User cases

Hyperspectral imaging has applications in many fields, including agriculture, environment, medicine and industry. This so-called “chemical” imaging makes it possible to characterize the chemical composition of products at each point of an image, thanks to the near-infrared spectrum measured for each pixel. It can thus be used to classify objects according to their composition, or to quantify compounds present on the surface and show their spatial distribution.

This scientific study presents the methodology for analyzing hyperspectral images in the laboratory, describing best practices in image acquisition, signal processing, image processing and finally chemometrics / Machine Learning methods applicable to hyperspectral images.

Images wee acquired with the SPECIM IQ (SPECIM – Konica Minolta) portable visible – near-infrared hyperspectral camera. The Machine LEarningmethods were developped on the PLS_Toolbox^® and MIA_Toolbox^® (Eigenvector Research Inc.).

In NIR spectroscopy, the most common calibration method is the PLS regression. It is a linear multivariate calibration method, efficient on spectral data and easy to implement. However, this method reaches its limits when the data to predict is complex, for instance when many products or recipes are analyzed or when non-linear correlations are present.

In this study, conducted by Bruker Optics and Ondalys, two spectroscopic data sets were acquired with Bruker FT-NIR instruments MPA or TANGO. In each dataset, the heterogeneity of the products and the non-linearity between the parameters of interest and the spectra justified the use of advanced ML methods.

The Support Vector Machines (SVM) for quantitative analysis, called Support Vector Regression (SVR) were applied on PLS scores.

This simple and fast approach gave very satisfactory results, with lower errors with the SVR models on the PLS scores than the PLS models built with the optimal pretreatment and variable range selection. Moreover, SVR models were also built on the PLS scores without variable selection in order to reduce the optimization time.

In this case, the performances of the PLS models were strongly degraded whereas those of the SVR models remained satisfactory.

The computing power of computers and volumes of data to be processed are increasing significantly. This makes Machine Learning (ML) more and more popular. Many methods exist and can be adapted to many fields.  But what about Machine Learning for the spectroscopic data analysis?

To answer this question, several Machine Learning methods were compared on a spectroscopic data set (Near infrared Spectroscopy). “Classical” methods, such as PLS – Partial Least Squares Regression –  et LWR – Locally Weighted Regression – are compared with 3 ML algorithms: SVM  – Support Vector Machines – , ANN – Artificial Neural Networks and  CART/RF – Classification and Regression Trees / Random Forest.

The B.I.P. – French interprofessional institute of the prune, wanted to develop a non-destructive decision aid-tool, applicable directly to the orchard in order to better estimate the maturity of the fruits and predict the optimal harvest date of the plums.

From large databases (more than 6000 samples) scanned on a laboratory spectrometer, the ASD LabSpec 4 (Malvern Panalytical), Ondalys has developed models for quantitative prediction of the sugar (° Brix) and acidity levels of plums in order to transfer the calibration on a micro-spectrometer MicroNIR^TM (Viavi Solutions)

Following encouraging results, the study continues to transfer the models obtained to the portable MicroNIR^TM spectrometer so that the fruit can be measured directly in the orchard.

Study funded by France AgriMer

Online process control has become essential in many industries. This monitoring improves the product quality and reduces costs, thanks to better production supervision and rapid intervention in case of drifts or anomalies.

Support Vector Machines (SVM) are parts of the supervised Machine Learning methods. SVM were originally developed for classification objectives (pattern recognition), especially to discriminate between convex or hardly-separable classes. But, they are also very effective for quantitative prediction purposes.

This technique is very interesting to model non-linear relationships between the data or for intricate situations (e.g. complex parameter or concentrations close to the detection threshold).

In order to implement SVMs, 3 parameters must be optimized: a regularization parameter, the margin size and the non-linearity degree of the model, which is much simpler than training Artificial Neural Networks.

When predicting several quantitative parameters based on near-infrared spectroscopic data, SVM models provided significantly better results than PLS regression one (errors divided by 2).

This Machine Learning method brought a significant performance improvement vs PLS regression, due to the non-linearity present in the data sets (fat, protein and humidity contents in meat with NIR spectroscopy).

This project also demonstrated that with a rather small training set size, SVMs could show high performance and generalization power on an independent test set.

In the frame of online monitoring of its polymerization processes, CERDATO, ARKEMA‘s research center, encountered problems in updating its spectroscopic calibrations. After several classical attempts, by adding new samples, their problem persisted.

They called Ondalys to train and support them in the use of different orthogonalization methods, in particular the Dynamic Orthogonal Projection (DOP) method.

Model corrected by DOP obtained better results than those developed with classical methods, while making it possible to diagnose problems that occurred on the production line.

In order to gain time and efficiency for the product characterization, one of our customers wished to develop a method for identifying his Raw Materials (powders and liquids) using a Near infrared spectrometer.

This major player in the Pharmaceutical Industry, asked Ondalys to apply robust classification methods and develop robust identification models, that can be used systematically on incoming batches of Raw Materials.

Over 100 different Raw Materials (liquids and powders) are now characterized while arriving, thanks to near infrared spectroscopy and identification models developed by Ondalys.