Prediction of physicochemical properties of strawberry fruits using convolutional neural network-regression models

Abstract

Timely monitoring and precise estimation of physicochemical properties, such as pH, total soluble solids (TSS), and firmness, are crucial for assessing the quality and ripeness of strawberries. Therefore, this study examined the application of convolutional neural network (CNN)-regression models for predicting pH, TSS, and firmness of strawberries based on image data captured by RGB camera. Three CNN architectures, namely a typical single branch convolutional neural network (CNNtl), a parallel convolutional neural network (CNNpl), and a series convolutional neural network (CNNsl) architectures were developed, and their performance were compared. To develop these models, 600 fruits in six different ripening stages were collected and indexed for enabling the measurement of pH, TSS, and firmness levels, as well as the acquisition of images. Through statistical analysis, significant correlations were obtained among pH, TSS, and firmness in strawberries, suggesting valuable insights into the physicochemical changes that occurred during the ripening process. The pH and TSS levels exhibited a continuous increase from the early to late ripening stages, while fruit firmness significantly decreased throughout the ripening process. Among the tested models, CNNsl outperformed CNNtl and CNNpl in predicting the physicochemical properties of strawberries, which precisely explained the relationship between the image data and the targeted properties. For pH prediction, CNNsl achieved an R2 greater than 0.74 and an RMSE below 0.20. The CNNsl model demonstrated better performance in predicting TSS, with a 9.65% increase in R2 and reductions of 14.34% and 14.51% in RMSE and MAE, respectively, compared to the CNNtl model. Furthermore, the CNNsl architecture achieved the best results for firmness prediction, with an increase inss R2 of 2.74% and 6.92%, and reductions of 9.13% and 16.38% in RMSE, and 8.34% and 15.33% in MAE, compared to the CNNpl and CNNtl models, respectively. The consistency assessment of these models indicated that CNNsl exhibited the highest consistency among the tested models with minimal decreases in R2 and small increases in RMSE and MAE, followed by CNNpl and CNNtl. However, in terms of detection speeds, CNNtl required the shortest prediction time compared to CNNpl and CNNsl. Overall, this study demonstrated the potential of CNN-regression models in precisely predicting the physicochemical properties of strawberries based on image data. The findings may contribute valuable insights in determining physicochemical characteristics of strawberries, emphasizing the importance of advanced deep learning techniques in agricultural applications.