Convolutional Neural Network (CNN)-based transfer learning framework for cherry tomato production

As crop harvesting becomes more difficult in environments affected by climate change, the application of artificial intelligence technology to crop management through accurate yield prediction is receiving worldwide attention. This study proposes a convolutional neural network (CNN)-based transfer learning framework to increase the productivity and improve the economic feasibility of cherry tomatoes (solanum lycopersicum) in South Korea. You-Only-Look-Once 10 Nano (YOLOv10n) is adopted as a CNN-based algorithm. The source model for transfer learning is trained using cherry tomato imagery from the Tomato Plantfactory Dataset, while the target model is trained based on field survey data collected by the National Institute of Horticultural & Herbal Science, Rural Development Administration, Korea. In that process, an image segmentation technique is developed to improve the prediction accuracy, which reduces the root-mean-square deviation of the existing YOLOv10n from 32.3 to 19.8, a 38.7% reduction. Also, the devised economic feasibility analysis method finds the cost of producing cherry tomatoes in South Korea to be 11.12 USD/m2, while the maximum revenue can reach 22.44 USD/m2. As a result, the proposed transfer learning framework helps general farms where it is difficult to collect big data to use machine learning techniques to predict crop or vegetable production. Keywords: transfer learning, smart farming, cherry tomatoes, yield estimation, convolutional neural network, computer vision. DOI: 10.25165/j.ijabe.20251805.9827 Citation: Lim H, Kim Y, Kim S, Kim S. Convolutional Neural Network (CNN)-based transfer learning framework for cherry tomato production. Int J Agric & Biol Eng, 2025; 18(5): 90–101.