Predicting cadmium accumulation in carrot (Daucus carota L.) using reflectance spectroscopy and machine learning

Abstract

Agricultural commodities such as root vegetables are subject to strict regulations regarding Trace Metal Elements (TME) for food safety reasons. Assessing the compliance of these commodities with authorized limits is usually achieved through costly and time-demanding traditional analytical techniques. As an alternative, we propose a new, rapid-and-scalable approach based on reflectance spectroscopy to assess TME content in root vegetables. The latter relies on exploiting the reflectance spectra of root samples to predict either the absolute concentration of TMEs or their compliance with a certain threshold using machine learning regression and classification. Our approach was successfully applied to predicting cadmium accumulation in the roots of two carrot varieties under controlled conditions, achieving up to 95% accuracy by exploiting the reflectance of root cross-sections (R² = 0.95 and F1-score ≥0.96 in regression and classification, respectively). We also explored non-destructive models using either carrot leaves or unpeeled roots, which achieved moderate to high accuracy for the prediction of root cadmium, respectively (0.48 < R² < 0.87 and 64 < F1-score <90). Our models showed consistent accuracy against varying cadmium limits and allowed identifying wavelengths in the visible and short-wave infrared regions of the spectrum as main contributors of predictions. Our study thus opens encouraging perspectives to assess TME compliance in agricultural commodities, from the field to harvest.