Automated Dimension Reduction in Archaeometry Using Autoencoder Neural Networks

By Samudrapom DamReviewed by Susha Cheriyedath, M.Sc.Feb 20 2024

In an article recently published in the journal Scientific Reports, researchers investigated the effectiveness of the autoencoder neural network as the dimension reduction technique based on the X-ray fluorescence (XRF) spectra reconstruction.

Dimensionality reduction in archaeometry

Energy dispersive X-ray fluorescence (EDXRF) spectrometry is extensively used as an analytical technique for analyzing materials that were used to make cultural heritage objects. The EDXRF analytical technique provides multivariate results, both when used with the full spectrum or with selected informative peaks associated with the elements in the material composition.

Thus, the analytical results are subjected to multivariate analysis for archaeometry studies to decrease the dimensionality of the initial data based on informative features. Several chemometric dimension reduction techniques and deep learning methods can be employed for dimensionality reduction of the analytical results.

However, developing models using the chemometric approach requires high expert involvement and is time-consuming. Applying neural networks (NNs) and deep learning techniques can successfully address the dimension reduction problem. For instance, autoencoders are crucial in learning efficient data representations. They compress the input data into a reduced-dimensional space and then reconstruct it to capture the input data's essential features.

Studies have also shown that autoencoder neural networks enable robustness and speed by effectively decreasing input data dimensions and obtaining only advanced features. Moreover, unlike the classical machine learning (ML) chemometric methods, deep neural networks can learn critical patterns from the raw spectra. This leads to fewer human interventions in feature selection and preprocessing, resulting in higher model robustness and accuracy.

The proposed approach

In this study, researchers designed an autoencoder neural network and used it as a dimension reduction tool of initial 40 × 2048 data obtained in the raw EDXRF spectra, containing information about the elemental composition of the selected points on the canvas paintings' surface. Overall, 40 yellow-colored spots of various nuances were analyzed using portable EDXRF spectrometry.

The study's objective was to investigate the effectiveness of the autoencoder neural network as the dimension reduction technique based on the XRF spectra reconstruction. Researchers performed dimension reduction to examine the input dataset structure to explore the possible variance in the paintings' creation period and identify a low-dimensional space that is suitable for undoubted attribution. All EDXRF spectra were pretreated using the peaks balancing procedure to minimize the contributions of the experimental setup.

The number of neurons per layer and number of layers were altered during the NN design, while simultaneously, various activation functions were employed to obtain the highest between-class separability within the reduced space. The natural logarithm, raw, and square root transformed spectral data were utilized to evaluate the NN efficiency and design to ensure a reliable procedure. Bhattacharyya's distance was employed to measure class separability.

The autoencoder's fundamental structure consisted of three primary components, including a hidden layer/bottleneck layer, an output layer, and an input layer. In the input layer, the initial input data, such as numerical values, images, or text, is received, while the hidden layer learns the encoded or compressed representation of the input data.

This hidden layer has fewer nodes/neurons compared to the output and input layers, which prompts the network to identify a compressed input representation. In the output layer, the input data is reconstructed from the encoded representation generated by the bottleneck layer. The objective is to create an output that closely resembles the input data.

Significance of the work

Researchers performed a non-destructive EDXRF analysis of painted layered materials. The elemental composition of the measured yellow-colored spots was utilized to attribute artistic paintings to late, middle, and early periods of creation. The use of multiple synthetic yellow pigments of various chemical compositions began during the creation of these works, which was the first half of the nineteenth century and beyond.

Pigment composition analysis displayed that at least five chemically different yellow pigments, including yellow ochre, chrome yellow, Naples yellow, lemon yellow, and the massicot pigment, were utilized in the analyzed spots. Among them, yellow ochre was used in all three periods, while chrome yellow and Naples yellow were used in the early and middle periods.

These findings indicated that deep learning methods could effectively reveal informative data regarding the chemical composition of the pigments used for the paintings' time attribution. The proposed autoencoder network design enabled the best possible original EDXRF spectrum reconstruction and the most informative feature extraction, which was used for dimension reduction.

Specifically, the central layer/bottleneck layer of the stacked autoencoder neural network (SAENN) had a dimension of two, which enabled instant reduction of the dimensions into a space that provides maximal visualization.

Additionally, the SAENN, with two neurons in the bottleneck layer and 70 neurons in the decoder/encoder layers, possessed the highest Bhattacharyya distance value. This is important as satisfactory separability was achieved using small numbers of neurons, enabling the training to be conducted reliably and quickly on the CPU without expert involvement.

To summarize, the findings of this study demonstrated the feasibility of the proposed autoencoder neural network-based efficient automated procedure for dimension reduction, feature extraction, and class separation based on the elemental composition of the pigments used and several painting creation periods.