Weed Stem Location from Overhead Imagery using Artificial Neural Networks

Nexus Robotics developed a robust autonomous robot to work in outdoor agricultural settings to perform tasks such as the precision weeding of field crops. The goal was to provide affordable alternatives to costly human labor and environmentally damaging herbicides that were traditionally required to eliminate weeds from farmed crops. To achieve this, Nexus Robotics combined the latest in robotics, artificial intelligence, and power supply systems. A key aspect of their robotic solution was computer vision software components that could accurately identify objects from video images captured by the robot’s camera(s). Nexus Robotics worked to identify the key computer vision problems and the best machine learning solutions that could be used.

 

Dataset

Data type: Images                               Instances: 1100                                             Associated Task: Recognition                       

Subject Area: Agriculture, Computer Vision, Robotics

Additional information:

Nexus Robotics created a robot named R2-Weed2 to make weeding cheaper and faster than using human labor or herbicides. R2-Weed2 moved through crop rows on its own, using a gripper arm to pull weeds. It had two cameras: one on a ‘delta’ arm and another on the base, taking pictures every 2 seconds. These pictures were analyzed by a neural network to find where weed stems emerged from the soil, which the delta arm then removed. The main challenge was accurately identifying the weed stems, a task called image segmentation, using deep learning.

The overhead camera images were sent overseas for manual labeling with Nexus’s special software. In the labelled images, the green part indicated the plant and red part indicated the weed. It had been further enhanced by pointing out the stem emerging point of the plant and weed.

Demonstration

The video below demonstrates 100 stitched images: the original image on the left, the hand-labeled segmented image on the centre, and the stem emergence points on the right.

Related work

Date Publication Author
September 2020 CROP AND WEED STEM CLASSIFICATION WITH CONVOLUTIONAL LONG SHORT-TERM MEMORY NETWORKS

Dieudonné N. D’Arnall