Non-destructive classification of sugarcane milling feasibility using deep learning: A comparative study of VGG19 and ResNet50

Assessing sugarcane quality is crucial for ensuring both economic value and processing efficiency in sugar production. Conventional approaches, such as refractometer-based Brix measurements, are destructive, labor-intensive, and unsuitable for large-scale or rapid field evaluations. This study proposes a non-destructive deep learning framework for classifying sugarcane internodes into two quality categories (< 16 °Bx and ≥16 °Bx) to address existing limitations. Two convolutional neural network architectures, VGG19 and ResNet50, were evaluated utilizing a defined transfer learning and data augmentation methodology. Because of its residual connections, which enable deeper and more stable feature learning, ResNet50 consistently outperformed VGG19, achieving the highest accuracy of 78.85% on the Luar2_Putih dataset. This comparative finding demonstrates that modern residual-based networks provide superior robustness for subtle visual classification tasks in agricultural imaging, while also validating the stability of the proposed two-phase training framework. The study advances AI-driven non-destructive quality assessment by offering a scalable, field-deployable solution that supports sustainable, efficient sugarcane processing in line with the UN Sustainable Development Goals (SDG 2, 9, and 12).