GrapeSense: A Comparative Study of Residual Transfer Learning Models for Grape Aging Classification Using Drone Images

Grape harvesting is a crucial process in the wine-making industry, and the accurate classification of grape aging stages can significantly improve the efficiency and quality of grape harvesting. This paper presents a comprehensive comparative study of residual transfer learning models for grape aging classification using drone images. With the advent of drone technology, it is now possible to collect high-resolution images of grapes in a vineyard setting, providing a rich source of data for machine learning models. GrapeSense is an ongoing project that utilizes an aerial drone and leverages residual transfer learning deep convolutional neural networks (DCNNs) to enable the classification of different ripeness stages of green and purple grapes. The study evaluated the performance of several state-of-the-art residual transfer learning models, including ResNet-50, ResNet-152, DenseNet-201, and ResNetRS-420, on a dataset of drone images collected from a vineyard. The results show that the ResNet-152 model achieved the highest accuracy for the green grapes and the ResNet-50 model achieved the highest accuracy for the purple grapes with ripeness classification using transfer learning. The images used to train and evaluate the model were taken on a biweekly basis across the span of the entire grape-growing period (6 months) allowing for the classification of six different grape aging stages. GrapeSense empowers small-scale farmers and vineyards with automated detection technology, significantly enhancing the efficiency of grape harvesting while minimizing grape yield wastage.