About 15 years ago as a kid I remember driving with my dad/sister to chat with a farmer near Lake King about their latest 'GPS' technology.
Now, driving just near Hyden dad comes across a
@SwarmFarm
robot spot spraying weeds in fallow.
What will we see in 15 years time?
Life is strange.
In the gap between finishing the PhD and starting a new job at University of Copenhagen, I've been diagnosed with early stage Hodgkin's Lymphoma.
Fortunately, surrounded by my amazing family, friends and medical professionals to get me cured with chemo.
So how many spot spraying companies are there now?
With accessibility of tech and the clear margins involved, the spot spraying market has flourished.
Here's a (non exhaustive) thread of all that I've come across.
Using colour to discriminate weeds from ginger in a drone video.
No ML, just computer vision techniques, though suggests opportunities for high accuracy ML detection rates.
@AgriFuturesAU
@Sydney_Science
OpenWeedLocator 2.0 is here!
Features:
1. Full support for picamera2, picamera so you can now use the latest Global Shutter and V3 cameras on the Pi 5.
2. Entirely new enclosure designs, both 3D printable and aluminium.
3. Up to 120FPS processing speed on the Pi 5 at 320 x 416.
Real-time detection of ryegrass in wheat.
This example shows that technology isn't the barrier anymore, it's the availability of opensource datasets for Australian conditions.
E.g. this example would only work well at
@pbinarrabri
@theGRDC
@GRDCNorth
Started in a pandemic and ended in chemo, but this PhD has been such an incredible experience
@SiaSydney
.
Challenging but very rewarding, I'm very grateful for everyone that's helped along the way.
You can read the full thesis here:
Green-on-green (GoG) camera spraying tech is moving quickly. Better algorithms, commercialisation, more crops and more weeds.
So where are we now for large-scale ag?
A quick summary of tech, companies and the future below. 1/
📷🌱🎯🧵
DIY green-on-brown weed detection system in 3 main parts for ~US$200.
1. Raspberry Pi 4 - US$55
2. HQ Camera + lens - US$75
3. OWL Driver Board - ~US$45
(4. enclosure/wiring - ~US$25)
Solenoids/nozzles/plumbing are then the most expensive parts.
New OpenWeedLocator PCB!
Combines power supply (7 - 26V), relays and Pi protection into one board.
Short demo assembly video below.
Designed by my brother, all files available here if you want to build:
A week in, can confirm that chemo is pretty miserable, but also amazing.
Lumps in my neck you could easily see have shrunk considerably.
Impressive science behind it all! Buzz cut next.
Life is strange.
In the gap between finishing the PhD and starting a new job at University of Copenhagen, I've been diagnosed with early stage Hodgkin's Lymphoma.
Fortunately, surrounded by my amazing family, friends and medical professionals to get me cured with chemo.
Very excited to have won first place in the
@ASA_CSSA_SSSA
Precision Agriculture Systems community grad presentations at the
#ACSmtg
this week.
Massive thanks to
@MBagavathiannan
and the TAMU weed science lab for all their help and
@AustAmFulbright
for the opportunity.
We have liftoff!
Green on brown spot-spraying is a go, using some nifty
#OpenCV
@PyImageSearch
contour detection and object tracking.
Transfer learning and 6m expansion next
#agchatoz
After 5 months stuck in a 5km radius of home/hospital, getting to Kalgoorlie/Salmon Gums/Esperance has been fantastic. Have sorely missed the wheatbelt.
Cherry on top seeing
@SwarmFarm
robot with the OWL in person
The OpenSourceAg newsletter is here!
As a starting point for an open-source (OS) community in ag, it will feature:
🌱summaries of latest tech for data, algos, field use
🌱stories from OS founders
🌱projects
🌱community comments
Signup here:
2 years ago
@williamtsalter
and I launched the OpenWeedLocator.
Since then it's been built in 8 countries, adapted for uses well beyond what we envisaged.
Open-source lets everyone innovate.
And lots more to come still!
Seems strange this needs to be said - but agtech shouldn't cost what it is expected to save you.
This only reinforces a need for open source options.
Parts for an OpenWeedLocator are around US$250
@MillerFarm06
@Deere_Salesman
@JohnDeere
The way we are thinking about it is starting with the value creation for the farmer. Think about your chemical program and how much cost you'd save if S&S can reduce that by 60-70%. Most growers tell me "fair" would be Deere charging ~1/3 of that value...
