Bay Vision - Spring 2018

on these chemicals, because they’re paired with genetically modified crops that are naturally resistant to the herbicide. The idea is that the herbicide doesn’t damage the crop, but it kills the weeds. However, these chemicals are expensive and the broadcast spraying technique is inefficient – the whole field is being sprayed when only a small proportion of the ground is covered in weeds, so there is a huge potential to save. However, these chemicals are expensive and the broadcast spraying technique is inefficient – the whole field is being sprayed when only a small proportion of the ground is covered in weeds, so there is a huge potential to save. This is where Blue River comes in. They estimate 75-90 per cent chemical savings by using intelligent shot-spray techniques. This technique has solved another problem - herbicide resistance. As you can spray on the weeds directly, you can use chemicals that traditionally would not be used in broadcast spraying because they kill the crops. Jim tells us that as they moved into weeding they hoped to use some of the machine learning techniques they used for lettuce thinning; however, they found they weren’t achieving their high accuracy targets and so shifted their focus on to deep learning: “ That’s really been successful. We have used bounding box type methods like YOLO or SSD - the Single Shot MultiBox Detector - and also semantic segmentation, so pixel labelling. We find there are advantages and disadvantages in both, but with both of them we’re seeing good progress. ” Jim adds that there’s another thing that often gets lost in translation between academic papers and commercial work: the notion of generalizability. How do you get your system to work in lots of fields and lots of areas? Many academic papers exist that develop results based data from a field. They can produce high 90 percentile accuracy rates in terms of identification classification, but when you move on to the next field, it gets harder. He explains: “ We spent a lot of time with what we call n+1 testing, holdout testing basically, where we’ll collect data from a hundred different fields - we only need a little bit of data from each field - and do rotating holdout tests where we’ll train on, say, 99 fields and test on that one holdout field. Understanding the performance as you go into a field you’ve never seen before is one of the keys to our success and one of the challenges in getting systems to work, not just from an academic setting of squeezing out an extra percentage or two on a given dataset, but actually having it work in practice as you go into new fields. ” In conclusion, Jim says something which will come as no surprise to people who work on neural networks is that they spend a lot of time working on data management: “ Data management, collecting data and labelling data , is in many ways the secret sauce in deep learning these days. There’s figuring out good ways to build and train the model, but I’d say data collection and maintenance and management are just as big a part of the story for us. ” “How exciting the field of agriculture is for computer vision and machine learning (and data science and deep learning in particular)! There's a lot of opportunity to expand the way in which computer vision can enable smart machines, so it's a great area to be working in right now.” Blue River 19 Blue River Technology Bay Vision

RkJQdWJsaXNoZXIy NTc3NzU=