learning modalities and research from the image processing community. In addition, a number of fusion techniques are investigated in this work to optimize individual classification results. Extensive experiments have been carried out, and their results demonstrate the effectiveness of our proposed framework. In comparison with the existing state-of-the-art approaches, our proposed framework achieves superior performance in terms of not only the standard visual object classification criteria, but also the exploitation of transfer learning. For the convenience of research dissemination, we make the source code publicly available for downloading at GitHub ( https://github.com/aneeg/dual-modal-learning ).

Multimodal Depression Detection: Fusion Analysis of Paralinguistic, Head Pose and Eye Gaze Behaviors

An estimated 350 million people worldwide are affected by depression. Using affective sensing technology, our long-term goal is to develop an objective multimodal system that augments clinical opinion during the diagnosis and monitoring of clinical depression. This paper steps towards developing a classification system-oriented approach, where feature selection, classification and fusion-based experiments are conducted to infer which types of behaviour (verbal and nonverbal) and behaviour combinations can best discriminate between depression and non-depression. Using statistical features extracted from speaking behaviour, eye activity, and head pose, we characterise the behaviour associated with major depression and examine the performance of the classification of individual modalities and when fused. Using a real-world, clinically validated dataset of 30 severely depressed patients and 30 healthy control subjects, a Support Vector Machine is used for classification with several feature selection techniques. Given the statistical nature of the extracted features, feature selection based on T-tests performed better than other methods. Individual modality classification results were considerably higher than chance level (83 percent for speech, 73 percent for eye, and 63 percent for head). Fusing all modalities shows a remarkable improvement compared to unimodal systems, which demonstrates the complementary nature of the modalities. Among the different fusion approaches used here, feature fusion performed best with up to 88 percent average accuracy. We believe that is due to the compatible nature of the extracted statistical features.

SOM-based partial labeling of imbalanced data stream

Data streams are found in many large-scale systems such as security, finance, and internet. In many of the data streams, the class distribution is imbalanced, and hence most of the traditional classification modeler fails to produce high accuracy for the samples from minority class. In addition, data streams are changing and the model should be updated to maintain the classification performance. However, obtaining the true class labels of the data samples is not an easy task since labeling process is extremely time-consuming and very often class labels are not available immediately after classification. The goal of this research is to reduce the labeling for an imbalanced data stream, and to produce high classification performance when compared to fully labeling setting. In an imbalanced data stream, the challenging part is to find and label minority class samples. In this paper, we propose RLS-SOM (Reduced labeled Samples-Self Organizing Map) framework for classification of the imbalanced data stream in a non-stationary environment. RLS-SOM locates the minority class samples in the feature space using SOM. It maintains an ensemble of the classifiers and builds a new model when the changes occur, using only partial labeled samples. In RLS-SOM, the classification results are obtained from the ensemble, as well as each individual model in the ensemble. An individual model classification results are selected over ensemble results, if its performance is higher than the ensemble's performance. This comparison is performed to improve the performance as there may be one model in the ensemble that produces higher performance than the ensemble. Our experimental results demonstrate that RLS-SOM obtains higher performance when it is compared with several partially labeling techniques over benchmark data sets. In addition, the experimental results with other state of the art fully labeling methods such as UCB, SERA, SEA, and Learn++.CDS shows RLS-SOM maintains equivalent classification performance by using 10–30% labeling, on average.

HUMAN ACTIVITY RECOGNITION BY FUSING MULTIPLE SENSOR NODES IN THE WEARABLE SENSOR SYSTEMS

This paper presents a novel method for recognizing human daily activity by fusion multiple sensor nodes in the wearable sensor systems. The procedure of this method is as follows: firstly, features are extracted from each sensor node and subsequently reduced in dimension by generalized discriminant analysis (GDA), to ensure the real-time performance of activity recognition; then, the reduced features are classified with the multiclass relevance vector machines (RVM); finally, the individual classification results are fused at the decision level, in consideration that the different sensor nodes can provide heterogeneous and complementary information about human activity. Extensive experiments have been carried out on Wearable Action Recognition Database (WARD). Experimental results show that if all the five sensor nodes are fused with the adaptive weighted logarithmic opinion pools (WLOGP) fusion rule, we can even achieve a recognition rate as high as 98.78%, which is far more better than the situations where only single sensor node is available or the activity data is processed by state-of-the-art methods. Moreover, this proposed method is flexible to extension, and can provide a guideline for the construction of the minimum desirable system.