A prototype spot spraying blue lupins in narrow leaf lupins, using early season data with the
@NVIDIAEmbedded
Jetson Nano.
Early days still, but an exciting prospect for the 2020 season with lots to improve.
Our first published results on laser weeding for control of annual ryegrass are now up!
100% control of the 3 and 7 leaf ryegrass plants at 76.4J/mm2 (60 secs @ 25W), which quickly disappeared for the older plants.
@WeedControlLab
@Sydney_Science
What an experience this has been.
Can't thank
@bryceives
@evokeAG
and the Agrifutures team enough for giving me this opportunity.
Onwards and upwards for open source agtech and the OWL.
Public funds. Public recipes.
Currently involved in R&D of autonomous weed control platforms, Guy Coleman is passionate about the importance of open-source technology, and how public tech ‘recipes’ can help everyone innovate.
We’re excited to count
@geezacoleman
as an evokeAG 2024 Future Young Leader 👏
It's fairly often I hear ag sales people justify the > $300k cost of camera sprayers with 'but the farmers will pay it off in a couple of years'
Pricing systems based on estimated ROI in a risky industry just undermines adoption and innovation. Irks me big time.
Recognition of Palmer amaranth in some Texan cotton across eight different growth stages with the YOLOv5 architecture. Data collected for training on a FLIR camera, but run on video a phone.
Always amazing to see how it turns out.
@AustAmFulbright
@SiaSydney
Weed detection in just two lines of code.
Just released the openweedlocator-tools software package for easier integration of both green-on-green and green-on-brown detection.
Check it out here:
Open-source data collectors almost ready to head to Esperance. Ending up on SP sprayers and
@SwarmFarm
robots for large scale image data collection.
@ezi_farmdata
Some awesome work from
@danbloomer1
and team in New Zealand getting an ATV-mounted OWL sprayer going.
Added a display/controller to solenoid nozzle activation.
Very excited to be taking my PhD research into plant and
#machinelearning
interactions to
@TAMU
with
@MBagavathiannan
as part of an
@AustAmFulbright
Future Scholarship.
Lots of opportunity to collaborate and improve weed control opportunities globally through ML weed detection.
Six of our science researchers, students and alumni have been awarded Fulbright Scholarships.⭐️
Prof Paul McGreevy,
@ProfZdenka
,
@GeezaColeman
, Nicholas Hindley, Alice Yan + Ruebena Dawes will undertake further study and research in the US.
Read more: 👏
Recognising blue lupins in narrow leaf lupins near Geraldton, WA
YOLOv5 L trained on 217 images from
@WeedAI_news
dataset. Training images are younger/yellower soil, so quite a few false detections here.
@nick_mckenna1
There's something very satisfying about turning code into real-world movement (one of the reasons I enjoy Python so much!).
Not too long now until the first prototype implement for ginger site-specific weed control is set up!
@AgriFuturesAU
@SiaSydney
@Sydney_Science
Three years ago today,
@williamtsalter
and I launched the OpenWeedLocator as a new approach to offering weed detection to farmers.
DIY, low-cost and accessible.
New OpenWeedLocator PCB!
Combines power supply (7 - 26V), relays and Pi protection into one board.
Short demo assembly video below.
Designed by my brother, all files available here if you want to build:
Found a local stash of blue lupins on the afternoon walk.
34 images annotated before dinner (27 used for training). Quite overfitted (more data needed).
Still, pretty cool YOLOv8 result:
@grassrootsag
@WeedAI_news
@WeedControlLab
@WeedSmartAU
@williamtsalter
Here's a bit of testing we did on some tracks around where we work.
The algorithm is picking out green colour and spraying just those instead. No machine learning involved - rough cost of a single detection unit is ~$300 (Raspberry Pi, camera, relay board, power supply).
Research on the performance of image-based weed detection on grass/broadleaf up to 30km/h now up online.
Using the OpenWeedLocator, the Arducam camera could find up to 79% of oats and 92% of radish at 30km/h.
@Macintyre_an
@williamtsalter
@SiaSydney
About 50 years ago, the tech behind reflectance-based 'green-on-brown' sprayers was published.
Since then a lot (!) has changed.
We went through ~175 papers from the last 50 years of weed recognition research for this review.
Challenge
First farmer/farmer group to send 1k images of a weed in-crop, I'll annotate for free, upload to WeedAI and use to train a public, benchmark model for the OWL.
Rough requirements:
1. about 1 - 5 weeds per image
2. top down (examples below)
3. decent quality/consistent
2m spot spraying boom in the making for the
@MingenewExpo
next week, with
@eyres_N
welding up the frame.
