Agricultural product drying is a critical process for ensuring food safety and enhancing added value. From grains to fruits and vegetables, fresh agricultural products are prone to spoilage due to high moisture content. In China, improper drying leads to annual grain losses of up to 21 million tons. Traditional drying techniques, reliant on manual experience or physical models, often face issues such as “over-drying causing quality degradation” and “insufficient drying leading to mildew”, coupled with persistently high energy consumption. With the development of artificial intelligence, can artificial neural networks (ANNs) bring intelligent transformation to this traditional industry?
Qing Wei and colleagues from the College of Engineering, China Agricultural University, systematically elaborated on the innovative applications of neural networks in agricultural product drying, offering new insights to address industry pain points. The related article has been published in Frontiers of Agricultural Science and Engineering (DOI: 10.15302/J-FASE-2025620).
Traditional drying models depend on mathematical formulas or physical equations, struggling to handle nonlinear changes during agricultural product drying. For instance, in hot air drying, parameters like temperature, air velocity, and humidity interact, and manual adjustments often result in uneven moisture removal. In contrast, neural networks simulate the connection patterns of human brain neurons, enabling them to learn patterns from vast experimental data and achieve precise predictions.
Studies have shown that neural networks excel in moisture ratio prediction. Taking peppermint drying as an example, traditional mathematical models yield large prediction errors, while neural network models achieve a coefficient of determination (R²) of 0.998, nearly identical to actual measurements. In drying shiitake mushrooms, hawthorns, and other agricultural products, neural networks can monitor moisture changes in real time, keeping prediction errors within a low range. This dynamic sensing capability allows drying equipment to automatically adjust parameters based on material status, preventing cracking or nutrient loss caused by over-drying.
Drying not only removes moisture but also preserves the quality of agricultural products. Traditional methods often cause fruit and vegetable browning and vitamin loss due to improper temperature control. By integrating multi-source data such as image recognition and sensor data, neural networks enable comprehensive quality control. For example, in kiwifruit drying, researchers used neural networks to develop a color prediction model, which optimizes temperature parameters by analyzing color changes during drying, thereby improving the retention rate of nutritional components.
High energy consumption is another major challenge in the agricultural product drying industry. Neural networks balance energy consumption and efficiency by optimizing drying processes. In garlic infrared drying experiments, researchers adjusted radiation intensity and air velocity using neural network models, reducing drying time by 6.5% and energy consumption by 36%.
Notably, the integration of neural networks with traditional control technologies is transforming production models. For instance, combining neural networks with PID controllers enables real-time regulation of temperature and humidity in drying chambers, addressing the issues of low precision and unstable product quality in traditional manual control. In tobacco curing, the deep learning-based TobaccoNet model automatically sets temperature and humidity parameters based on tobacco leaf images, achieving a prediction error of only 1.62% and significantly reducing manual intervention.
Despite these achievements, neural networks in agricultural product drying face challenges. Most current models rely on large volumes of labeled data, and the lack of experimental data for new agricultural products limits technology promotion. Additionally, the “black box” nature of deep neural networks raises doubts among some producers regarding reliability. The research team pointed that future efforts should focus on developing more concise and interpretable models, integrating Internet of Things (IoT) technology for real-time data collection and model updates.
DOI: 10.15302/J-FASE-2025620