Vegetation classification on Prathong Island, Phang Nga, Thailand

Habitat mapping as defined by plant communities is a common component of the planning and monitoring for conservation management. This study provided the test of two classifiers for producing the vegetation mapping in a tropical island. The selected area is Prathong Island, Phang Nga Province, Southern part of Thailand. The performance of the support vector machines (SVMs) image classification technique for vegetation classifying was assessed and compared with maximum likelihood classifier (MLC). The vegetation was grouped into 6 categories as Grassland, Beach forest, Mangrove forest, Swamp forest, Sand and Water. The overall and individual classification results were compared to find out the suitable classifier for producing a vegetation map. THEOS multispectral image with 15-m resolution, achieved on 19 th January 2009, was used for this study. The results are useful to identify the boundary of each ecosystem on Prathong Island. Additional research is needed to assess the full potential of both classifiers and THEOS imagery for exploring potential applications on other tropical environments.

A comparison of classification techniques to support land cover and land use analysis in tropical coastal zones

This study assesses the performance of the support vector machine image classification technique in the context of a tropical coastal zone exhibiting low to medium scale development. The overall and individual classification results of this approach were compared to the maximum likelihood classifier and the artificial neural network techniques. A 15-m spatial resolution ASTER image of Koh Tao in Thailand was used for the test, and support vector machine was found to offer only limited improvements in classification accuracy over the other methodologies. The support vector machine did, however, show promise in dealing with the difficult challenge of separating human infrastructure such as buildings from other land cover types such as coastal rock and sandy beach which have very similar spectral signatures. The medium resolution ASTER image also proved highly suited to classifying coastal landscapes with this mix of land cover types. Additional research is needed to assess the full potential of the support vector machine in a weighted or layered classification, and to explore potential applications of this classification tool in other tropical environments.

Decision-Based Fusion for Improved Fluvial Landscape Classification Using Digital Aerial Photographs and Forward Looking Infrared Images

Riverine landscapes associated with large dynamic flood-plains provide a complex array of habitats for fish. Mapping and quantitative assessment of the habitats poses a major challenge. This study uses high spatial resolution airborne digital photographs and forward-looking infrared (FLIR) images, simultaneously acquired in spring 2005 over a 12 river kilometer section of the Unuk River in Southeast Alaska to map macro habitat indicators for Pacific salmon such as large woody debris (LWD), water channels, sand/gravel bars, and riparian vegetation. Image processing revealed that LWD could best be extracted using contextual information from the digital photos. River channel water had prominent shadows cast from neighboring trees and could not be accurately classified using digital photos, but could be well-delineated using unsupervised classification of the FLIR images. All other classes showed up well using supervised classification of digital photos. Using a decision-based fusion approach, the best individual classification results obtained from the digital photos and FLIR images to generate an improved fluvial landscape classification map (land-cover map). Using a decision-based fusion method resulted in an overall classification accuracy of the study area to 84.29 percent, compared to 77.00 percent using supervised classification of aerial photos alone. This appears to be first time that high-resolution airborne thermal images have been used for fluvial landscape classification, and the study clearly demonstrates the value of using thermal images and decision-based fusion approach for improved land-cover classification.

Integration of classifiers for improvement of vegetation category identification accuracy based on image objects

Though the multi‐classifier system (MCS) has been used widely in the area of machine learning, it has rarely been used in processing remotely sensed data. To test whether MCS could improve land cover classification results from remote sensing imagery, conservative voting, majority voting, and plurality voting were processed to combine individual classification results in this study. Considering the characteristics of high spatial resolution imagery, object‐oriented classification methods were employed. Firstly, the objects were acquired by segmenting the original images at two scales; then taking the “objects” as the classification units, five kinds of classifiers—iterative self‐organising data analysis (ISODATA), Mahalanobis distance, maximum likelihood, neural network, and expert system—were selected to classify urban vegetation categories using different samples and feature spaces to maintain independence among individual classifiers. The three voting rules were then used to combine the individual classification results. For unclassified objects, the ISODATA method was used again to identify their categories. The accuracy assessment showed that the three combined classification results gained higher accuracy than the individual classifier, and the plurality voting rules obtained the best identification results.

