Prediction of Physico-chemical Properties in Tomatoes Using Deep Neural Architecture

This study develops an image-based prediction model for the physico-chemical composition of tomatoes based on their surface characteristics. The experiment starts with a collection of 120 tomatoes of hybrid variety, analysis of their physico-chemical properties, and image acquisition of the respective tomatoes. Physico-chemical parameters assessed in this study include firmness, color, lycopene content, titratable acidity (TA), total soluble solids (TSS), and pH. The parameters are analyzed against their maturity classes, and correlation is established with a statistical tool, Jeffreys’s Amazing Statistics Program (JASP). One-way AVONA analysis indicates a statistically significant difference (p ≤ 0.05) in mean lycopene, TSS, and TA values. Among the parameters, lycopene showed the highest correlation with TSS (R = 0.921) and TA (R = − 0.858). On the other hand, the least correlated parameter is pH (R = 0.393). Lycopene, being the highest correlated chemical parameter, correlation with other physical parameters such as firmness (R = − 0.910) and color (R = 0.909) is established. The attained physico-chemical results are then mapped against their respective images of three maturity classes. A total of 900 mapped images are then trained in a VGG19 model with 16 convolutional layers and three fully connected layers pre-trained on ImageNet. In classifying tomatoes into their maturity classes, the proposed model achieves a validation accuracy of 92% after 50 epochs with a batch size of 32. The achieved accuracy shows a clear indication of its applicability in industries and researches for real-time maturity detection and prediction of physico-chemical properties from tomato image.