Random Forest Ensemble Classifier Research Articles

A Novel Feature-Selection Algorithm in IoT Networks for Intrusion Detection.

The Internet of Things (IoT) and network-enabled smart devices are crucial to the digitally interconnected society of the present day. However, the increased reliance on IoT devices increases their susceptibility to malicious activities within network traffic, posing significant challenges to cybersecurity. As a result, both system administrators and end users are negatively affected by these malevolent behaviours. Intrusion-detection systems (IDSs) are commonly deployed as a cyber attack defence mechanism to mitigate such risks. IDS plays a crucial role in identifying and preventing cyber hazards within IoT networks. However, the development of an efficient and rapid IDS system for the detection of cyber attacks remains a challenging area of research. Moreover, IDS datasets contain multiple features, so the implementation of feature selection (FS) is required to design an effective and timely IDS. The FS procedure seeks to eliminate irrelevant and redundant features from large IDS datasets, thereby improving the intrusion-detection system's overall performance. In this paper, we propose a hybrid wrapper-based feature-selection algorithm that is based on the concepts of the Cellular Automata (CA) engine and Tabu Search (TS)-based aspiration criteria. We used a Random Forest (RF) ensemble learning classifier to evaluate the fitness of the selected features. The proposed algorithm, CAT-S, was tested on the TON_IoT dataset. The simulation results demonstrate that the proposed algorithm, CAT-S, enhances classification accuracy while simultaneously reducing the number of features and the false positive rate.

A hybrid model for post-treatment mortality rate classification of patients with breast cancer

Terminal cancer is not curable and eventually results in death. Breast cancer (BC) is a prevalent malignancy affecting women. Although there are prognostic indicators, BC prognosis is still challenging because of the intricate connections between various survival factors and influencing factors. This study proposes an ensemble classifier for predicting BC survivability using a new BC post-treatment dataset for the number of survivals. However, the classes survival cases dataset for BC is skewed, which caused a sub-optimal classification performance. Hence, a hybrid sampling scheme of Synthetic Minority Over-Sampling TEchnique (SMOTE) and Wilson's Edited Nearest Neighbor (ENN) is employed to treat the class imbalance in the dataset. Random Forest (RF) ensemble classifier is for classifying the dataset. The proposed framework performs well in terms of accuracy, recall of the two classes, Receiver Operating Characteristics (ROC) and Kappa Statistic (KS) metric on the dataset. The results demonstrated that the RF, with 97.0% accuracy on the holdout sample, is the best predictor. This prediction accuracy is superior to any noted in the literature, compared with Logistic Regression (LR) and Bagging classifiers.

Delineation of agricultural fields in arid regions from Worldview-2 datasets based on image textural properties

Barren lands are being transformed into agricultural fields with the growing demand for agriculture-based products. Hence, monitoring these regions for better planning and management is crucial. Surveying with high-resolution RS (remote sensing) satellites like Worldview-2 provides a faster and cheaper solution than conventional surveys. In the study, the arid region comprising cropland and barrenlands are efficiently and autonomously delineated using its spectral and textural properties using state-of-the-art random forest (RF) ensemble classifiers. The textural information window size is optimized and at a GLCM (gray-level co-occurrence matrix) window size of 13, a stable trend in classification accuracy was observed. A further rise in window sizes did not improve the classification accuracy; beyond GLCM 19, a decline in accuracy was observed. Comparing GLCM-13 RF with the no-GLCM RF classifier, the GLCM-based classifiers performed better; thus, the textural information assisted in removing isolated crop-classified outputs that are falsely predicted pixel groups. Still, it also obscured information about barren lands present within croplands. Delineation accuracy was 93.8% for the no-GLCM RF classifier, whereas, for the GLCM-13 RF classifier, an accuracy of 97.3% was observed. Thus, overall, a 3.5% improvement in accuracy was observed while using the GLCM RF classifier with window size 13. The textural information with proper calibration over high-spatial resolution datasets improves crop delineation in the present study. Henceforth, a more accurate cropland identification will provide a better estimate of the actual cropland area in such an arid region, which will assist in formulating a better resource management policy.

