Blueberry Drone AI at SURP

Rowan’s students in the Computer Science Department are hard at work all year round. Even during the summer, you can find them learning and expanding their skills through Rowan’s Summer Undergraduate Research Program (SURP for short). This past July, students involved in SURP presented the results of their research at the SURP Poster Session. This event was home to a multitude of innovative projects created by Rowan students. One such project, made by Anthony Thompson, Harper Zappone, and Brandon McHenry dealt with the development of drone AI for scanning blueberry farms. 

In order to fully understand this research project on drone AI, it’d be best to first understand the problem that Anthony, Harper, and Brandon are tackling. When farming blueberries, or any crop for that matter, it is important to know how much produce your farm is yielding. However, it can be difficult and tedious to count crops in large quantities. This can leave many farmers wondering if there’s a more efficient way to get it done. Well, according to Anthony, Harper, and Brandon, there is!

These students are using deep learning models to train an AI that can be used to scan and count a blueberry farm’s blueberries much more efficiently. Instead of having a person go out into the field and count the blueberries by hand, you could have a drone fly overhead, scan the field, and use an AI program to pick out blueberries from the scanned images and count them up.

If this AI were to succeed at its job, surveying blueberry farms would become a synch, which means a lot less work for the farmers themselves. However, there’s a lot of different factors to consider when developing this AI. For example, not only does the drone AI need to be trained to recognize individual blueberries, but it also needs to recognize blueberry bushes. That way, the drone knows where to fly and aim its camera. It also needs to be able to recognize rows of bushes, so that it can fly between the rows and get close enough to the ground to take photos. By using these different detection models in tandem, they can ensure that the drone is counting as accurately as possible.

The technology isn’t perfect at the moment, but Anthony, Harper, and Brandon are working hard to make changes and improvements. For example, it is effectively impossible to count every single blueberry with 100% accuracy, partly due to the drone’s limited battery life. There will always be a handful of berries that the AI will miss. So, how do you ensure that the final berry count is accurate? Well, one potential solution they’ve proposed is to manually count the berries in some fields and compare them to the AI’s count. Using that data, they could develop a rough estimate of how many berries the AI misses on average, and then add a constant to the AI’s count to make up the difference.

Building on this, they also plan to test a few different sampling methods, which could be used to estimate the average amount of blueberries on each bush, and by extension, the average amount of blueberries in the entire field. Current methods for sampling include random sampling, stratified random sampling, flying the drone down a few randomly selected rows, and flying to bushes uniformly spread throughout the field.

This research project doesn’t stop at drone AI, however. They have also developed a website and a mobile app in the interest of helping test and train the AI ever further. With all of these developments and more in the training of drone AI through deep learning models, the future of this research project is looking bright!

What we’ve shared here today only scratches the surface of what this summer’s SURP poster session has to offer. There were lots of other talented students at the event who put their research projects on display. If you’re interested in hearing more about these SURP projects, stay tuned for more articles like this in the near future.

Written by Cole Goetz  |  Posted 2022.9.8