ENSIO GmbH designs, develops and supplies image-processing and camera systems for agricultural weeding tines under the brand name OKIO which make real-time identification of crop and weeds possible while farm vehicles are in motion. The company is carrying out a joint research project with Software Competence Center Hagenberg (SCCH) that uses deep learning methods to develop plant models that can not only identify crop but can also precisely locate stems and roots.

Over the past three years, use of cameras for guidance in hoeing technology has become the standard in organic farming; most new machinery is now supplied with integrated camera technology. These systems currently only assist tracking rather than precise identification of individual plants which is, however, essential for effectively removing weeds from amid the crop. For this reason, ENSIO GmbH and Software Competence Center Hagenberg (SCCH) are now employing deep learning methods to develop plant models that use the globally known OKIO camera systems to distinguish seedlings from weeds precisely and in real time. They are also capable of locating the stem and rootstock of seedlings.

Maximizing crop yield

What is new about this next step in the development of precision farming is the deep learning methods applied by SCCH which train individual, particular plant models and neural networks so that they are not only capable of recognizing the leaf shape, density and contours of the crop but also locate its stem and rootstock – no matter what the weather and the light are like. The plant model trained by SCCH therefore supplies information about the location of crop, stems and rootstock to prevent their being damaged by a wayward cut of a blade or other hoeing or weeding implement. These plant models therefore help to minimize crop damage and maximize crop yield. What is more, the plant model can also be used for selective spraying and fertilizing, ensuring that this is carried out only where it is actually necessary. This reduces the amounts of sprays used enormously.

“The combination of ‘organic and conventional plant protection’ is also gaining ground in areas other than organic farming,” says Michael Mayr of ENSIO, looking ahead to the coming years with optimism. “It engenders many intriguing developments that will take account of social, political and environmental requirements.”


Autonomous field robots of the future

Another challenge facing the neural networks is that recognition of individual plants, which has to function while vehicles are travelling at speeds of up to 15 km/h, should also be used to guide the weeding tines autonomously between the rows with the aid of several cameras. To achieve this, individual images have to be processed in less than one hundredth of a second and the results transmitted to the control unit which controls the weeding tines and triggers removal of weeds.

As an aside, in the medium term this will also be the key technology that makes field robots a reality. Initial trials, in which robots take over certain tasks autonomously and without a tractor, are scheduled for the coming year.

Some successes have already been achieved with first crop recognition models which show that convolutional neural networks (CNN) are particularly well suited for this task. The obligatory difficulty is currently the acquisition of the large amount of image data needed to train the neural networks.

Farming without artificial chemicals

Crop and weed recognition that works using artificial intelligence represents another step towards ecological precision farming. It can help reduce the amount of pesticides and fertilizers used.

“In this way, several of the United Nation’s Sustainable Development Goals can be reached,” says Theodorich Kopetzky, project manager at SCCH. “For instance, Sustainable Development Goal 15 requires that soil productivity be increased. We researchers and scientists can contribute to achieving that by developing new technologies.”