Artificial Intelligence helps identify sick plants

In Brazil, equipment that allows capturing and simulating brain signals began to be tested this year to detect early-stage diseases in soybean crops using Artificial Intelligence (AI). Embrapa maintains a partnership with the companies Macnica DHW and InnerEye, the latter developer of BrainTech, a technology that should speed up decision-making, reducing losses in rural enterprises and rationalizing the use of natural resources.

The technology can have several applications, being used when it is necessary to generate models that solve a classification problem, coming in to accelerate the data labeling process, and thus generating AI models that work when the traditional approach does not meet the classification requirements.

The system simulates the brain functioning of experts when they view images of diseased plants, automating labeling and making the stage faster and more efficient. The Braintech equipment captures the specialists' neural signals using a helmet with electrodes, in a similar way to an electroencephalogram.

“This is a pioneering initiative by Embrapa that is combining disruptive BrainTech technology, brought exclusively by Macnica DHW to Brazil. By associating EEG neural signals and AI, it is possible to create a machine that imitates the human brain with high reliability”, observes the IoT & AI solutions manager at Macnica DHW, Fabrício Petrassem.

The testing and validation of the system had the participation of developer Yonatan Meir, from InnerEye, who came from Israel last August, especially for this purpose. “By capturing brain waves, InnerEye’s solution is capable of identifying the judgment and classification of an image observed by a person, allowing the image to be labeled automatically and immediately,” explains Yonatan.

The system is already used in European airports to identify dangerous objects in suitcases. In 2019, Macnica DHW approached Embrapa to, in partnership, explore the technology in the agricultural sector with possible new applications. Early detection of plant diseases was where the experiments began, in April.

Experiment Details

“AI tools have evolved a lot and, with good quality data, can solve almost any problem”, says Embrapa Digital Agriculture researcher, Jayme Barbedo, who leads the project for the Company. The challenge, as he points out, is obtaining this 'quality data', which in addition to being collected needs to be labeled by experts. A costly and time-consuming process in which the equipment will assist.

The first results of the experiment were positive, as the equipment helped to identify, with high accuracy, diseased leaves (powdery mildew and soybean rust) and healthy ones. Now, the project must go beyond the detection of diseased/non-diseased plants and move forward in identifying the type of disease present in soybean cultivation, starting with the most commercially significant ones. The inclusion of corn and coffee crops in experiments with the respective Embrapa research centers is also being coordinated.

In April, the equipment was brought to Brazil to the headquarters of Macnica DHW, a Japanese multinational, located in Florianópolis (SC). There, the structure was set up for the experiment to capture the brain signals of phytopathologists Cláudia Godoy and Rafael Soares (photo) from Embrapa Soya. Both evaluated around 1.500 images of diseased and healthy leaves for testing with the collecting helmet.

The proof of concept stage showed that the models generated from the experts' electroencephalograms are capable of handling images well, allowing the machine to be trained to identify diseased plants. “The combination of labeled images – sick/healthy – with the experts’ brain signals resulted in an improvement in the model’s performance, indicating the feasibility of using AI”, points out Barbedo.

First results

“The experience was very interesting, because the system learns to identify images of diseased leaves based on the count that is done silently when viewing diseased and healthy leaves that pass quickly on a computer screen by identifying brain signals”, highlights Claudia Godoy (photo).

“Based on the evolution of this training, these recognition technologies can be used by people who do not have much knowledge of diseases, helping to manage them”, he reinforces.

According to Soares, two diseases were chosen for this experiment: Asian rust - the most economically important disease in the crop - and powdery mildew - relevant in the southern region of Brazil. “These diseases were chosen because, in addition to their importance for soybean cultivation, they cause two distinct types of foliar symptoms on the plant and there was adequate availability of images for evaluation”, explains Soares.

For the researcher, improving soybean disease management tools is important because “detecting and diagnosing diseases is one of the greatest difficulties encountered in the crop, and innovative technologies that add information to these practices are desirable and necessary”, he highlights.

System operation

The system “imitates” the brain functioning of experts when viewing images of diseased plants, automating labeling and making the step faster and more efficient. The idea is to simulate, as closely as possible, the brain process of a specialist when identifying something or making a decision, as was done with phytopathologists.

The first step is to calibrate the model, adjusting the helmet with the electrodes on the specialist's head to identify their brain signals. “Each person has a different brain pattern, the brain's electrical signals are different from person to person, so it is necessary to calibrate each person so that the model can understand what they are thinking”, explains Barbedo.

Once the system has ‘learned’ how the person works, the database labeling process begins. The instructions for specialists are to list (1, 2, 3...) the diseased leaves when they see them on the screen, which displays three images per second. The system captures the brain signals emitted with each new stimulus, different from when viewing a healthy leaf.

According to the project leader, the counting process is not mandatory, but it reinforces brain signals, making it easier to differentiate between what is sick and what is healthy. The system allows the presentation of up to ten images per second.

Reliable results

Lasting an average of half an hour, each session made it possible to label more than a thousand images, a task that would take days in the manual system. In addition to the gain in agility in the labeling process, Barbedo highlights the reliability of the system, “which has mechanisms for correcting possible errors, making the model that is trained more reliable”.

The system can identify whether the specialist blinked or was losing attention in the process of viewing images in sequence through neural signals. In these cases, the system discards the result and redisplays the image later. The BrainTech system generates an indicative curve of attention, pausing the experiment to rest when it drops to a level critical to the reliability of the results.

Furthermore, the system is capable of detecting the expert's level of certainty when viewing the image, which is called a softlabel. The use of this parameter allows for better calibration of the AI ​​model according to the level of experience of each specialist, which consequently brings greater accuracy in the decision of the AI ​​model.

Applications in agriculture

The technology opens up several possibilities for application in the agricultural sector. The trained models could be embedded in agricultural machinery, cell phone applications and working in activities with a lack of specialized labor.

A more rational application of pesticides, with less economic cost and less environmental impact, and food production in a cleaner and more sustainable way would be possible with trained models embedded in machinery, identifying in real time and in specific plots the need to apply pesticides when passing on production lines.

“Embedding this model in a cell phone application would give the producer agility in decision-making when diseases and symptoms of pathologies are identified, accelerating the adoption of the necessary measures”, indicates Barbedo.

The researcher also points out the relevance of using technology in the pasture rotation strategy for dairy farming, an area in which there is a lack of specialists. The choice of the most appropriate paddocks to maximize milk production is made by a technician experienced in identifying the best location and ideal number of animals. “The system could simulate the activity of this specialist to create a technological location. Most properties do not have someone with this expertise”, he concludes.

Watch the video with researcher Jayme Barbedo, from Embrapa Digital Agriculture.