Conventional Multivariate Methods Research Articles

Deep convolution neural network with weighted loss to detect rice seeds vigor based on hyperspectral imaging under the sample-imbalanced condition

• This study proposes the weight loss for DCNN to achieve batch and non-destructive vigor detection of rice seeds under the sample-imbalanced condition. • The proposed DCNN with weighted loss outperforms the DCNN with balanced loss and the conventional multivariate methods. • Modern hyperspectral imaging technology was combined with traditional germination test and biochemistry analysis method. • Visualization analysis results interpret the aggregation characteristics between the feature points outputted by the proposed DCNN with weighted loss. Vigor detection of crop seeds before putting them on the market is important for ensuring the yield and quality of the crops. In the actual production, however, different levels tend to vary in the number of samples, leading to the problem of sample imbalance. This study proposed a deep convolution neural network (DCNN) with weighted loss to achieve batch and non-destructive vigor detection of rice seeds based on hyperspectral imaging (HSI) under the sample-imbalanced condition. The true vigor state of seeds with different degrees of artificial aging was labeled by traditional analysis methods. The seeds were first classified into six categories according to the aging time using a constructed DCNN, which was proved to be unreasonable. Then four categories were merged, and the seeds were reclassified into three new categories by a rebuilt DCNN. However, merging categories caused the problem of sample imbalance, leading to much confusion between two aged categories. Thus, a DCNN with weighted loss was further proposed focusing on assigning appropriate weight to each category. Obtaining the highest accuracy and Macro F1 of 97.69% and 97.42%, respectively, it outperformed the DCNN with balanced loss and conventional models. The visualization analysis was conducted using PCA and t-SNE to inspect the aggregation between feature points. The overall results indicated the effectiveness of the proposed DCNN with weighted loss in the vigor detection of rice seeds under the sample-imbalanced condition, which would be conducive to online grading according to seed vigor and other qualities in the actual production.

Fault Detection of Tennessee Eastman Process using Kernel Dissimilarity Scale Based Singular Spectrum Analysis

Singular spectrum analysis (SSA) has become a popular and widely used forecasting and pre-processing technique in time series analysis which is currently exploited in chemical process monitoring and fault detection. Given its increased application and superior performance in comparison to conventional multivariate methods such as Principal Component Analysis (PCA) and Wavelets and its nonlinear extensions, it is relevant to study the variants of SSA and its applications in process monitoring. In this study SSA is combined with Kernel Multidimensional Scaling called Kernel Dissimilarity Scale Based Singular Spectrum Analysis (KDSSA) and is used to detect the faults in the Tennessee Eastman Process (TEP). The methodology is focused on three particular faults which were not observable with conventional multivariate methods and its no nlinear extensions. The monitoring results showed that the proposed method is efficient in detecting those faults in reduced number of modes. A unified monitoring index combined T2 statistics with Q statistics is used to simplify the fault detection task.

Transformation of cover-abundance values for appropriate numerical treatment – Alternatives to the proposals by Podani

Abstract Two alternatives are offered to Podani's proposals, based on the claim that Braun-Blanquet cover-abundance estimates cannot be properly analysed by conventional mul-tivariate methods. 1. The ordinal transform scale, based on an extended Braun-Blanquet cover-abundance scale, comes close to a metric cover percentage scale after (1) the abundance values r (very few individuals), + (few ind.), 1 (abundant) and 2m (very abundant, cover < 5%) are replaced by cover percentage estimates and (2) the higher Braun-Blanquet values, notably 4 and 5, with cover intervals 50-75% and 75-100%, respectively, are interpreted as estimates of considerably higher cover values than the usual visual projection on the ground (because of the position of stems and leaves in several layers). I propose the equation ln C = (OTV -2)1 a, where C = Cover%, OTV is the 1 to 9 Ordinal Transfer Value and a is a factor weighting the cover values. With this equation cover values in a geometric series are achieved for the nine values i...

On the evaluation of ordinal data with conventional multivariate procedures

AbstractIn a recent Forum paper, it is argued that, in most studies, ordinal data such as the Braun‐Blanquet abundance/dominance scale are not properly treated by multivariate methods. This is because conventional multivariate methods are generally adequate for ratio‐scale variables only, while for ordinal variables differences between states and their ratios are not interpreted. Conversely, in this paper it is shown that using conventional multivariate procedures for evaluating ordinal data should imply a shift from a metric space to a topological data space; as such the use of ordinal data does not represent a serious methodological error, provided that results are interpreted accordingly.

