Spectral preprocessing and instrumentation for the identification of camelid and goat fibers using artificial intelligence

Luxury fibers, derived from South American camelids (SAC)—alpaca and vicuña—and goats (cashmere and mohair), are highly valued in fashion and haute couture for their softness, lightness, thermal properties, and scarcity. They are obtained through a meticulous process and holdgreat cultural value. For this reason, they are closely linked to key factors such as price, quality ,and comfort. It is therefore essential to identify their origin and characteristics in order to carryout proper evaluation and quality control. In a previous study, the authors developed and validated two machine learning models: deep neural networks (DNN) and support vector machine (SVM), to quickly, easily, and automatically identify Fourier-Transform Infrared (FTIR) spectra of SAC and caprine fibers. However, in such effort, the traditional spectral band range with no pre-processing was used. In addition, only an FTIR spectrometer in Attenuated Total Reflectance (ATR) mode was used. Thus, in this paper, the effects of, on one hand, pre-process the spectra and, on the other hand, using a FTIR microspectrometer in ATR mode are analyzed. For the former, two separate actions were tested: removing the CO2 band and selecting a characteristic spectral band (usually referred to as “the fingerprint region”). Combining all these options (i.e. machine learning algorithm, spectral pre-processing and FTIR equipment), five additional different experiments were carried out, which were compared with the previously achieved results. Samples from SAC (alpaca: n ¼ 51, llama: n ¼ 50, vicuña: n ¼ 50) and goats (mohair: n ¼ 35, cashmere: n ¼ 20) were evaluated, yielding valid FTIR-ATR spectra (n ¼ 1236) and m-FTIR-ATR spectra (n ¼ 1220). A 100% accuracy was achieved when evaluating m-FTIR-ATR spectra of alpacas with the DNN model, and a 98.3% accuracy with the SVM model. After removing the CO2 band from the FTIR-ATR spectra, accuracy dropped to 95.12%, which is 1.63 percentage points lower than the 96.75% reported in a previous study by the authors, where the whole spectra was evaluated. On the other hand, evaluating the fingerprint region with the models resulted in an accuracy of up to 89.43%, indicating that important information is lost when the rest of the spectrum region is removed. Regardless of the spectrometry technique used, DNN models provided higher accuracy than SVM models, although performance improved with FTIR-ATR spectra. In conclusion, machine learning models can be used to identify and assess spectral differences between the fibers of diverse SAC and Caprine species.