Training a computer to detect weeds

[gavztheouch]

Has anyone tried to train a machine learning model to detect weeds? I'm currently trying to retrain a tensorflow object recognition model to detect broad leaved docks in grassland with the intension of building a land drone that uses rtk GPS to autosteer and autospray weeds.

Looking to start an open source project to compete with John deeres new company Blue Rivers new tech 'see and spray' which also uses machine learning for the detection side of things.

Recent advances in machine learning have made this application relatively low hanging fruit for AI. Would be cool to see a land done seeking out and spraying any weed you want.

 

[Northern farmer]

I know nothing about the topic but we had a program on the TV talking about future food production and a robot developed at the University of Sydney. It was detecting weed and spraying each one individually. Perhaps a university would share their knowledge to some degree at least.

 

[Pheasant Surprise]

Has anyone tried to train a machine learning model to detect weeds? I'm currently trying to retrain a tensorflow object recognition model to detect broad leaved docks in grassland with the intension of building a land drone that uses rtk GPS to autosteer and autospray weeds.

Looking to start an open source project to compete with John deeres new company Blue Rivers new tech 'see and spray' which also uses machine learning for the detection side of things.

Recent advances in machine learning have made this application relatively low hanging fruit for AI. Would be cool to see a land done seeking out and spraying any weed you want.

I read about the JD acquisition when it was announced a few months ago and posted it here. JD are clearly very serious about this tech ~ $300M is not a trivial sum of money.

How far along your project are you? Sounds very cool. Did you get anywhere with UoS as @Northern farmer suggested?

 

[Slick]

https://agriculture.trimble.com/precision-ag/products/weedseeker/

 

[Pheasant Surprise]

https://agriculture.trimble.com/precision-ag/products/weedseeker/

Do Trimble do open source now? Cor blimey things must be tough.

 

[gavztheouch]

I read about the JD acquisition when it was announced a few months ago and posted it here. JD are clearly very serious about this tech ~ $300M is not a trivial sum of money.

How far along your project are you? Sounds very cool. Did you get anywhere with UoS as @Northern farmer suggested?

Haven't followed up the UoS lead yet. Im currently trying to detect broad leaved docks against a grass field background. I have taken some photographs about 300 in total and marked them up with coordinates so the computer can train on them. I guess this will be proof of concept, if this works then the rest could all be done with some work.

Some big hurdles are.

1. Gathering labelled images of the desired weeds to spray.
2. Finding a way to take fairly clear photos, don't need to be high res just somehow get them in focus with all the shaking, maybe some sort of suspension mount.
3. Propulsion thats cheap, we might need some sort of battery breakthrough or maybe solar? think slow and steady. Time is maybe less of a concern if you can leave it be
4. Decent gps, prob basestation with path planning and steering angle readout.
5. Some sort of chassis to make it autonomous. 3 meter spray width?

If all the parts all fairly cheap say less than £2000 in total and you can fix it yourself by replacing, rebuilding and uploading then you might be willing to watch from a far without too much concern of it going haywire.

 

[Richard Smyth]

Seems drone mapping is being trialled down here at the moment. A vr map can be produced from the drone images to allow a standard sprayer to spot spray.

I believe it is being used to spot spray broadleaf weeds in lentils. Have no links but shall put them here if can find them.

Weedseeker posted above has been on the market for years. Didn’t think it was Trimble so I assume they have bought it recently.

 

[Pheasant Surprise]

I know nothing about the topic but we had a program on the TV talking about future food production and a robot developed at the University of Sydney. It was detecting weed and spraying each one individually. Perhaps a university would share their knowledge to some degree at least.

Some stuff here from Mashable...

 

[Slick]

Seems drone mapping is being trialled down here at the moment. A vr map can be produced from the drone images to allow a standard sprayer to spot spray.

I believe it is being used to spot spray broadleaf weeds in lentils. Have no links but shall put them here if can find them.

Weedseeker posted above has been on the market for years. Didn’t think it was Trimble so I assume they have bought it recently.

Weedseeker has been a Trimble product for at least five years.

 

[Richard Smyth]

Well there you go. Don’t hear much about it any more as the weedit has taken over here

 

[Briantee]

Look up OpenCV. Haven't had time, but am trying to do much the same thing.

 

[gavztheouch]

Hi Braintee, I have been using Opencv to output images from the webcam into TensorFlow. The plan was to use a neural network/deep learning to classify weeds and then use open cv to tell if the weed is within the spray heads reach. Opencv can run on a raspberrypi.

A Nvidia jetson tx2; like a raspberry pi but with cuda cores for matrix manipulation would be ideal.

 

[gavztheouch]

So far so good. Definitely has potential as this model was only trained for a short time with a small number of images. You can see it thinks my boots are a docken and it misses one of the small purplish ones.

 

[farmerm]

Can the processing be done in real time on board or do the images have to go to a data centre for processing?

 

[gavztheouch]

Real time, but it is limited quite a bit by the speed of the CPU or GPU (GPU is much faster). There are two things happening right now in terms of speed. The algorithms are getting more efficient and the compute is getting cheaper/more efficient. I would guess it won't be long until we have a single board computer that will be able to take multiple camera inputs and return decent detection at high frame rates, which is what we need for spraying weeds. Machines available on the market right now will be traditional computer vision based and not machine learning based.

 

[farmerm]

Real time, but it is limited quite a bit by the speed of the CPU or GPU (GPU is much faster). There are two things happening right now in terms of speed. The algorithms are getting more efficient and the compute is getting cheaper/more efficient. I would guess it won't be long until we have a single board computer that will be able to take multiple camera inputs and return decent detection at high frame rates, which is what we need for spraying weeds. Machines available on the market right now will be traditional computer vision based and not machine learning based.

Once you solve the weed challenge can you create a molehill tracking algorithm fitted to a drone that will alert me when and where a new hill is being pushed up so I can blast it...

 

[Pheasant Surprise]

Once you solve the weed challenge can you create a molehill tracking algorithm fitted to a drone that will alert me when and where a new hill is being pushed up so I can blast it...

That might actually be the one useful task drones are good for!