Optimal Feature Selection through Search-Based Optimizer in Cross Project

Cross project defect prediction (CPDP) is a key method for estimating defect-prone modules of software products. CPDP is a tempting approach since it provides information about predicted defects for those projects in which data are insufficient. Recent studies specifically include instructions on how to pick training data from large datasets using feature selection (FS) process which contributes the most in the end results. The classifier helps classify the picked-up dataset in specified classes in order to predict the defective and non-defective classes. The aim of our research is to select the optimal set of features from multi-class data through a search-based optimizer for CPDP. We used the explanatory research type and quantitative approach for our experimentation. We have F1 measure as our dependent variable while as independent variables we have KNN filter, ANN filter, random forest ensemble (RFE) model, genetic algorithm (GA), and classifiers as manipulative independent variables. Our experiment follows 1 factor 1 treatment (1F1T) for RQ1 whereas for RQ2, RQ3, and RQ4, there are 1 factor 2 treatments (1F2T) design. We first carried out the explanatory data analysis (EDA) to know the nature of our dataset. Then we pre-processed our data by removing and solving the issues identified. During data preprocessing, we analyze that we have multi-class data; therefore, we first rank features and select multiple feature sets using the info gain algorithm to get maximum variation in features for multi-class dataset. To remove noise, we use ANN-filter and get significant results more than 40% to 60% compared to NN filter with base paper (all, ckloc, IG). Then we applied search-based optimizer i.e., random forest ensemble (RFE) to get the best features set for a software prediction model and we get 30% to 50% significant results compared with genetic instance selection (GIS). Then we used a classifier to predict defects for CPDP. We compare the results of the classifier with base paper classifier using F1-measure and we get almost 35% more than base paper. We validate the experiment using Wilcoxon and Cohen’s d test.

Audio-visual multi-modality driven hybrid feature learning model for crowd analysis and classification.

The high pace emergence in advanced software systems, low-cost hardware and decentralized cloud computing technologies have broadened the horizon for vision-based surveillance, monitoring and control. However, complex and inferior feature learning over visual artefacts or video streams, especially under extreme conditions confine majority of the at-hand vision-based crowd analysis and classification systems. Retrieving event-sensitive or crowd-type sensitive spatio-temporal features for the different crowd types under extreme conditions is a highly complex task. Consequently, it results in lower accuracy and hence low reliability that confines existing methods for real-time crowd analysis. Despite numerous efforts in vision-based approaches, the lack of acoustic cues often creates ambiguity in crowd classification. On the other hand, the strategic amalgamation of audio-visual features can enable accurate and reliable crowd analysis and classification. Considering it as motivation, in this research a novel audio-visual multi-modality driven hybrid feature learning model is developed for crowd analysis and classification. In this work, a hybrid feature extraction model was applied to extract deep spatio-temporal features by using Gray-Level Co-occurrence Metrics (GLCM) and AlexNet transferrable learning model. Once extracting the different GLCM features and AlexNet deep features, horizontal concatenation was done to fuse the different feature sets. Similarly, for acoustic feature extraction, the audio samples (from the input video) were processed for static (fixed size) sampling, pre-emphasis, block framing and Hann windowing, followed by acoustic feature extraction like GTCC, GTCC-Delta, GTCC-Delta-Delta, MFCC, Spectral Entropy, Spectral Flux, Spectral Slope and Harmonics to Noise Ratio (HNR). Finally, the extracted audio-visual features were fused to yield a composite multi-modal feature set, which is processed for classification using the random forest ensemble classifier. The multi-class classification yields a crowd-classification accurac12529y of (98.26%), precision (98.89%), sensitivity (94.82%), specificity (95.57%), and F-Measure of 98.84%. The robustness of the proposed multi-modality-based crowd analysis model confirms its suitability towards real-world crowd detection and classification tasks.

