Using Neural Networks to Improve Agriculture

As an ex-researcher in signal processing, I’d like to share how some of this type of research is being used to improve agriculture. I’ve written both a bit about the science and the agriculture benefits.

Researchers are using sophisticated image processing techniques to better identify plants so that less herbicide can be used on our conventionally grown crops. This is obviously a good thing, as herbicides can be quite nasty to the environment and human health.

Robots can identify plants by collecting and processing images

Researchers at the University of Illinois Urbana-Champaign and the United States Department of Agriculture have developed a vision system to identify types of plants. The vision system is mounted on a robot, which captures images as it moves through the field. On the robot is also a computer to which the images are transferred.

After collecting two sets of images at different times, they are analyzed using algorithms by my favorite old tool, MATLAB.

The algorithm – Neural Networks (sciency, but not too sciency – do read on)

The algorithm used is a neural network algorithm.

The idea of a neural network is quite interesting. The human brain is amazing at recognizing patterns – faces, foods, etc – in a split second, you can recognize a face you haven’t seen in years; however, it is very slow at doing computations like addition or multiplication. On the other hand, a computer can do thousands of computations per second, but even the best pattern recognition programs fall short of the human brain’s power.

Neural networks use artificial ‘neurons’, which are simply mathematical models of an actual neuron. Mathematically, the artificial neuron receives a series of inputs, and based on these inputs, it outputs a value. For a simple neuron model, it would either output 0 or 1 – modelling an actual neuron firing or not firing.

The outputs of the artificial neurons can be ‘taught’, much like a human’s brain. By giving the neuron information (the inputs, images in this example) and looking at its output (type of plant) then giving it feedback on whether its response is correct or not, the network of artificial neurons can learn.

How this relates to the plants and agriculture

The algorithm used was trained using various images of plants so that when actual data is collected, the system is able to identify the plants it sees compared to what it has learned.

At testing, the system was able to identify 94% of the corn plants correctly. Not too bad given that this is in the very early research stages.

What a system like this could potentially do is allowed farmers to use such a robot to identify where there are more or less weeds in order to use herbicide more efficiently rather than treating all crops with herbicides as a preventative measure.

Photo: Flickr Creative Commons by Sebastianlund

Source: vision systems




Leave a Reply

Your email address will not be published. Required fields are marked *