A while back, I watched Astro Teller’s wired interview and came across Mineral, a moonshot project aimed to collect huge amounts of data and use it to improve the agriculture sector.While watching the video, I was a bit surprised when Astro Teller talked about the difficulty in imaging melons, as they are generally covered with leaves. It struck me how reliant we (and the Mineral rover, which is very cool, BTW) is on images.

I looked up Mineral after that and resonated with its fundamental ideas. Agriculture is definitely ripe for moonshot thinking, and there is a huge pay-off for developing new data-centric methods for performing precision agriculture. The Mineral rover, I felt, is a good instantiation of the idea. It captures huge image datasets which are then used to evaluate plant health, weeds, and map locations and ripeness of fruits (at least that is what I gathered from this video). Mineral seems to use this in conjunction with satellite images to guide what and where to devote agriculture effort (weeding, gathering fruit, maintenance, etc.).

One of the aspects of precision agriculture is controlling for pests. This too is probably part of their image dataset, but it is not clear if this worked well. I suspect this is a hard problem because it is likely that visible signs of pest infestation happen long after substantial damage is done to the plant. The ideal solution is to catch pest infestation before most of the damage is done, that is, before visible signs.

Therefore, for detecting pests, using images might not work because often it is the damage that needs to be prevented that is visible as a sign of pest infestation. I have been wondering about this from the point I watched Astro Teller’s interview.

Maybe a better way to detect pests is to listen in on the vibrations in a plant.1 Pests moving around on or inside the plant will cause vibrations that will be isolated in the plant. Listening to the plant through microphones would allow listening in and checking if there are pests scurrying on the plant, drinking sap, or burrowing through its stem.

Using microphones and laser Doppler vibrometers for detecting plant vibrations might be a good way of detecting pests in plants and fruits. This can be combined with DNNs for continuous signal separation to pull out relevant features to detect the presence of pests in acquired vibration signals.2

Footnotes

  1. I got the idea from An Immense World which describes Rex Cocroft’s research on plant (substrate) vibrations caused by insects - insects using it as a communication channel and plants using it to detect pests and responding to them. ↩

  2. The tricky part is generating training datasets to pull out pest-specific features. But I think this can be done in collaboration with agriculture labs that study pests and maybe pull out a general set of features for different species, genera, and orders of pests. ↩