Detecting Malware Families and Subfamilies using Machine Learning Algorithms: An Empirical Study

Machine learning algorithms have proved their effectiveness in detecting malware. This paper conducts an em-pirical study to demonstrate the effectiveness of selected machine learning algorithms in detecting and classifying Android malware using permissions features. The used dataset consists of 9000 different malicious applications from the CIC-Maldroid2020, CIC-Maldroid2017 and CIC-InvesAndMal2019 datasets collected by the Canadian Institute for Cybersecurity. Meta-Multiclass and Random Forest ensemble classifiers are used based on different machine learning classifiers to overcome the imbalance in the data classes. Moreover, a genetic attribute selection technique and SMOTE are used to classify Ransomware sub-families to handle the small size of the dataset and underfitting problem. The results show that optimization and ensemble approaches are successful in treating dataset issues, with 95% accuracy in classifying big malware families and 80% in Ransomware subfamilies.

Electronic Recommendation Based on Customer Review Using Text Mining

In machine learning, high-dimensional data classification is one of the significant challenges. For dimensionality reduction, the traditional classifiers enhance the variety of classifiers. The conventional method has some restrictions; information loss is caused during dimensionality reduction, minimizing accuracy. The selection of the sample is vulnerable to redundant features and noise. The proposed method, Hybrid Dimensionality Reduction Forest (HDRF) with Random Forest (RF) ensemble classifier and kappa measure, were used to overcome those restrictions. Initially, the Kappa measure is used for pruning, and the higher degree is selected from the forest. For partitioning the features, a tree-based selection method is used. Principal Component Analysis (PCA) is used for feature extraction, noise reduction, and dimensionality reduction. The proposed method removes the weak classifiers and eliminates redundancy. Also, it reduces the unselected structures and the fundamental structures into a new system. The evaluation results on 25 high-dimensional data, the proposed method outperforms with Random Forest ensemble classifier methods and provides enhanced results obtained on 21 out of 25.

Hybrid bio-inspired model for fraud detection with correlation based feature selection

The credit card is one of the most used forms of online transactions accepted by almost all the service payment systems. The increasing popularity of credit card use is increasing the challenge and fraud cases. This paper proposes a hybrid model based on the feature selection method, firefly bio-inspired algorithm, and random forest ensemble classifier (namely HBRF) for fraud detection and overcomes the imbalanced dataset issue in the banking domain. The firefly bio-inspired algorithm and the correlation-based feature selection (CFS) method have been applied to optimize the predictive features. The proposed HBRF model has been implemented on the Brazilian credit card dataset, and its performance has been compared with machine learning techniques such as Naive Bayes, KNN, Logistic Regression, Support Vector Machine, and existing methods. The proposed HBRF model has achieved an accuracy of 96.23% with a low error rate of 3.77%. It has been identified that the HBRF model outperforms state-of-the-art classifiers and existing techniques.

Adaptive Continuous Authentication System for Smartphones using Hyper Negative Selection and Random Forest Algorithms

As smartphones have become a part of our daily lives, including payment and banking transactions; therefore, increasing current data and privacy protection models is essential. A continuous authentication model aims to track the smartphone user's interaction after the initial login. However, current continuous authentication models are limited due to dynamic changes in smartphone user behavior. This paper aims to enhance smartphone user privacy and security using continuous authentication based on touch dynamics by proposing a framework for smartphone devices based on user touch behavior to provide a more accurate and adaptive learning model. We adopt a hybrid model based on the Hyper Negative Selection Algorithm (HNSA) as an artificial immune system (AIS) and the random forest ensemble classifier to instantly classify a user behavior. With the new approach, a decision model could detect normal/abnormal user behavior and update a user profile continuously while using his/her smartphone. The proposed approach was compared with the v-detector and HNSA, where it shows a high average accuracy of 98.5%, a low false alarm rate, and an increased detection rate. The new model is significant as it could be integrated with a smartphone to increase user privacy instantly. It is concluded that the proposed approach is efficient and valuable for smartphone users to increase their privacy while dynamic user behaviors evolve to change.

