Unsupervised analysis of NIRS spectra to assess complex plant traits: leaf senescence as a use case.

BACKGROUND: As a rapid and non-destructive method, Near Infrared Spectroscopy is classically proposed to assess plant traits in many scientific fields, to observe enlarged genotype panels and to document the temporal kinetic of some biological processes. Most often, supervised models are used. The signal is calibrated thanks to reference measurements, and dedicated models are generated to predict biological traits. An alternative unsupervised approach considers the whole spectra information in order to point out various matrix changes. Although more generic, and faster to implement, as it does not require a reference data set, this latter approach is rarely used to document biological processes, and does requires more information of the process. METHODS: In our work, an unsupervised model was used to document the flag leaf senescence of durum wheat (Triticum turgidum durum). Leaf spectra changes were observed using Moving Window Principal Component Analysis (MWPCA). The dates related to earlier and later spectra changes were compared to two key points on the senescence time course: senescence onset (T0) and the end of the leaf span (T1) derived from a supervised strategy. RESULTS: For almost all leaves and whatever the signal pre-treatments and window size considered, the MWPCA found significant spectral changes. The latter was highly correlated with T1 (0.59 ≤ r ≤ 0.86) whereas the correlations between the first significant spectrum changes and T0 were lower (0.09 ≤ r ≤ 0.56). These different relationships are discussed below since they define the potential as well as the limitations of MWPCA to model biological processes. CONCLUSION: Overall, our study demonstrates that the information contained in the spectra can be used when applying an unsupervised method, here the MWPCA, to characterize a complex biological phenomenon such leaf senescence. It also means that using whole spectra may be relevant in agriculture and plant biology.

Reduction of repeatability error for analysis of variance-Simultaneous Component Analysis (REP-ASCA): Application to NIR spectroscopy on coffee sample.

A method to reduce repeatability error in multivariate data for Analysis of variance-Simultaneous Component Analysis (REP-ASCA) has been developed. This method proposes to adapt the acquisition protocol by adding a set containing repeated measures for describing repeatability error. Then, an orthogonal projection is performed in the row-space to reduce the repeatability error of the original dataset. Finally, ASCA is performed on the orthogonalized dataset. This method was evaluated on NIR spectral data of coffee beans. This study shows that the repeatability error due to physical variations between measurements can alter results of the analysis of variance. These effects are predominant in factors analysis and can be seen on spectra as constant or non-constant baselines. By reducing repeatability error with REP-ASCA, baselines are removed and factor analysis provides more information about chemical content of the factors of interest.

Using spatial information for evaluating the quality of prediction maps from hyperspectral images: A geostatistical approach.

Applying a calibration model onto hyperspectral (HS) images is of great interest because it produces images of chemical or physical properties. HS imaging is widely used in this way in food processing industries for monitoring product quality and process control. In this context, one of the main difficulties in the application of regression models to HS images is to evaluate the error of the obtained predictions, since in a proximal imaging set up, the size of the pixels is usually much smaller than the area required to obtain a wet chemical reference. Moreover, the selection of regression model parameters, such as the number of latent variables (LV) in a partial least squares (PLS) model, can modify the appearance of the prediction maps. The objective of this work is to propose an approach based on geostatistical indices to use spatial information of prediction maps for supporting the evaluation of regression models applied to HS images. This work stablishes a theoretical connection between linear regression model performance estimates and the spatial decomposition of variance in prediction maps, when the ground truth to estimate is spatially structured. This approach was tested in a simulated dataset and two real case studies. Geostatistical indices of the prediction maps were compared to model performance metrics for PLS models with increasing number of LV. The theoretical framework was proven by the results on the simulated dataset. In particular, the evolution of the nugget effect, C0, corresponded with the evolution of the random error of the model. Conversely, the error term of the model related with the slope of the model corresponded with the evolution of the structured variance observed in the prediction maps. On the real case studies, geostatistical indexes, extracted from the prediction maps, allowed to quantitatively evaluate the spatial structure of the estimations and complement the Root Mean Standard Error of Cross Validation (RMSECV) for the choice of optimal number of LV to consider in the model. The main advantage of this approach is that it does not require ground truth values. It could be used as a source of information for supporting the choice of optimum calibration parameters, such as the number of latent variables, or the choice of pre-treatments, complementing the traditional visual inspection of prediction maps with quantitative and objective metrics.