Will be demoing the
#OpenWeedLocator
and talking about how open source tech can help bring the next generation of students into ag!
As someone who has never used tkinter before,
#ChatGPT
let me build a basic Python 'app' for a niche problem in about 3 hours.
It makes lots of errors, but the efficiency gained is huge.
Full code to use this (for any sort of colour-based analysis):
Can anyone point me in the direction of software which can differentiate colours into percentages. Looking to calculate losses from Gilgai’s
@jakeonfarm
Weed detection technology is really going to give us some pretty unparalleled insights into weed movement.
Combine that with species, growth stage layered with weather/field data...
@BartParks
Dude there’s so much blowing in off neighbors in the Golden Ghetto. We mainly treated edges this week and if you look around you can see exactly where it’s coming from. See and Spray map tells the story
Seeing some decent control of large ryegrass plants with the 25W laser.
4 mins on the left; untreated on the right.
4 mins is clearly too long, but a 1000W laser for example, would see the same energy delivered in just 6 seconds.
Currently getting 30-40cm weeds, but looking for 10-20cm weeds in the future with satellite imagery
@data_farming
@TimNeale1
Impressive opportunities for reducing capital expenditure, at trade off of getting smaller weeds
All files for V2.1 of the OWL driver board now up.
Four channel control via GPIO/PWM. Neat hat for robotics automation too.
Designed by my brother, a big fan of open hardware.
Exciting times - new OWL enclosures arrived.
Rough draft of the picamera2/Raspberry Pi 5 version running now too. Will merge it once tested/properly written up.
Give it a whirl here:
Gordon (the robot) is looking for a new home!
A short history of the construction and why you could be the proud owner of a 2 x 1.5 x 1.5 m robot🤖
Features:
> 350W motors x 2
> thin wires/dodgy connections
> 400W solar
> eBay parts
> Bunnings aluminium
> Here+ RTK GPS/Pixhawk
Real time detection with the OWL at 16 kph.
60 FPS at 640 x 480 resolution. 20cm gap between OWL and the light.
Red lines indicate field of view of camera.
Open-source development and investment can transform agricultural research.
Super appreciative of
@evokeAG
for giving a platform to this idea at one of the largest agtech events in Australia.
Introducing
@GeezaColeman
, a
#2024FutureYoungLeader
and the brains behind the
@openweedlocator
(OWL) and Weed-AI platform.
The OWL is an open-source, DIY, low-cost fallow weed detector, transforming precision weed control.
Watch Guy's application video
The variability in soil type in NW NSW (near Bellata) is fascinating.
Demonstrates why precision agriculture is so useful for managing variability.
#WorldSoilDay
Images are everything in
#deeplearning
!
E.g.
#ImageNet
has >14 M images in 20k classes. Nothing of the same scale exists for weeds (
@WeedAI_news
is trying!) but there are open datasets.
For weed recognition, here's a 🧵of all weeds datasets I've found.
📸🌱🧵
At least 15 different commercial spot spray systems were on show at
#Agritechnica2023
.
Well up from 5 prototypes in 2019.
Check out the summary thread below
Introducing the web-based Coverage Mapping Tool!
A colour-based method for calculating % coverage of two-classes in images.
1. Go to:
2. Upload images
3. Set thresholds
4. Check images
5. Save stats/download CSV.
Built with ChatGPT/streamlit.
As someone who has never used tkinter before,
#ChatGPT
let me build a basic Python 'app' for a niche problem in about 3 hours.
It makes lots of errors, but the efficiency gained is huge.
Full code to use this (for any sort of colour-based analysis):
Raspberry Pi AI weed control sandwich.
OWL board on top,
@PineberryPi
AI HAT with Coral Dual TPU underneath and global shutter camera.
Missed the memo that RPi 5 has different size camera cable.
The
#OpenWeedLocator
is almost here! Entirely opensource software, hardware for community-driven, site-specific weed detection and control.
Build your own green-on-brown weed detector from all cheap/simple/off-the-shelf parts.
Stay tuned for the release in the next week!
Far too long since my last kanelsnegl og rugbrød 🇩🇰
Very much looking forward to the next 3.5 years as a postdoc with
@KU_PLEN
in Denmark working on the OpenWeedLocator and adaptation of weeds to image-based detection. Plus getting back to learning Danish again.
I think the accessibility of these high performing algorithms just reinforces why 'green on green' tech should never be hidden behind million dollar projects.