Advantage of Combining OBIA and Classifier Ensemble Method for Very High-Resolution Satellite Imagery Classification

Accurate and timely collection of urban land use and land cover information is crucial for many aspects of urban development and environment protection. Very high-resolution (VHR) remote sensing images have made it possible to detect and distinguish detailed information on the ground. While abundant texture information and limited spectral channels of VHR images will lead to the increase of intraclass variance and the decrease of the interclass variance. Substantial studies on pixel-based classification algorithms revealed that there were some limitations on land cover information extraction with VHR remote sensing imagery when applying the conventional pixel-based classifiers. Aiming at evaluating the advantages of classifier ensemble strategies and object-based image analysis (OBIA) method for VHR satellite data classification under complex urban area, we present an approach-integrated multiscale segmentation OBIA and a mature classifier ensemble method named random forest. The framework was tested on Chinese GaoFen-1 (GF-1), and GF-2 VHR remotely sensed data over the central business district (CBD) of Zhengzhou metropolitan. Process flow of the proposed framework including data fusion, multiscale image segmentation, best optimal segmentation scale evaluation, multivariance texture feature extraction, random forest ensemble learning classifier construction, accuracy assessment, and time consumption. Advantages of the proposed framework were compared and discussed with several mature state-of-art machine learning algorithms such as the k -nearest neighbor (KNN), support vector machine (SVM), and decision tree classifier (DTC). Experimental results showed that the OA of the proposed method is up to 99.29% and 98.98% for the GF-1 dataset and GF-2 dataset, respectively. And the OA is increased by 26.89%, 11.79%, 11.89%, and 4.26% compared with the traditional machine learning algorithms such as the decision tree classifier (DTC), support vector machine (SVM), k -nearest neighbor (KNN), and random forest (RF) on the test of the GF-1 dataset; OA increased by 32.31%, 13.48%, 9.77%, and 7.72% for the GF-2 dataset. In terms of time consuming, by rough statistic, OBIA-RF spends 223.55 s, SVM spends 403.57 s, KNN spends 86.93 s, and DT spends 0.61 s on average of the GF-1 and GF-2 datasets. Taking the account classification accuracy and running time, the proposed method has good ability of generalization and robustness for complex urban surface classification with high-resolution remotely sensed data.

Detection of abnormality in wireless capsule endoscopy images using fractal features

One of the most recent non-invasive technologies to examine the gastrointestinal tract is wireless capsule endoscopy (WCE). As there are thousands of endoscopic images in an 8–15 h long video, an evaluator has to pay constant attention for a relatively long time (60–120 min). Therefore the possibility of the presence of pathological findings in a few images (displayed for evaluation for a few seconds only) brings a significant risk of missing the pathology with all negative consequences for the patient. Hence, manually reviewing a video to identify abnormal images is not only a tedious and time consuming task that overwhelms human attention but also is error prone. In this paper, a method is proposed for the automatic detection of abnormal WCE images. The differential box counting method is used for the extraction of fractal dimension (FD) of WCE images and the random forest based ensemble classifier is used for the identification of abnormal frames. The FD is a well-known technique for extraction of features related to texture, smoothness, and roughness. In this paper, FDs are extracted from pixel-blocks of WCE images and are fed to the classifier for identification of images with abnormalities. To determine a suitable pixel block size for FD feature extraction, various sizes of blocks are considered and are fed into six frequently used classifiers separately, and the block size of 7×7 giving the best performance is empirically determined. Further, the selection of the random forest ensemble classifier is also done using the same empirical study. Performance of the proposed method is evaluated on two datasets containing WCE frames. Results demonstrate that the proposed method outperforms some of the state-of-the-art methods with AUC of 85% and 99% on Dataset-I and Dataset-II respectively.

