Soil-driven tea quality prediction via lightweight deep learning with rolled tensor representation

Traditionally, tea quality assessment has heavily depended on chemical testing methods, which are typically time-intensive, expensive, and necessitate specialized laboratory equipment. These drawbacks hinder large-scale, rapid, and cost-effective quality evaluations. To tackle these issues, this paper introduces the Rolled-Mobile-Tea Network (RMTN), a model designed for tea quality prediction based on soil data. The model incorporates an innovative rolling mechanism that converts one-dimensional tea soil parameter vectors into three-dimensional tensors, overcoming the constraints of traditional convolutional neural networks (CNNs) and facilitating the extraction of more intricate soil-quality relationships. Within the RMTN architecture, the Mobile-Tea (MT) module functions as a lightweight feature extractor, utilizing cost-effective computational operations and identity mappings to generate comprehensive feature representations, thereby simplifying the model while boosting its performance. The soil data analyzed encompasses nine key indicators: Ca, Cu, Fe, K, Mg, Mn, Na, P, and Zn. This study aims to predict six crucial tea quality components: gallic catechin (GC), epigallocatechin (EGC), catechin (C), theaflavin (TF), theaflavin gallate (TF-G), and theaflavin gallate 2 (TF-G2). Experimental results reveal model's strong predictive capabilities, as evidenced by the following performance metrics (RMSE, MAE, and R2): GC (0.07, 0.05, 0.73); EGC (0.09, 0.07, 0.90); C (0.25, 0.20, 0.76); TF (0.01, 0.01, 0.82); TF-G (0.02, 0.02, 0.78); and TF-G2 (0.02, 0.02, 0.83). On average, the model achieves an RMSE of 0.08, an MAE of 0.06, and an R2 of 0.80, underscoring its robustness across various quality indicators. This research presents a new technical solution for convenient, low-cost, and accurate tea quality detection, and also offers a methodological framework for quality assessment in other crops, thereby advancing quality control practices.