It's accessible to everyone, and support should be provided to assist that.
Found a local stash of blue lupins on the afternoon walk.
34 images annotated before dinner (27 used for training). Quite overfitted (more data needed).
Still, pretty cool YOLOv8 result:
The
#OpenWeedLocator
is now out!
All the code/instructions/results available below.
Massive thanks to
@williamtsalter
and Michael for helping get this done.
Looking forward to seeing some completed units!
The software, hardware, guides & supporting preprint for the OpenWeedLocator (OWL) are now available.
OWL is an entirely open-source, low cost, image-based green-on brown weed detection device run on the
@Raspberry_Pi
.
GitHub:
@SiaSydney
@Sydney_Science
New preprint on the OWL is up!
With honours student Angus Macintyre, we (
@williamtsalter
) looked at how the OWL performed with 4 cameras, on oats/radish 'weeds' at speeds up to 30 km/h.
Arducam found 95.7% of weeds @ 5kmh, 85.7% @ 30kmh.
@SiaSydney
A factory fitted camera spot spraying system for fallow spraying (not green on brown sensor based) just launched by
@JohnDeere
Claims of 77% reduction over blanket application from testing on ~ 30,000 ha.
Cameras leave the door open for in-crop use.
Exciting times - new OWL enclosures arrived.
Rough draft of the picamera2/Raspberry Pi 5 version running now too. Will merge it once tested/properly written up.
Give it a whirl here:
A good day out in the field helping
@SiaSydney
honours student Angus compare
#OpenWeedLocator
cameras at different speeds - 5-30kmh.
New settings seem to go alright at 30 kmh!
@williamtsalter
After quite a few iterations, finished the new OWL controller for small (< 4 camera) systems.
Toggle spot/blanket spraying, data recording, and sensitivity.
Status light indicates errors/weed detections/data recording.
If I had $1M to develop effective weed recognition, I'd invest it entirely in large scale, open source image and associated ag context data collection.
With the abundant evidence for data importance, still boggles my mind the lack of interest in funding large scale programs.
interesting post from an OpenAI employee claiming that all large language models reach the same endpoint regardless of training strategy or clever tricks
this is of course what the bitter lesson teaches us but useful to get an up to date confirmation that it still holds true
Very glad I could finally see the impressive high speed field camera for weed recognition by Mads and Søren at Aarhus University.
Super clear image collection at 50kmh (!) in all conditions
ChatGPT4-Vision is wild - correctly made the link between glyphosate - surviving weeds and the patch of green in the bottom corner.
Photo from Argentina with
@VientoSurAgro
and the crew
Dad has been using in field mulching for fallow weed control and increased organic matter for some time.
He reckons having a thick mat of OM keeps moisture around well into December (in Esperance, WA). No sand blowing either.
As the saying goes: if we had weeds for faces, weed detection would have been done/mastered years ago.
I spoke with
@theGRDC
recently about weeds, lasers, face recognition and how this is the future of weed management.
@Sydney_Science
@AgInstituteAus
A substantial change of scenery for the rest of the year, finally focusing on the PhD and working with
@MBagavathiannan
's group at
@TAMU
on palmer/ryegrass ID in cotton/wheat thanks to
@AustAmFulbright
!
Field trials going in and data collection kicking off this week!
@SiaSydney
Piecing together 20 years of herbicide resistance work by Paul Neve:
"Reducing reliance on herbicides the only viable option for future weed management"
Alternative weed control at
#Agritechnica
Besides vision-guided tillage, there was a surprisingly small group of non-chemical weed control options.
A short summary
Trained on 1011 images of blue lupins from the suburbs, the model does an alright job finding them in a paddock. Images collected with an S21, test video from a 2018 iPhone XS.
Some generalisability. Variable performance between paddock/growth stage - very poor in young crops.
Just finished assembling the compact OWL or OWLet, after some great feedback during the
@MingenewExpo
.
Fits inside 125 x 75 x 3mm RHS steel for smash protection, same features as the big OWL.
Designs/assembly guide to follow!
Had the opportunity to meet with the
@deepagroco
team in Argentina and see their SprAI system in action for fallow weed detection.
Inspired by
@nozzle_guy
, here's a thread on the system.
Hot off the press!
#OpenWeedLocator
driver boards going to make everything a lot easier to assemble.
The HATs sit on top of the Pi and:
- drive 4 solenoids (stackable too)
- GPIO or serial drive
- 12V in to power the Pi/board (no more separate converter!)
Designs to come.