The agriculture industry relies on technological advancement to continue to advance and serve future generations. A recent study conducted out of Jeonbuk National University in South Korea addressed the critical challenge of orchard management, providing a solution that uses deep learning models to conduct the precise segmentation of fruit trees (1). The study, led by Il-Seok Oh of Jeonbuk National University, and its findings was published in Agriculture (1).
Orchard management is essential for successfully growing, cultivating, and harvesting crops. In this study, the researchers examined apple trees. One of the main challenges when growing fruit trees in orchards is its layout, which often results in branches intertwining (1). Intertwined branches introduce a host of complications to important agricultural process, such as yield estimation, phenotyping, pruning, and spraying. Although this knowledge is nothing novel, there has been limited research on the topic to prescribe any practical solutions (1).
Il-Seok Oh and others attempted to contribute a solution to this issue. Their study introduced a novel data set labeled by human annotators, delineating branches and trunks of target apple trees (1). Unlike traditional rule-based image segmentation methods that overlook semantic considerations, the study used deep learning models. Five pre-trained deep learning models were adapted and fine-tuned to suit tree segmentation using this unique data set (1).
Among the five models that were tested, YOLOv8 emerged as the best model. It exhibited exceptional accuracy with an average precision of 93.7 box AP@0.5:0.95 and 84.2 mask AP@0.5:0.95 (1). This achievement signifies a significant leap forward in automating agricultural tasks crucial for efficient orchard management.
YOLOv8 (You Only Look Once version 8) is a deep learning object detection model, and it represents an evolution in the YOLO (You Only Look Once) series. The YOLOv8 model demonstrates strong performance with an average precision of 93.7% for bounding boxes and 84.2% for instance segmentation masks across a range of IoU(Intersection over Union) thresholds from 0.5 to 0.95. These metrics signify high precision in object localization and accurate pixel-level mask predictions, indicating the model's proficiency in object detection and segmentation tasks.
What made this study unique was that it extended tree segmentation to a natural condition where the branches of neighboring trees intertwine randomly. This complexity creates intricate and often ambiguous boundaries for tree regions, a challenge the research effectively addresses (1). The publicly available data set constructed in this study serves as a valuable resource for future research in this domain (1).
As a result, the deep learning models demonstrated its potential to segment fruit tree images captured in natural orchard conditions using low-cost RGB (red, green, blue) cameras (1). The precise segmentation achieved by YOLOv8 opens avenues for applications in yield estimation, where the segmented tree region aligns precisely with the area for fruit counting. This helps automate specific agricultural processes.
Despite the groundbreaking progress, the study identifies limitations in the data set size and diversity. With only 150 labeled images, expanding the data set to encompass variations in fruit types, tree ages, seasons, weather, and training methods is imperative (1). However, the labor-intensive nature of image labeling necessitates alternative approaches like few-shot learning to scale up the data set practically (1).
As a result, this study out of South Korea diagnoses a solution for improved orchard management by employing deep learning techniques that can achieve precise fruit tree segmentation. The promising results and future directions outlined in this research hold the key to enhancing efficiency and accuracy in various agricultural practices, marking a significant milestone in the intersection of technology and agriculture.
This article was written with the help of artificial intelligence and has been edited to ensure accuracy and clarity. You can read more about our policy for using AI here.
(1) La, Y.-J.; Seo, D.; Kang, J.; Kim, M.;Yoo, T.-W.; Oh, I.-S.Deep Learning-Based Segmentation of Intertwined Fruit Trees for Agricultural Tasks. Agriculture 2023, 13 (11), 2097. DOI: 10.3390/agriculture13112097
Deep Learning Advances Gas Quantification Analysis in Near-Infrared Dual-Comb Spectroscopy
May 15th 2024Researchers from Tsinghua University and Beihang University in Beijing have developed a deep-learning-based data processing framework that significantly improves the accuracy of dual-comb absorption spectroscopy (DCAS) in gas quantification analysis. By using a U-net model for etalon removal and a modified U-net combined with traditional methods for baseline extraction, their framework achieves high-fidelity absorbance spectra, even in challenging conditions with complex baselines and etalon effects.
AI-Based Neural Networks Revolutionize Infrared Spectra Analysis
May 13th 2024A Researcher from Lomonosov Moscow State University has developed a convolutional neural network (CNN) model for Fourier transform infrared (FT-IR) spectra recognition. This AI-based system is capable of classifying 17 functional groups and 72 coupling oscillations with remarkable accuracy, providing a significant boost to material analysis in fields like organic chemistry, materials science, and biology.