Automated systems are turning from the “way of the future” to our present lives. Artificial intelligence such as ChatGPT has shown there are impressive ways these programs can adapt and learn in the middle of a conversation or with new data input.
There are still plenty of kinks to iron out, however.
Xin Sun, associate professor of agricultural and biosystems engineering at North Dakota State University, was a part of a 2023 study looking at “deep learning,” or ways in which artificial intelligence is improving its ability to detect weeds.
He said this is still an evolving process, but significant progress is being made with site-specific weed management.
“Deep learning has contributed significantly in precision agriculture domains involving disease detection, crop plant and row detection, stress detection and weed management,” Sun said. “There are procedures that use weed location and density to determine the right rates and spray herbicide in real time based on the AI brain.”
The evolution of this technology has gone from what Sun called “traditional” image processing methods to more sensor-based mapping.
“It is a faster process that involves data collection and processing on the go,” he said. “All the processing is accomplished in real-time by leveraging the application of machine learning or deep learning techniques during in-field motion of ground or aerial technologies.”
Deep learning is a subset of artificial intelligence and is a “giant leap” toward improving precision agriculture, Sun said.
“Years ago, no one would have imagined that one day unmanned ground robots and unmanned aerial systems could be enabled to monitor crop plants and eliminate weeds,” he said.
Sun said the deep learning term is in reference to the many layers of data the machine is inputting and learning from. Unlike traditional machine learning, it is able to create more complex solutions.
“The key reason why industries and university researchers are adopting this approach is because deep learning has the ability to sift through unstructured and large-scale data,” Sun said. “This data is usually in the form of audio, video and images where the algorithm tends to perform classification and detection tasks on similar distribution.”
The effects of this technology on future advancements means a network may be able to be trained from scratch, Sun said. Many weed-detection approaches are using a pretrained model, but with a human-trained program there may be inherent bias brought in. Any data labeled incorrectly by a human or weight given to a certain plant may lead to mistakes.
“There could be added noise in a video or audio recording, or a pixelation of an image that could fool a model into classifying something inaccurately,” Sun said.
An AI-trained model would allow for fewer biases, but there are other issues at play.
“The challenge there is it will demand hefty hours for data collection, preprocessing and annotation process,” Sun said. “In some cases, training may also take weeks, provided the computational resources are high-end.”
Sun said there are many opportunities for artificial intelligence to be a part of agriculture, even more so than we are seeing today. He mentioned while weed detection and herbicide application is a part of their research, topics such as nitrogen management and soil health, along with other precision agriculture aspects, could be impacted sooner than people may realize.
