Codorniu Group’s Raimat picked our solution for their vineyards.
I’ve been doing some experimentation with Tensorflow (several custom ways) and YOLOv4 for a while on fruit detection & counting. Last week we’ve went out a walnut field, size of a 37Ha for a demonstration, this is it’s blog post.
These algorithms works without a working internet connection. The reason why I don’t use internet is simple, there are lots of scientists and engineers on all fields, that are claiming to solve agricultural production related problems with simple phone apps, but it is not possible because there are often times, there is no cell reception at the field.. where is no reception, there is no internet. That is why I am an AI at the Edge expert for Precision Agriculture.
I got seriously bored at home during this pandemic and started testing a lot of stuff that both can improve my knowledge of inner workings of object detection and NLP.. I of course abandoned NLP because it is boring.. so I focused on using YOLOv4 , it’s accuracy & precision & speed was off the chart and it was easy to train, but there was a catch, you had to label “everything”. literally.
So I did. I’ve generated several videos that you can see on yieldestimator page that works flawlessly.. that started to get some attention from Turkish farmers and companies. One of them was Serdar Dikbaş, who is Sales Manager of Antonio Carraro, orchard type of tractors, and it was a perfect fit for my experiments.
He is the host of Tarımsal Technology (Agricultural Technology) program on AGRO TV, and we had a skype call type of program a few weeks ago. At the program, he said, we must do a real life application of this stuff.. and so we decided to..
After some tinkering with the algorithm, I was able to collect & label & train & deploy the data within an hour..
So I thought I was ready..
Serdar sent the coordinates and we hit the road at 5:45 am
Arrived to the destination around 9:08 am and started collecting data..
Of course it wouldn’t be same as collecting data from a single tree or youtube videos but, the tests I’ve made on previously aired show about walnuts were really accurate so I thought, what could go wrong 🙂
I took my 2 UAVs, 6 Cameras, workstation and my laptop, to the field, set them on the table and started labeling..
At first I collected the data with little action cams but the thing supposed to carry them didn’t stick to the window.
So I had to deploy Plan B.. I connected my gaming webcam to my thinkpad and started walking around in the orchard..
Extracted the frames, started labeling…
When they were in the training, I charged my UAV and had a short flight above the orchard..
When training ended, I ported the model to jetson xavier, connected camera to the window, and put my Spectralix Tablet to work,
the problem was, I didn’t realized the colored glass of the tractor so accuracy was around 40% inside the tractor, 70% outside..
But fine tuned model reached above 90% so no problem..
It wasn’t all that bad.. I labeled and trained tree detection and its accuracy was 94% and it was excellent.
Since all of these done within an hour (per model) it might one and only demonstration on the world and its one of its kind 🙂
After successful demo, I had fun with other toys..
Now I have to fit the camera, processor at the size of my Spectralix Tablet to available for all farmers.
Why we done this demo? Answer is simple. To show that we can estimate yield on orchards per tree basis with AI.
At the end of the day, you can’t roam around the orchard all day long..
And we’re getting older.. Young generation is really fast to adapt AI and they will continue on creating next agricultural revolution.
To sum up..
Antonio Carraro is the best piece of engineering and artwork designed for orchards. I’m sure my tech will combine it very well.
Why I need to use really high-end computing? Because it is what works, for now..
And why UAVs is a must have for agriculture? Because they can scan entire field on a single charge..
I have to thank Serdar Dikbaş, Dr. Cem Aldağ and Ziya Meral for their support for this demonstration.
stop stealing my ideas and come up with your original idea. you know who you are, because I know who you are.