On the evaluation of ordinal data with conventional multivariate procedures

In a recent Forum paper, it is argued that, in most studies, ordinal data such as the Braun-Blanquet abundance/dominance scale are not properly treated by multivariate methods. This is because conventional multivariate methods are generally adequate for ratio-scale variables only, while for ordinal variables differences between states and their ratios are not interpreted. Conversely, in this paper it is shown that using conventional multivariate procedures for evaluating ordinal data should imply a shift from a metric space to a topological data space; as such the use of ordinal data does not represent a serious methodological error, provided that results are interpreted accordingly.

IDENTIFYING BATS FROM TIME-EXPANDED RECORDINGS OF SEARCH CALLS: COMPARING CLASSIFICATION METHODS

Recording ultrasonic echolocation calls of bats using bat-detectors is often used for wide-scale monitoring in studies on bat management and conservation. In Europe, the most important legal instrument for bat conservation is the Habitat Directive (43/92/EEC), which defines various levels of species (and habitat) protection for different bat species and/or genera. Thus for most management needs, the usefulness of bat-monitoring techniques depends on the possibility to determine to species/genus. We compared the discrimination performances of 4 statistical methods applied to identify bat species from their ultrasonic echolocation calls. In 3 different areas of northern Italy, we made recordings of 20 species of bat (60% of those occurring in Italy), 17 from the family Vespertilionidae and 3 from Rhinolophidae. Calls of bats identified to species level from morphological and genetic characters were time-expanded and recorded on release. We measured 7 variables from each call, and we developed classification models through both conventional tests (multiple discriminant analysis and cluster analysis) that were based on a classical statistical approach, and through 2 nonconventional classifiers (classification and regression trees, and neural networks) that relied on generalization and fuzzy reasoning. We compared the performance of the 4 techniques using the percentage of cases classified correctly in 5 classification trials at various taxonomic levels that were characterized by an increasingly difficult identification task: (1) family level (Rhinolophidae vs. Vespertilionidae), (2) species level within genus Rhinolophus, (3) genus level within Vespertilionidae, (4) species level within genus Myotis, and (5) all species. Multiple discriminant function analysis (DFA) correctly classified marginally more cases than artificial neural networks (ANN; 74–100% against 64–100%), especially at the species level (trial 4, species of genus Myotis; trial 5, all species). Both these techniques performed better than cluster analysis or classification and regression trees, the latter reaching only 56 and 41% in Myotis species and all species trials. Artificial neural networks do not yet seem to offer a major advantage over conventional multivariate methods (e.g., DFA) for identifying bat species from their ultrasonic echolocation calls.

Two Projection Methods for Use in the Analysis of Multivariate Process Data With an Illustration in Petrochemical Production

Principal components analysis (PCA) is often used in the analysis of multivariate process data to identify important combinations of the original variables on which to focus for more detailed study. However, PCA and other related projection techniques from the standard multivariate repertoire are not explicitly designed to address or to exploit the strong autocorrelation and temporal cross-correlation structures that are often present in multivariate process data. Here we propose two alternative projection techniques that do focus on the temporal structure in such data and that therefore produce components that may have some analytical advantages over those resulting from more conventional multivariate methods. As in PCA, both of our suggested methods linearly transform the original p-variate time series into uncorrelated components; however, unlike PCA, they concentrate on deriving components with particular temporal correlation properties, rather than those with maximal variance. The first technique finds components that exhibit distinctly different autocorrelation structures via modification of a signal-noise decomposition method used in image analysis. The second method draws on ideas from common PCA to produce components that are not only uncorrelated as in PCA, but that also have approximately zero temporally lagged cross-correlations for all time lags. We present the technical details for these two methods, assess their performance through simulation studies, and illustrate their use on multivariate output measures from a fluidized catalytic cracking unit used in petrochemical production, contrasting the results obtained with those from standard PCA.

Processing of three-dimensional microscopic X-ray fluorescence data

A novel polycapillary based confocal X-ray fluorescence (XRF) technique was applied for the first time at the ID18F beamline of the ESRF to obtain directly three-dimensional (3D) compositional information of an inclusion inside a natural diamond sample (KK200). Preliminary analysis of the results on the basis of the trace elements Sr, Y/Zr, and Th suggest three phases, two of which were identified by earlier micro-Raman spectroscopy as larnite (β-Ca2SiO4) and CaSiO3-walstromite, two minerals with significantly different Ca content. In order to support this multiphase model for the investigated inclusion the data set was analysed combining the conventional multivariate method of PCA and K-means clustering procedure after application of instrument specific routines such as spectral evaluation and normalization. Through the knowledge of the full spatial 3D structure of the different phases, it was possible to correct for absorption of the fluorescent radiation in the different phases of the inclusion and the surrounding diamond.

