Deep Learning Applications in Agriculture 🌽


(Francisco Ingham) #1

Hey!

I am interested in how deep learning is affecting the way we grow our food. It would be interesting to discuss current and potential applications of deep learning in the various nodes of the agricultural supply chain:

  1. Input (seeds, fertilizers, animal food)
  1. Production (from sowing to harvest)
  1. Processing and storage
  2. Distribution/logistics

Others:

If you are interested in this as well feel free to share your views!


(Brian Muhia) #2

In January, I fine-tuned Resnet50 with the PlantVillage dataset, slightly outperforming a previously published result. I’m building an app, similar to Plantix, to serve this and similar insights to farmers to help with #4, because it takes a while to get the right fungicides, and pesticides if you have to wait for people in Whatsapp groups to answer questions based on photos taken. It is in alpha.

I know the original PlantVillage team has since moved on to using a single-shot detector for real-time, on-device disease detection, but a web-service approach should work for people who don’t have a powerful enough smartphone. This is what I’m doing.


(Kerem Turgutlu) #3

Very interesting topic, thanks for starting it !


(Gavin Armstrong) #4

Hi there im intrested in deep learning for Agri. I am currently building a robot to selectively spray for weeds in a grassland field. My skills are in the building of the hardware but I have been learning a little about deep learning from the fast ai videos. I think it would be a cool project to do open source.


(Francisco Ingham) #5

This is super interesting! I read your post, and just as an idea: is it possible to have a drone recognize weeds AND spray selectively in real time (no GPS coordinates there). I suspect the drone should be a custom build… The major challenge I can think of is the weight of the herbicide (although some drones already have this functionality DJI MG1S) but one could devise a way so that the drone goes across the field and back and replenishes its stock with a moving reservoir that moves slowly across the field’s border as the drone progresses through the field.


(Jon Gold) #6

I’m super interested in this area too - appreciate you starting the thread and posting such great research :slight_smile:


(Francisco Ingham) #7

So… this happened:

Recognize weeds and spray in real time


(RobG) #8

Just this week Barack Obama made a speech at the Univeristy of Illinois. He opened with “It is good to be home. It’s good to see corn, beans. I was trying to explain to somebody as we were flying in, that’s corn. That’s beans. They were very impressed at my agricultural knowledge.”

Well, at the risk of ‘patting oneself on the back’, I scored 2nd place and a $4k prize in CrowdAnalytix’s “Agricultural Crop Cover Classification Challenge - Perform image segmentation to determine which fields are growing corn or soybeans” in Illinois producing predictions as shown below (1px = 30m). I don’t know how it would compare to Obama’s expert knowledge. https://twitter.com/crowdanalytix_q/status/1035147920050475008

Competition rules mean I can’t detail the method, suffice to say fast.ai and U-net and opencv were my friends. And I learnt a lot about opencv and remote sensing in the process.

I would say to fast.ai learners, do not be intimidated or constrained by subject matter: the transferability of what you learn here from dogs v cats to fashion to agriculture to medicine is a key lesson.