Multi-task deep convolutional neural network for weed detection and navigation path extraction

doi:10.1016/j.compag.2024.109776

Multi-task deep convolutional neural network for weed detection and navigation path extraction

2025-02-01

0

PRE

AI

DOI：10.1016/j.compag.2024.109776

Original

Original request for help

Save

Abstract

En 中文

Field management operations, such as spraying and weeding, necessitate precise execution along crop rows and traversable areas to minimize crop damage and operational losses. Vehicular real-time visual navigation, a widely adopted method, can identify crop rows and traversable areas but faces susceptibility in complex field environments and amidst intersecting crops. To overcome these challenges, this study integrates UAV (Unmanned Aerial Vehicle) measurements with the novel multi-task YOLO algorithm, proposing a managed navigation approach based on coordinate information derived from pineapple field navigation area extraction and weed detection. Initially, the improved multi-task YOLO algorithm segments the traversable region of the pineapple field in UAV imagery and detects weed positions. Subsequently, for navigation line extraction from the segmented traversable area, a series of operations, including regional edge extraction, framing the outline area, filtering operation, and navigation line extraction. To generate executable information for the weeder, recognized route pixels and weed localization pixels are converted to actual latitude and longitude information using the software. Validation against professional equipment indicates an error of 5.472 cm. The novel multi-task YOLO algorithm is enhanced by c2f and anchor-free modules. The multi-task YOLO satisfies multi-task prediction while maintaining the close prediction performance of single-task. Experimental results demonstrate the enhanced multi-task YOLO model, showing a 4.27 % increase in training accuracy and a 2.50 % increase in mIoU compared to the original model. The experimental results show that the proposed navigation method for the weeding machine and weed localization method can solve the problems of limited perception ability of ground robots and the weed localization problems.

Keywords:

Pineapple

Weed detection

Navigation line extraction

Deep learning

Fields management

Multi-task deep convolutional neural network for weed detection and navigation path extraction

Abstract

Journal

Organization