Exploring the potential of PROCOSINE and close-range hyperspectral imaging to study the effects of fungal diseases on leaf physiology.

The detection of plant diseases, including fungi, is a major challenge for reducing yield gaps of crops across the world. We explored the potential of the PROCOSINE radiative transfer model to assess the effect of the fungus Pseudocercospora fijiensis on leaf tissues using laboratory-acquired submillimetre-scale hyperspectral images in the visible and near-infrared spectral range. The objectives were (i) to assess the dynamics of leaf biochemical and biophysical parameters estimated using PROCOSINE inversion as a function of the disease stages, and (ii) to discriminate the disease stages by using a Linear Discriminant Analysis model built from the inversion results. The inversion results show that most of the parameter dynamics are consistent with expectations: for example, the chlorophyll content progressively decreased as the disease spreads, and the brown pigments content increased. An overall accuracy of 78.7% was obtained for the discrimination of the six disease stages, with errors mainly occurring between asymptomatic samples and first visible disease stages. PROCOSINE inversion provides relevant ecophysiological information to better understand how P. fijiensis affects the leaf at each disease stage. More particularly, the results suggest that monitoring anthocyanins may be critical for the early detection of this disease.

Near infrared hyperspectral dataset of healthy and infected apple tree leaves images for the early detection of apple scab disease.

This dataset presents two series of hyperspectral images of healthy and infected apple tree leaves acquired daily, from two days after inoculation until an advanced stage of infection (11 days after inoculation). The hyperspectral images were calibrated by reflection correction and registered to match the geometry of one reference image. On the last experiment day, scab positions are provided.

An iterative approach for compound detection in an unknown pharmaceutical drug product: Application on Raman microscopy.

Raman chemical imaging provides both spectral and spatial information on a pharmaceutical drug product. Even if the main objective of chemical imaging is to obtain distribution maps of each formulation compound, identification of pure signals in a mixture dataset remains of huge interest. In this work, an iterative approach is proposed to identify the compounds in a pharmaceutical drug product, assuming that the chemical composition of the product is not known by the analyst and that a low dose compound can be present in the studied medicine. The proposed approach uses a spectral library, spectral distances and orthogonal projections to iteratively detect pure compounds of a tablet. Since the proposed method is not based on variance decomposition, it should be well adapted for a drug product which contains a low dose product, interpreted as a compound located in few pixels and with low spectral contributions. The method is tested on a tablet specifically manufactured for this study with one active pharmaceutical ingredient and five excipients. A spectral library, constituted of 24 pure pharmaceutical compounds, is used as a reference spectral database. Pure spectra of active and excipients, including a modification of the crystalline form and a low dose compound, are iteratively detected. Once the pure spectra are identified, multivariate curve resolution-alternating least squares process is performed on the data to provide distribution maps of each compound in the studied sample. Distributions of the two crystalline forms of active and the five excipients were in accordance with the theoretical formulation.

Application-Dedicated Selection of Filters (ADSF) using covariance maximization and orthogonal projection.

Visible and near-infrared (Vis-NIR) spectra are generated by the combination of numerous low resolution features. Spectral variables are thus highly correlated, which can cause problems for selecting the most appropriate ones for a given application. Some decomposition bases such as Fourier or wavelet generally help highlighting spectral features that are important, but are by nature constraint to have both positive and negative components. Thus, in addition to complicating the selected features interpretability, it impedes their use for application-dedicated sensors. In this paper we have proposed a new method for feature selection: Application-Dedicated Selection of Filters (ADSF). This method relaxes the shape constraint by enabling the selection of any type of user defined custom features. By considering only relevant features, based on the underlying nature of the data, high regularization of the final model can be obtained, even in the small sample size context often encountered in spectroscopic applications. For larger scale deployment of application-dedicated sensors, these predefined feature constraints can lead to application specific optical filters, e.g., lowpass, highpass, bandpass or bandstop filters with positive only coefficients. In a similar fashion to Partial Least Squares, ADSF successively selects features using covariance maximization and deflates their influences using orthogonal projection in order to optimally tune the selection to the data with limited redundancy. ADSF is well suited for spectroscopic data as it can deal with large numbers of highly correlated variables in supervised learning, even with many correlated responses.

