Lactose prediction in dry milk with hyperspectral imaging: A data analysis competition at the "International Workshop on Spectroscopy and Chemometrics

In April 2024, the Vistamilk SFI Research Centre organized the fourth edition of the "International Workshop on Spectroscopy and Chemometrics - Spectroscopy meets modern Statistics". Within this event, a data challenge was organized among workshop participants, focusing on hyperspectral imaging (HSI) of milk samples. Milk is a complex emulsion comprising of fats, water, proteins, and carbohydrates. Due to the widespread prevalence of lactose intolerance, precise lactose quantification in milk samples became necessary for the dairy industry. The dataset provided to the participants contained spectral data extracted from HSI, without the spatial information, obtained from 72 samples with reference laboratory values for lactose concentration [mg/mL]. The winning strategy was built using ROCKET, a convolutional-based method that was originally designed for time series classification, which achieved a Pearson correlation of 0.86 and RMSE of 9.8 on the test set. The present paper describes the approaches and statistical methods adopted by all the participants to analyse the data and develop the lactose prediction models.