Hacker News Reader

Comment by MayeulC | Hacker News Reader

I don't think you really need AI/CV if you have multispectral or full-spectrum imaging. Different species will have different absorption/emission at multiple wavelength. This makes it easy to identify crops, and seems to be commonly used for aerial imaging[1-3].

That kind of sensor is expensive though, and while you could probably do it for cheap with something like a DLP wheel (edit: or an array of different light emitters) plus a B/W camera, ML might be more price-effective, though probably more error-prone, so it isn't a given if you want a high match rate.

Also, isn't this a fire hazard?

[1] https://en.wikipedia.org/wiki/Hyperspectral_imaging#Agricult...

[2] https://en.wikipedia.org/wiki/Multispectral_image

[3] https://en.wikipedia.org/wiki/Full-spectrum_photography

View on HN

> This makes it easy to identify crops

...until they evolve to emit the same spectrum of light as the crops being harvested! Cuz you know that's going to happen.

Well, that's right, and I didn't think of it. But besides what the sibling comments pointed out, that's an issue with any kind of weeding, including manual.

The answer is to select according to a wide range of criteria, and not rely on a single one. That way, weeds cannot progressively acquire resistance, and need to check all the boxes at once, which makes it highly unlikely that they will pass on their "slightly better" genes.

And of course, the larger the scale you employ a single weeding system at, the more risky it gets. It would be great for supplementing herbicide (while lowering doses) at first, for instance.

This is the part where Monsanto starts engineering crops to fluoresce in a particular frequency of light, so you can target weeds inversely.

That makes me think that this also makes crop-rotation obsolete. You can plant 20 different crops in the same field, in repeating sequence. The bot will take care of them as-needed, and next season, move everything down one slot.

This is a known process already https://en.wikipedia.org/wiki/Vavilovian_mimicry

Examples of this are Rye and Oats.

That’s a slow process, further you can always rotate between crops with different spectra.

It sounds like a slow process but plant life cycles are yearly compared to 20+ years for human generations to spawn, it will happen faster than you think. There are already round up resistant weeds and it's only been 50 years.

Even 30 years after first implementation would be several generations of better hardware and software. Spectra is just one way to get v1 out the door, there are many ways you can improve things.

Crops are planted in a specific pattern, at a specific time, grow at a specific rate, and have a unique shape. So you have a lot of information to work with.

Bacteria and viruses do that, but weeds first don't reproduce as fast and second humans can see them and find out something like treating crop seeds with fluorescent color.

Eventually, the only organisms that remain will be mutualistic ones, if we do it right.

We just need the weedbots to test plants for flavour and nutritional value; cf. xkcd.com/810

I think the AI solution is actually pretty well proven. I use PlantNet to identify weeds in my yard and it's plenty accurate and doesn't require any sort of advanced sensor.

Could be, but having more discriminating data as an input can only improve speed and accuracy. I don't know what technology they picked for the first iteration, but can only imagine they'll end up using every trick in the book going forward.

A slightly more advanced sensor or lighting apparatus wouldn't cost much more, proportionately to the cost of the whole system.

The cost would be acquiring a training data set that includes data from the more advanced sensor, versus existing catalogs with many millions of images and tens of thousands of species.

> Also, isn't this a fire hazard?

Unlikely. Heating a plant to kill it is considerably lower power than heating a plant to combust it and turn it to ash. It would also be trivial to add a water sprayer as part of the process.

> you have multispectral or full-spectrum imaging

...that's exactly what CV needs to handle.

Well, right, but it doesn't need particularly impressive algotithms to achieve something with that data. Those algorithms were here 20 years ago, with less processing power and camera resolution.

A lot of modern ML is not that impressive either.

I can imagine discriminating species of a hundred leaves in a very high resolution images to be very challenging.

The leaves are 3-dimensional objects can be twisted, tangled, curled up, broken, half rotten, sunburnt...

Reply on news.ycombinator.com