Cost-sensitive metaheuristic technique for credit card fraud detection

To prevent online financial losses, credit card frauds need to be addressed and it should be controlled. To minimize the credit card frauds, fraudulent credit card transactions should be identified and minimized. Keeping this fact in view, in this paper, a cost-sensitive metaheuristic technique (CSFPA: cost-sensitive learning flower pollination metaheuristic algorithm) is proposed to minimize the misclassification cost of credit card transactions from class imbalance data using the correlation-based feature section method (CFS), flower pollination algorithm (FPA), and cost-sensitive classifier. To identify the fraudulent transactions, the random forest (RF) ensemble classifier is used as a base learner in the cost-sensitive classifier for classification tasks and the proposed CSFPA technique has been tested on the Brazilian bank dataset. The performance of proposed CSFPA has been compared with cost insensitive techniques and cost-sensitive techniques such as NB (Naive Base), KNN (K-Nearest Neighbors), LB (Logit Boost), CSForest (Cost-sensitive decision forest), NBBFS (NB Best First Search) KNNGSS (KNN Greedy Stepwise Search), and CSFCS (CSForest Cuckoo Search). The experimental outcomes revealed that the proposed CSFPA technique has obtained encouraging outcomes for handling misclassification costs (Total cost and Avg. cost) and also has outperformed (FPR, Precision and Recall) compared to all the other approaches.

Wrist in Motion: A Seamless Context-Aware Continuous Authentication Framework Using Your Clickings and Typings

In this paper, we propose enhanced continuous authentication by supplementing keystroke and mouse dynamics with wrist motion behaviors. Our method bridges the security gap when neither the mouse nor the keyboard is being used, such as during transitions from mouse to keyboard and vice versa, or during intermittent pauses when wrist movement is captured. Context-aware keystroke latency feature cell generation improves performance and solves latency fluctuation - different diagraphs have different latencies in different words. Based on two Random Forest Ensemble Classifiers (RFECs) recognizing the mouse and keystroke actions with corresponding wrist motions and one Sequential Sampling Analysis (SSA) or SSA Dynamic Trust Model (SSA-DTM), the identity of the user can be continuously verified no matter the operation mode - mouse clicking or keyboard typing. Experimental results, based on 44 subjects, show that the proposed approach can reach an FRR of 0.92% for genuine users and an FAR of 0 for attackers. The approach is shown to be more superior in efficient and timely authentications by making an authentication decision within only 35 mixed actions - mouse clicks or keystrokes, compared with conventional methods solely based on the mouse geometry and locomotion features or keystroke latency features.

Structured approach to downsize data administration and improve fault recognition accuracy for large power systems

This study presents a structured pattern-recognition framework for fault pattern recognition in large power systems. The main novelty of the study is in the efficient use of techniques to reduce data acquisition and administration requirements and limiting the computational overhead without sacrificing fault recognition accuracy and reliability. Optimal fault recorder placement based on the criterion of ‘network observability’ has been suggested for real-time monitoring of voltage and current at strategic network locations. An efficient feature-screening method has been adopted to select only the recorders whose data are assessed to be critical for fault recognition. ‘Wavelet’-based multi-resolution decomposition facilitated the extraction of important fault features from recorded waveforms and feature-screening was leveraged, once more, to eliminate irrelevant features to ensure that the data presented to the fault classifier is precise and feature-enriched. Random forest ensemble classifier was used for final class prediction and location of faults through voting of several individual learners to achieve high prediction accuracy. The proposed approach proved to be computationally efficient and consistent when tested on IEEE 118-bus system.

Mapping and Assessing the Dynamics of Shifting Agricultural Landscapes Using Google Earth Engine Cloud Computing, a Case Study in Mozambique

