Technology enables precision viticulture

Sensors and platforms for data collection

Data acquisition through sensors is the first step in digital viticulture. Poblete-Echeverría lists several non-invasive technologies for use in sensors. The most familiar is digital photography based on visible light. “Normal cameras are a great low-cost technology,” he says. “Plenty of applications exist for simple cameras, especially when you attach them to a drone or vehicle. You can also use time-lapse photography.” Other digital imaging technologies, for example, spectroscopy, multi- and hyperspectral imaging, and infrared thermography, rely on non-visible light, while chlorophyll fluorescence estimates photosynthesis by measuring the light emitted by leaves. Poblete-Echeverría has published several papers investigating these technologies in vineyards, including his most recent work mapping vine water status with drone-collected multispectral images. Besides light-based sensors, some sensors detect electrical resistance and conductance, which are used for soil analysis. For sensors to collect data from a whole vineyard, they must be mounted on platforms. Options include satellites, manned or unmanned aircraft, ground-based vehicles, and robots. According to Poblete-Echeverría, drones represent a radical shift for air-based sensors. “We are almost in a moment where unmanned aerial vehicles are fully automatic,” he says. “And we can have excellent resolution with the images we obtain.” Ground-based sensors can be mounted on tractors or quad bikes and linked to global positioning systems to obtain accurate spatial data. “Robots are coming – it’s certain that they will perform specific tasks and capture information,” says Poblete-Echeverría. The Department of Viticulture and Oenology at SU, in collaboration with The Council of Scientific and Industrial Research (CSIR), has already implemented a robotic platform specifically designed for vineyards. Called Dassie, the robot trundles up and down the work rows, carrying sensors to collect data about each vine it passes.

Putting artificial intelligence to work

“Today artificial intelligence is a hot topic – we’re living in a revolution with artificial intelligence,” says Poblete-Echeverría. “It’s amazing what we can do.” Artificial intelligence is the umbrella term applied to tasks in which computers acquire and apply data and skills. The computers are programmed to simulate human reasoning to draw conclusions from data. “In the traditional programming method we used in the past, you generated a model based on the physics, physiology, and rules of a process. You passed that model and your data to the computer, and it gave you a result,” explains Poblete-Echeverría. “Today, with artificial intelligence and machine learning, we give the computer data and expected results, and it creates the model for us. So, the machine extracts the patterns – it’s not given rules – and we can use its model to give us final results as an output.” Machine learning is the subset of artificial intelligence that mimics human learning. Likewise, computer vision is analogous to human vision: the machine acquires, processes, analyses, and extracts data from images and converts the data to information. Data is the raw material for artificial intelligence. Computers need it to train themselves, and researchers need it to test the computers. “Cleaning, preparing, and manipulating the data is very important,” stresses Poblete-Echeverría. “Especially in agriculture because we tend to have databases in different formats.”

Smart applications for yield and quality

To illustrate how data acquisition and processing can be deployed in real-life applications, Poblete-Echeverría discusses yield forecasts based on bunch detection, using an artificial intelligence algorithm called YOLO (You Only Look Once) based on technology similar to facial recognition software. The YOLO algorithm can be trained to detect anything in an image. “We’re training the system to detect bunches on grapevines, and we’re exploring the effect of occlusion,” says Poblete-Echeverría. “Not all the bunches are totally visible, so we’re training the model to estimate the number of bunches per vine.” Other research groups have done similar work, in one case using a version of the YOLO software that can run on a smartphone. “They tested this system in seven cultivars,” says Poblete-Echeverría, “and reported 92% accuracy in detecting bunches.” Researchers have also trained models that can classify bunches as optimal or damaged. Taking technology a step further, scientists have combined artificial intelligence and virtual reality to assist thinning. A human wearing a virtual-reality headset sees a bunch as a collection of yellow berries, with ones earmarked for removal coloured red. Poblete-Echeverría believes that new technologies have great potential for assessing berry quality. He is involved in a project looking at multi- and hyperspectral imaging as non-invasive tools for determining total soluble solids and total acidity in table grapes. The images can be taken in the vineyard without disturbing the developing bunch. The results highlighted the critical role of the light spectrum used to take the image, so Poblete-Echeverría has initiated a SATI-funded project focusing on physical and chemical bunch characterisation using machine vision.

Disease monitoring and beyond

Unsurprisingly, the scope of smart appli-cations extends far beyond fruit. A new paper by Poblete-Echeverría and co-workers addresses the detection of downy mildew by artificial intelligence. “The symptoms are very small, especially in a big canopy, so they’re difficult to capture in the field,” he says. “But in the end, we trained the machine to recognise those patterns, so it can identify the affected areas in detail within one or two seconds, which would take a human expert one to two hours.” In future, a camera mounted on a vehicle or robot could survey an entire vineyard, identify mildewed areas, and facilitate variable-rate applications of crop protectants. Technology for similar fine-grained nutritional management is already in use. Although Poblete-Echeverría is confident about the value of new technologies, he has a word of caution. “Artificial intelligence is not something magic. We need to follow certain steps to produce correct models, and models should be trained and tested properly before being released to the market.” When it comes to implementation, companies have started developing products and services, and growers can look forward to augmenting traditional viticulture practices with innovative technologies. “We’ve done all the science. Once you have the databases and the models, the application of the models is not difficult,” says Poblete-Echeverría. “We can help the table-grape industry improve quality and productivity. I think the future is very promising.”