Setting local rank constraints by orthogonal projections for image resolution analysis: application to the determination of a low dose pharmaceutical compound.

Raman chemical imaging provides chemical and spatial information about pharmaceutical drug product. By using resolution methods on acquired spectra, the objective is to calculate pure spectra and distribution maps of image compounds. With multivariate curve resolution-alternating least squares, constraints are used to improve the performance of the resolution and to decrease the ambiguity linked to the final solution. Non negativity and spatial local rank constraints have been identified as the most powerful constraints to be used. In this work, an alternative method to set local rank constraints is proposed. The method is based on orthogonal projections pretreatment. For each drug product compound, raw Raman spectra are orthogonally projected to a basis including all the variability from the formulation compounds other than the product of interest. Presence or absence of the compound of interest is obtained by observing the correlations between the orthogonal projected spectra and a pure spectrum orthogonally projected to the same basis. By selecting an appropriate threshold, maps of presence/absence of compounds can be set up for all the product compounds. This method appears as a powerful approach to identify a low dose compound within a pharmaceutical drug product. The maps of presence/absence of compounds can be used as local rank constraints in resolution methods, such as multivariate curve resolution-alternating least squares process in order to improve the resolution of the system. The method proposed is particularly suited for pharmaceutical systems, where the identity of all compounds in the formulations is known and, therefore, the space of interferences can be well defined.

Distribution of a low dose compound within pharmaceutical tablet by using multivariate curve resolution on Raman hyperspectral images.

In this work, Raman hyperspectral images and multivariate curve resolution-alternating least squares (MCR-ALS) are used to study the distribution of actives and excipients within a pharmaceutical drug product. This article is mainly focused on the distribution of a low dose constituent. Different approaches are compared, using initially filtered or non-filtered data, or using a column-wise augmented dataset before starting the MCR-ALS iterative process including appended information on the low dose component. In the studied formulation, magnesium stearate is used as a lubricant to improve powder flowability. With a theoretical concentration of 0.5% (w/w) in the drug product, the spectral variance contained in the data is weak. By using a principal component analysis (PCA) filtered dataset as a first step of the MCR-ALS approach, the lubricant information is lost in the non-explained variance and its associated distribution in the tablet cannot be highlighted. A sufficient number of components to generate the PCA noise-filtered matrix has to be used in order to keep the lubricant variability within the data set analyzed or, otherwise, work with the raw non-filtered data. Different models are built using an increasing number of components to perform the PCA reduction. It is shown that the magnesium stearate information can be extracted from a PCA model using a minimum of 20 components. In the last part, a column-wise augmented matrix, including a reference spectrum of the lubricant, is used before starting MCR-ALS process. PCA reduction is performed on the augmented matrix, so the magnesium stearate contribution is included within the MCR-ALS calculations. By using an appropriate PCA reduction, with a sufficient number of components, or by using an augmented dataset including appended information on the low dose component, the distribution of the two actives, the two main excipients and the low dose lubricant are correctly recovered.

Application of independent component analysis on Raman images of a pharmaceutical drug product: pure spectra determination and spatial distribution of constituents.

Independent component analysis (ICA) was used as a blind source separation method on a Raman image of a pharmaceutical tablet. Calculations were performed without a priori knowledge concerning the formulation. The aim was to extract the pure signals from the initial data set in order to examine the distribution of actives and major excipients within the tablet. As a method based on the decomposition of a matrix of mixtures of several components, the number of independent component to choose is a critical step of the analysis. The ICA_by_blocks method, based on the calculation of several models using an increasing number of independent components on initial matrix blocks, was used. The calculated ICA signals were compared with the pure spectra of the formulation compounds. High correlations between the two active principal ingredient spectra and their corresponding calculated signals were observed giving a good overview of the distributions of these compounds within the tablet. Information from the major excipients (lactose and avicel) was found in several independent components but the ICA approach provides high level of information concerning their distribution within the tablet. However, the results could vary considerably by changing the number of independent components or the preprocessing method. Indeed, it was shown that under-decomposition of the matrix could lead to better signal quality (compared to the pure spectra) but in that case the contributions due to minor components or effects were not correctly identified and extracted. On the contrary, over-decomposition of the original dataset could provide information about low concentration compounds at the expense of some loss of signal interpretability for the other compounds.