Land cover maps obtained at high spatial and temporal resolutions are necessary to support monitoring and management applications in areas with many smallholder and low-input agricultural systems, as those characteristic in Mozambique. Various regional and global land cover products based on Earth Observation data have been developed and made publicly available but their application in regions characterized by a large variety of agro-systems with a dynamic nature is limited by several constraints. Challenges in the classification of spatially heterogeneous landscapes, as in Mozambique, include the definition of the adequate spatial resolution and data input combinations for accurately mapping land cover. Therefore, several combinations of variables were tested for their suitability as input for random forest ensemble classifier aimed at mapping the spatial dynamics of smallholder agricultural landscape in Vilankulo district in Mozambique. The variables comprised spectral bands from Landsat 7 ETM+ and Landsat 8 OLI/TIRS, vegetation indices and textural features and the classification was performed within the Google Earth Engine cloud computing for the years 2012, 2015, and 2018. The study of three different years aimed at evaluating the temporal dynamics of the landscape, typically characterized by high shifting nature. For the three years, the best performing variables included three selected spectral bands and textural features extracted using a window size of 25. The classification overall accuracy was 0.94 for the year 2012, 0.98 for 2015, and 0.89 for 2018, suggesting that the produced maps are reliable. In addition, the areal statistics of the class classified as agriculture were very similar to the ground truth data as reported by the Serviços Distritais de Actividades Económicas (SDAE), with an average percentage deviation below 10%. When comparing the three years studied, the natural vegetation classes are the predominant covers while the agriculture is the most important cause of land cover changes.

Improving Malware Detection Classification Accuracy with Feature Selection Methods and Ensemble-based Machine Learning Methods

Malware is evolving serious threats to internet security. The classification of malware is extremely crucial in recent days. The traditional models are failed to achieve to get effective accuracy rate and the machine learning models are the basic models that accomplish the task of classification in a certain way, but in recent decades malware attacks are very drastic and difficult to achieve zero-day attacks. To compete with new malware, ensemble methods are highly effective and give better results of accuracy. In this paper, we propose a framework that combines the exploit of both feature selection methods and ensemble learning classifiers and gives better results of classification. In the experimental results, we prove that this combination of methods gives better classification with high accuracy of 100% with the Random Forest ensemble classifier.

Two-stage feature selection using ranking self-adaptive differential evolution algorithm for recognition of acceleration activity

Widespread research on activity recognition is becoming an imperative topic for improving the quality of human health. The fast development of sensing technology has become a fundamental platform for researchers to implement a system that could fulfill human needs. Due to privacy interests and low cost, wearable sensing technology is used in numerous physical activity monitoring and recognition systems. While these systems have proved to be successful, it is crucial to pay attention to the less relevant features to be classified. In such circumstances, it might happen that some features are less meaningful for describing the activity. Less complex and easy to understand, feature ranking is gaining a lot of attention in most feature dimension problems such as in bioinformatics and hyperspectral images. However, the improvement of ranking features in activity recognition has not yet been achieved. On the other hand, an evolutionary algorithm has proven its effectiveness in searching the best feature subsets. An exhaustive searching process of finding an optimal parameter value is another challenge. Consequently, this paper proposes a ranking self-adaptive differential evolution (rsaDE) feature selection algorithm. The proposed algorithm is capable of selecting the optimal feature subsets while improving the recognition of acceleration activity using a minimum number of features. The experiments employed real-world physical acceleration data sets: WISDM and PAMAP2. As a result, rsaDE performed better than the current methods in terms of model performance and its efficiency in the context of random forest ensemble classifiers.

WLAN Fingerprint Indoor Positioning Strategy Based on Implicit Crowdsourcing and Semi-Supervised Learning

Wireless local area network (WLAN) fingerprint positioning is an indoor localization technique with high accuracy and low hardware requirements. However, collecting received signal strength (RSS) samples for the fingerprint database is time-consuming and labor-intensive, hindering the use of this technique. The popular crowdsourcing sampling technique has been introduced to reduce the workload of sample collection, but has two challenges: one is the heterogeneity of devices, which can significantly affect the positioning accuracy; the other is the requirement of users’ intervention in traditional crowdsourcing, which reduces the practicality of the system. In response to these challenges, we have proposed a new WLAN indoor positioning strategy, which incorporates a new preprocessing method for RSS samples, the implicit crowdsourcing sampling technique, and a semi-supervised learning algorithm. First, implicit crowdsourcing does not require users’ intervention. The acquisition program silently collects unlabeled samples, the RSS samples, without information about the position. Secondly, to cope with the heterogeneity of devices, the preprocessing method maps all the RSS values of samples to a uniform range and discretizes them. Finally, by using a large number of unlabeled samples with some labeled samples, Co-Forest, the introduced semi-supervised learning algorithm, creates and repeatedly refines a random forest ensemble classifier that performs well for location estimation. The results of experiments conducted in a real indoor environment show that the proposed strategy reduces the demand for large quantities of labeled samples and achieves good positioning accuracy.

