Detection and localization of ripe tomato in greenhouse environment using Keras-based deep learning models

The global tomato industry faces increasing pressure to improve productivity and reduce labor costs, particularly in developing countries where manual harvesting remains the norm and is prone to inefficiencies. The development and adoption of automated tomato harvesting systems can significantly reduce reliance on manual labor. In response to the challenges in this field such as accurate detection and localization of ripe tomatoes amidst varying lighting conditions and occlusions, this study focused on developing state-of-the-art deep-learning models for detecting and localizing ripe tomatoes in greenhouses. Five deep-learning models were developed using pre-trained architectures based on MobileNetV2, DenseNet121, Xception, InceptionV3, and ResNet50. The Xception-based model emerged as the best performer, achieving the highest accuracy (96.97%), precision (96.32%), recall (96.99%), and F1-score (96.65%), and demonstrating superior generalization and stability in both detection and localization tasks. The ResNet50, DenseNet121, and InceptionV3-based models also performed well, though exhibited greater variability during testing. The MobileNetV2-based model, while faster in prediction (3.14 ms per image) and location extraction time (33.33 ms per image), showed reduced precision, making it more suitable for speed-critical applications. The findings highlight the potential of deep learning models, particularly Xception, in improving the accuracy and efficiency of precision farming practices in tomato harvesting. • Deep learning models were used to detect and localise ripe tomatoes efficiently. • Transfer learning improves model performances using pre-trained Keras architectures. • Xception-based model provides the highest accuracy for detection and localization. • MobileNetV2-based model is the fastest but shows reduced precision in tasks.