Reinforcing Web-object Categorization Through Interrelationships

Existing categorization algorithms deal with homogeneous Web objects, and consider interrelated objects as additional features when taking the interrelationships with other types of objects into account. However, focusing on any single aspect of the inter-object relationship is not sufficient to fully reveal the true categories of Web objects. In this paper, we propose a novel categorization algorithm, called the Iterative Reinforcement Categorization Algorithm (IRC), to exploit the full interrelationship between different types of Web objects on the Web, including Web pages and queries. IRC classifies the interrelated Web objects by iteratively reinforcing the individual classification results of different types of objects via their interrelationship. Experiments on a clickthrough-log dataset from the MSN search engine show that, in terms of the F1 measure, IRC achieves a 26.4% improvement over a pure content-based classification method. It also achieves a 21% improvement over a query-metadata-based method, as well as a 16.4% improvement on F1 measure over the well-known virtual document-based method. Our experiments show that IRC converges fast enough to be applicable to real world applications.

Integration of multiple neural classifiers for heart beat recognition

PurposeThis paper presents new approach to the integration of neural classifiers. Typically only the best trained network is chosen, while the rest is discarded. However, combining the trained networks helps to integrate the knowledge acquired by the component classifiers and in this way improves the accuracy of the final classification. The aim of the research is to develop and compare the methods of combining neural classifiers of the heart beat recognition.Design/methodology/approachTwo methods of integration of the results of individual classifiers are proposed. One is based on the statistical reliability of post‐processing performance on the trained data and the second uses the least mean square method in adjusting the weights of the weighted voting integrating network.FindingsThe experimental results of the recognition of six types of arrhythmias and normal sinus rhythm have shown that the performance of individual classifiers could be improved significantly by the integration proposed in this paper.Practical implicationsThe presented application should be regarded as the first step in the direction of automatic recognition of the heart rhythms on the basis of the registered ECG waveforms.Originality/valueThe results mean that instead of designing one high performance classifier one can build a number of classifiers, each of not superb performance. The appropriate combination of them may produce a performance of much higher quality.

Fusion of image classifications using Bayesian techniques with Markov random fields

This study investigates whether combining several different image classifications together with an a priori image model of the expected spatial distribution of the classes can produce a better classification. A maximum likelihood classifier and the cascade-correlation neural network architecture are used to generate various classification maps for satellite image data by varying the input features and network parameter settings. A likelihood for each pixel's class label is derived from the source classifications and combined with a Markov random field spatial image model to produce the final image classification. The method is applied to a ground cover type study based on Landsat Thematic Mapper (TM) imagery. It was found that a carefully selected combination could significantly improve individual classification results.

Off-line recognition of large-set handwritten characters with multiple hidden Markov models

There are many uncertainties in handwritten character recognition. Stochastic modeling is a flexible and general method for modeling such problems and entails the use of probabilistic models to deal with uncertain or incomplete information. This paper presents an efficient scheme for off-line recognition of large-set handwritten characters in the framework of stochastic models, the first-order hidden Markov models (HMMs). To facilitate the processing of unconnected patterns and patterns with isolated noises, four types of feature vectors based on the regional projection contour transformation (RPCT) are employed. The recognition system consists of two phases. For each character, in the training phase, multiple HMMs corresponding to different feature types of RPCT are built. In the classification phase, the results of individual classifiers to produce the final recognition result for an input character are integrated, where each individual HMM classifier produces one score that is the probability of generating the test observation sequence for each character model. In this paper, several methods for integrating the results of different classifiers are considered so that a better result could be obtained. In order to verify the effectiveness of the proposed scheme, the most frequently used 520 types of Hangul characters in Korea have been considered in the experiments. Experimental results indicate that the proposed scheme is very promising for the recognition of large-set handwritten characters with numerous variations.