Supervised Classification Tools Research Articles

Quantifying the causes and consequences of variation in satellite‐derived population indices: a case study of emperor penguins

AbstractVery high‐resolution satellite (VHR) imagery is a promising tool for estimating the abundance of wildlife populations, especially in remote regions where traditional surveys are limited by logistical challenges. Emperor penguins Aptenodytes forsteri were the first species to have a circumpolar population estimate derived via VHR imagery. Here we address an untested assumption from Fretwell et al. (2012) that a single image of an emperor penguin colony is a reasonable representation of the colony for the year the image was taken. We evaluated satellite‐related and environmental variables that might influence the calculated area of penguin pixels to reduce uncertainties in satellite‐based estimates of emperor penguin populations in the future. We focused our analysis on multiple VHR images from three representative colonies: Atka Bay, Stancomb‐Wills (Weddell Sea sector) and Coulman Island (Ross Sea sector) between September and December during 2011. We replicated methods in Fretwell et al. (2012), which included using supervised classification tools in ArcGIS 10.7 software to calculate area occupied by penguins (hereafter referred to as ‘population indices’) in each image. We found that population indices varied from 2 to nearly 6‐fold, suggesting that penguin pixel areas calculated from a single image may not provide a complete understanding of colony size for that year. Thus, we further highlight the important roles of: (i) sun azimuth and elevation through image resolution and (ii) penguin patchiness (aggregated vs. distributed) on the calculated areas. We found an effect of wind and temperature on penguin patchiness. Despite intra‐seasonal variability in population indices, simulations indicate that reliable, robust population trends are possible by including satellite‐related and environmental covariates and aggregating indices across time and space. Our work provides additional parameters that should be included in future models of population size for emperor penguins.

Neural net pattern recognition based auscultation of croup cough and pertussis using phase portrait features

• We propose a novel approach to distinguish croup cough (CC) and pertussis (PT). • The signature features of CC and PT are investigated using mathematical tools. • The complex airflow dynamics is unwrapped through nonlinear time series features. • FFT and fractal analyses helps to understand the internal morphology of air ways. • Cubic KNN and neural net pattern recognition are used for prediction. Cough signal analysis for understanding the pathological condition has become important from the outset of the exigency posed by the epidemic COVID-19. The present work suggests a surrogate approach for the classification of cough signals - croup cough (CC) and pertussis (PT) – based on spectral, fractal, and nonlinear time-series techniques. The spectral analysis of CC reveals the presence of more frequency components in the short duration cough sound compared to PT. The musical nature of CC is unveiled not only through the spectral analysis but also through the phase portrait features – sample entropy (S), maximal Lyapunov exponent (L), and Hurst exponent (H b ). The modifications in the internal morphology of the respiratory tract, giving rise to more frequency components associated with the complex airflow dynamics, get staged through the higher fractal dimension of CC. Among the two supervised classification tools, cubic KNN (CKNN) and neural net pattern recognition (NNPR), used for classifying the CC and PT signals based on nonlinear time series parameters, NNPR is found better. Thus, the study opens the possibility of identification of pulmonary pathological conditions through cough sound signal analysis.

Improved Discrimination of Disease States Using Proteomics Data with the Updated Aristotle Classifier.

Mass spectrometry data sets from omics studies are an optimal information source for discriminating patients with disease and identifying biomarkers. Thousands of proteins or endogenous metabolites can be queried in each analysis, spanning several orders of magnitude in abundance. Machine learning tools that effectively leverage these data to accurately identify disease states are in high demand. While mass spectrometry data sets are rich with potentially useful information, using the data effectively can be challenging because of missing entries in the data sets and because the number of samples is typically much smaller than the number of features, two challenges that make machine learning difficult. To address this problem, we have modified a new supervised classification tool, the Aristotle Classifier, so that omics data sets can be better leveraged for identifying disease states. The optimized classifier, AC.2021, is benchmarked on multiple data sets against its predecessor and two leading supervised classification tools, Support Vector Machine (SVM) and XGBoost. The new classifier, AC.2021, outperformed existing tools on multiple tests using proteomics data. The underlying code for the classifier, provided herein, would be useful for researchers who desire improved classification accuracy when using their omics data sets to identify disease states.

