Sugarcane stem node detection using computer vision and convolutional transfer learning

Abstract

Abstract Quantity of seed needed for sugarcane planting can be reduced by employing the single bud planting method. This approach involves utilizing the stem node or bud extracted from harvested sugarcane for replanting. Currently, this process is manually done using traditional planting techniques. To automate node/bud planting, this study developed a computer vision system based on artificial intelligence. The purpose of this system is to automate the recognition of the stem node, where the sugarcane buds are naturally located. By automating the recognition process, this system can facilitate the automation of node cutting equipment and prevent bud damage during the cutting phase. In this study, a transfer learning-based approach was employed, harnessing the capabilities of a pre-trained convolutional neural network to adapt to the specific application of stem node detection. Transfer learning was chosen to avoid the need for developing and training a new model from scratch, thereby saving time in model development, and enhancing accuracy. Transfer learning allows us to leverage the knowledge acquired from a previously trained machine learning model and apply it to a different but related problem. The goal is to exploit the learned knowledge from one task to improve generalization in another. Specifically, the weights learned by a network during "task A" are transferred to a new "task B." Four different transfer learning models were trained and tested in this work: VGG16, ResNet-50, EfficientNetb7, and MobileNet. Practical tests demonstrated that VGG16 exhibited the best performance in stem node detection accuracy, achieving approximately 98%, with an average image processing time of 0.182 seconds per detection. The developed model and methodology hold broad applicability for automating stem node/bud-based sugarcane replanting.