Comparing discriminant analysis, neural networks and logistic regression for predicting species distributions: a case study with a Himalayan river bird

We assessed the occurrence of a common river bird, the Plumbeous Redstart Rhyacornis fuliginosus, along 180 independent streams in the Indian and Nepali Himalaya. We then compared the performance of multiple discrimant analysis (MDA), logistic regression (LR) and artificial neural networks (ANN) in predicting this species’ presence or absence from 32 variables describing stream altitude, slope, habitat structure, chemistry and invertebrate abundance. Using the entire data (=training set) and a threshold for accepting presence in ANN and LR set to P≥0.5, ANN correctly classified marginally more cases (88%) than either LR (83%) or MDA (84%). Model performance was assessed from two methods of data partitioning. In a ‘leave-one-out’ approach, LR correctly predicted more cases (82%) than MDA (73%) or ANN (69%). However, in a holdout procedure, all the methods performed similarly (73–75%). All methods predicted true absence (i.e. specificity in holdout: 81–85%) better than true presence (i.e. sensitivity: 57–60%). These effects reflect species’ prevalence (=frequency of occurrence), but are seldom considered in distribution modelling. Despite occurring at only 36% of the sites, Plumbeous Redstarts are one of the most common Himalayan river birds, and problems will be greater with less common species. Both LR and ANN require an arbitrary threshold probability (often P=0.5) at which to accept species presence from model prediction. Simulations involving varied prevalence revealed that LR was particularly sensitive to threshold effects. ROC plots (received operating characteristic) were therefore used to compare model performance on test data at a range of thresholds; LR always outperformed ANN. This case study supports the need to test species’ distribution models with independent data, and to use a range of criteria in assessing model performance. ANN do not yet have major advantages over conventional multivariate methods for assessing bird distributions. LR and MDA were both more efficient in the use of computer time than ANN, and also more straightforward in providing testable hypotheses about environmental effects on occurrence. However, LR was apparently subject to chance significant effects from explanatory variables, emphasising the well-known risks of models based purely on correlative data.

Optimization of Crushing Strength and Disintegration Time of a High-Dose Plant Extract Tablet by Neural Networks

Optimization of crushing strength and disintegration time of a high-dose plant extract tablet was reached after extensive experimentation. Effects of the processing parameters, like compression force and tooling, and also of the excipients were found to be significant. Best results for both disintegration time and crushing strength were obtained with a plant extract that was granulated by roller compaction before compression. To gain more information about the different effects, artificial neural networks (ANNs) and a conventional multivariate method (partial least squares [PLS]) were used for data analysis. The topologies of the neural networks of the feed-forward type were optimized manually and by pruning methods. All methods were tested for contemplated parameters, crushing strength, and disintegration time. In general, ANNs were found to be more successful in characterizing the effects that influence crushing strength and disintegration time than the conventional multivariate methods.

A geometric approach to the analysis of physiological flow data.

Physiological flow data are common in various medical fields. Examples include urinary, blood and expiratory flows. They are widely used in assessing functions in the urinary, circulatory, or pulmonary systems, respectively. Current statistical methods for analysing these flow data in clinical trials are either univariate analyses, which do not utilize all the information together, or some conventional multivariate methods (such as regression analyses) which yield results that do not render clear medical interpretations. This paper presents a new approach to analysing the flow data, using urinary flow as the primary focus. The basic idea and technical steps are applicable to other flow data as well. The proposed method aims to transform the flow measurements back to the shape of the flow graphs. Since the whole geometric pattern of the flow graph provides more information about the patient's flow condition than any individual flow parameter alone, the method is a meaningful way of combining and analysing the flow data in both statistical and clinical senses. The method is a three-stage procedure. Patients are classified into three classes in the first stage and then ranked in sequence in the second stage, according to the geometry of the shape pattern and some clinical criteria. The classification procedure is shown to be very reliable when compared with the clinician's visual evaluation, and hence can be implemented by computer programming to aid clinical trials involving many patients. The whole ranking score is then readily analysed at the third stage for comparing treatment effects by the analysis of covariance method based on ranks, with the post-treatment score as the response variable and the baseline score as the covariate. An example of a urinary flow data set is provided to illustrate the use of the procedure.