Using unmanned aerial vehicle for identifying the vegetative vigor and quantify the area occupied by eucalyptus sprouts after chemical weeding in the state of Bahia, Brazil

Currently, the efficiency of chemical weeding for controlling eucalyptus sprouts is measured by field sampling, but the inefficiency of the sampling methods has led to the investigation of new technologies, such as using unmanned aerial vehicle (UAV) to help to identify the vegetative vigor of eucalyptus after chemical weeding. This study, therefore, used aerial images obtained by a UAV embedded with a sensor to identify the vegetative vigor and quantify the area occupied by eucalyptus sprouts 90 days after the chemical weeding. The study was conducted in three fields planted with eucalyptus whose sprouts had been previously controlled by the chemical weeding with the Scout® herbicide in November 2016. The vegetative vigor of the eucalyptus sprouts was evaluated from the aerial images obtained by the UAV with embedded sensor, during flights conducted in November 2016 and February 2017, that were used to calculate the normalized difference vegetation index and later, a random sample grid was constructed for each image by supervised classification of the area (m2) to determine the percentage occupied by the sprouts. The used chemical control method neither eradicated the sprouts nor reduced the sprout occupied area. The normalized difference vegetation index and supervised classification tools allowed determining with high precision sprout health status and size, generating interpretable data on the different evaluated fields and periods. The processing of the images obtained by the UAV provided a viable alternative of management to evaluate sprout status in reforestation areas.

Near infrared spectroscopy for counterfeit detection using a large database of pharmaceutical tablets

Medicine counterfeiting is one of the current burdens of the pharmaceutical world. Reliable technologies have become available for the chemical analysis of suspect medicines. Near infrared spectroscopy (NIRS) allows for instance fast, specific and non-destructive authentication of pharmaceutical products. In this paper, a NIRS method is presented for the identification of 29 different pharmaceutical product families of tablets, one family containing one or more formulation (s), e.g. different dosages. This selection represents the whole tablet portfolio of our firm. The high number of product families constituted a challenge, given that the measurement of the samples, made on two similar instruments, generated a dataset of 7120 spectra. Several chemometric tools proved efficient for the identification of these medicines. The dataset was first investigated with a Principal Component Analysis (PCA) in order to provide an overview of the distribution of the samples. The K-Nearest Neighbors (KNN), the Support Vector Machines (SVM) and the Discriminant Analysis (DA) supervised classification tools were successfully applied and generated an outstanding classification rate of 100% of correct answer. The methods were then fully validated with an independent set of spectra. The DA was selected as the method for the routine analysis of suspect tablets with the Mahalanobis distance as acceptance criterion for identification. Counterfeits, generics and placebos samples, constituting a second validation set, were tested and rejected by the method. NIRS has thus been demonstrated as an efficient tool for the quick identification of a large dataset of pharmaceutical tablets and the detection of counterfeit medicines.

Assessing biases in computing size spectra of automatically classified zooplankton from imaging systems: A case study with the ZooScan integrated system

Body size constrains prey–predator interactions and physiology, therefore plankton size spectra have been appointed as synthetic descriptors of plankton community structure and functioning. Recently developed imaging systems and supervised classification tools provide size measurements of any object in situ or in net samples and automatically classify them into previously defined categories. But because the nature of objects detected by these imaging systems is diverse, from non-living detritus to organisms of different plankton taxa, and because the steps in the analysis could introduce specific biases, a careful analysis of such plankton size spectra is needed before going deeper into ecological considerations. Using a WP2 net time series, we propose a general framework to analyze and validate zooplankton size spectra collected with nets and analyzed with the ZooScan integrated system that includes supervised classification. Size spectra were controlled, at each step of the procedure, to assess the modification of their shape due to several possible biases: (i) the effect of objects touching each other during the image acquisition, (ii) the error of the automatic classification differing among size classes and (iii) the choice of model to estimate body biovolume.