Given where ML and CV are today, I'd bet on open source models trained on weeds within 3-5 years. This company can scale, but their unit price is going to plummet. Also, it sounds like they're using cameras, where it's concievable that other future sensors could be more efficient. An iteration of this with higher resolution cameras and small flying drones seems like an intuitive next step.
This is a super interesting problem because the confusion matrix (fp/fn/tp/tn) rate that makes this economical is going to be variable across both crops, and market demand.
If there suddenly there was a demand jump for peas, you could afford to use a model with less accuracy, because you are optimizing betwee a sunk labor cost and margin on your yield. You could literally tune your detection parameters based on futures price data, since if if prices were high, you could optimize compute on your model. Anyway, spoken as a total outsider, but what a cool and interesting set of problems.
This is a super interesting problem because the confusion matrix (fp/fn/tp/tn) rate that makes this economical is going to be variable across both crops, and market demand.
If there suddenly there was a demand jump for peas, you could afford to use a model with less accuracy, because you are optimizing betwee a sunk labor cost and margin on your yield. You could literally tune your detection parameters based on futures price data, since if if prices were high, you could optimize compute on your model. Anyway, spoken as a total outsider, but what a cool and interesting set of problems.