Weeding Through New Technology
The demands of the digital age have inspired and challenged the spectrum of innovation. The rapid rate of progress and technological advancements force us to think harder and faster. As a result, we’re churning through more questions, ideas, problems and solutions than ever before – and at record speeds.
I spent my entire career in weed science. As I studied chemicals and plants, I worked to discover active ingredients that could consistently control grass and broadleaf weeds and, as a result, protect crops and secure yield and profit. Over the course of my career, we’ve made great strides in weed control, and the science and innovation behind our work generated lots of ideas about the future of farming.
Now, I dedicate my time and research to our latest innovation in farming technology: using image recognition to identify weeds.
This isn’t the first time we’ve considered alternative methods to identify weeds. The idea to have a system to characterize these plants has been floating around agriculture and horticulture for decades. Weed identification is the initial step in weed control – but it’s also tedious, time-consuming work.
Weed Scientist at Bayer
Misidentification can lead to choosing the wrong herbicide(s), resulting in insufficient weed control as well as partial loss of the financial investment into weed control. Consequently, this error can reduce crop yield and ultimately farm profitability.
At Bayer, we’ve taken the reins on weed control by developing specific software containing an algorithm capable of image recognition to identify weeds. Our current work involves feeding the software system thousands of images of different plant species. The objective of this step in development is training the software to recognize the specific weed in the image based on color, shape, morphology and other visible characteristics.
To “train” the software system, we amass photos from our colleagues and experts in botany that can verify plant species. We receive additional pictures sent in by users of our mobile application, which are checked to ensure proper annotation. To secure best input quality, we rely on the experts to identify images and keep everything we put into the system highly accurate.
Only they can see the details beyond what can be captured with an image: When an expert takes a photo, he/she will also observe the physical plant – outside just the visual characteristics – to ascertain additional clues like stem shape, leaf texture, root structure, scent, latex in the stem, and leaf secretions that all can aid identification. We need this kind of scrutiny to ensure correct documentation and ultimately, to provide correct education to the system. This is absolutely mandatory to beat the “garbage in, garbage out” principle. Then, we monitor whether the system identifies the catalog of photos correctly.
Currently, we’re trying to understand how variations in photo quality affect the system’s output, which is weed identification. Our objective is to allow farmers to feed images taken with their smartphones into our system without requiring any special equipment. In other words, whatever device the farmer uses to access the app should not affect the system’s accuracy. Farmers collect photos from all kinds of mobile devices, and we want to make sure that identification is not compromised by the quality of the image their device takes. To accomplish this, we’re training the system with images from myriad devices so it can identify from any reasonable quality of an image.
Our intention behind this technology is to create a system that offers farmers a reliable, feasible means to identify weeds, so that they can effectively deal with these crop threats.
The first application of our new technology has been incorporated into an app called WEEDSCOUT which is available for both iPhone and Android. And that’s only the beginning. In the future, it may spread to aerial platforms, such as drones, to monitor the density and composition of weed infestation in real time and uncover “hotspots” in the field efficiently. This would support more targeted crop protection applications and better assessment of treatment effectiveness. As we consider future plans, it’s important to keep things like these in mind: frequent and close problem monitoring, targeted applications and rapid success control with limited time input.
Numerous opportunities are opening up in farming with the incorporation of technology. We are dedicated to finding innovative ways to apply these advances, particularly in weed control. Farmers are open to these kinds of innovations, and they see the value of tools such as Weedscout to protect their crops and continue managing their farms sustainably.
As agriculture and technology collide, scientists are pushed to come up with new solutions and keep up with the rapid rate of technological advancement.
Things have changed a lot over the course of my career, but the challenges of the digital age spur more and more innovation in science. I’m excited to be involved in weed science during a time of such tremendous opportunity.