Using random forests for assistance in the curation of G-protein coupled receptor databases

BackgroundBiology is experiencing a gradual but fast transformation from a laboratory-centred science towards a data-centred one. As such, it requires robust data engineering and the use of quantitative data analysis methods as part of database curation. This paper focuses on G protein-coupled receptors, a large and heterogeneous super-family of cell membrane proteins of interest to biology in general. One of its families, Class C, is of particular interest to pharmacology and drug design. This family is quite heterogeneous on its own, and the discrimination of its several sub-families is a challenging problem. In the absence of known crystal structure, such discrimination must rely on their primary amino acid sequences.MethodsWe are interested not as much in achieving maximum sub-family discrimination accuracy using quantitative methods, but in exploring sequence misclassification behavior. Specifically, we are interested in isolating those sequences showing consistent misclassification, that is, sequences that are very often misclassified and almost always to the same wrong sub-family. Random forests are used for this analysis due to their ensemble nature, which makes them naturally suited to gauge the consistency of misclassification. This consistency is here defined through the voting scheme of their base tree classifiers.ResultsDetailed consistency results for the random forest ensemble classification were obtained for all receptors and for all data transformations of their unaligned primary sequences. Shortlists of the most consistently misclassified receptors for each subfamily and transformation, as well as an overall shortlist including those cases that were consistently misclassified across transformations, were obtained. The latter should be referred to experts for further investigation as a data curation task.ConclusionThe automatic discrimination of the Class C sub-families of G protein-coupled receptors from their unaligned primary sequences shows clear limits. This study has investigated in some detail the consistency of their misclassification using random forest ensemble classifiers. Different sub-families have been shown to display very different discrimination consistency behaviors. The individual identification of consistently misclassified sequences should provide a tool for quality control to GPCR database curators.

Object-based detection of Arctic sea ice and melt ponds using high spatial resolution aerial photographs

High resolution aerial photographs used to detect and classify sea ice features can provide accurate physical parameters to refine, validate, and improve climate models. However, manually delineating sea ice and melt ponds is time-consuming and labor-intensive. In this study, an object-based classification algorithm is developed to automatically extract sea ice and melt ponds efficiently from 163 aerial photographs taken during the Chinese National Arctic Research Expedition in summer 2010 (CHINARE 2010) in the Arctic Pacific Sector. The photographs are selected from 599 cloud-free photographs based on their image quality and representativeness in the marginal ice zone (MIZ). The algorithm includes three major steps: (1) the image segmentation groups the neighboring pixels into objects according to the similarity of spectral and textural information; (2) the random forest ensemble classifier distinguishes four general classes: water, general submerged ice (GSI, including melt ponds and submerged ice along ice edges), shadow, and ice/snow; and (3) the polygon neighbor analysis further separates melt ponds and submerged ice from the GSI according to their spatial relationships. The overall classification accuracy for the four general classes is 95.5% based on 178 ground reference objects. Furthermore, the producer's accuracy of 90.8% and user's accuracy of 91.8% are achieved for melt pond detection through 98 independent reference objects. For the 163 photos examined, a total of 19,438 melt ponds larger than 1m2 are detected, with a pond density of 867.2km−2, mean pond size of 32.6±0.03m2, and mean pond fraction of 0.06±0.006; a total of 42,468 ice floes are detected, with the mean floe size of 173.3±0.1m2 (majority in 1–30m2) and mean ice concentration of 46.1±0.5% (ranging from 18.6–98.6%). These results matched well with ship-based visual observations in the MIZ in the same area and time. The method presented in the paper can be applied to data sets of high spatial resolution Arctic sea ice photographs for deriving detailed sea ice concentration, floe size, and melt pond distributions over wider regions, and extracting sea ice physical parameters and their corresponding changes between years.