Predicting early mycotoxin contamination in stored wheat using machine learning

With continued global population growth and current rate of climate change, grain loss during storage remains a major contributor to postharvest losses of wheat (Triticum aestivum L.). Infection with mycotoxin leads to degradation or even discarding of stored grain, causing economic losses and risks to food security. Deep learning models have been used in the agricultural domain for detecting prevalent diseases or contamination; however, data scarcity remains a critical bottleneck for rapid implementation of computer vision in this field. Herein, a compact convolutional transformer (CCT)–based model was applied to classify contaminated wheat by deoxynivalenol (DON) and aflatoxins (AFB₁, AFB₂, AFG₁, and AFG₂), which was divided into three main classes: healthy, incipient, and contaminated. The classification was performed based on elevated CO₂ respiration rate (≥31.20 ± 0.62 mg CO₂ kg⁻¹ h⁻¹) and visual appearance of mold formation in initial and severe stage since the start the storage experiment. The proposed CCT model achieved an accuracy of 83.33%, with the contaminated class demonstrating the highest performance metrics, including precision (1.0), recall (0.90), and F1-score (0.95), followed by the healthy and incipient classes. At the same time, explicit classification between the healthy and incipient classes deserves further improvement because it is highly relevant for the timely detection of spoilage and prevention of proliferation of mycotoxins in stored wheat.