Deep learning — a technology that teaches computers to learn by example — allows cameras to recognize specific faces and driverless cars to distinguish a pedestrian from a lamppost.
Now, thanks to the work of crop physiologists, this kind of artificial intelligence is advancing into the area of agriculture, helping farmers easily spot disease in one of the world's most popular fruits.
Michael Selvaraj, a lead researcher from the Palmira, Colombia-based International Center for Tropical Agriculture (CIAT), developed a smartphone app with colleagues from the Bioversity International in Africa.
The technology, known as "Tumaini," uses machine learning to detect healthy — or unhealthy — bananas. With the app's 90-percent detection rate, Selvaraj hopes that Tumaini will save farmers millions of dollars in losses.
“This is not just an app,” said Selvaraj. “But a tool that contributes to an early warning system that supports farmers directly, enabling better crop protection and development and decision making to address food security.”
Pests and diseases like Xanthomanas wilt, Fusarium wilt, or Black sigatoka threaten to damage the healthy growth of bananas.
The Fusarium Tropical Race 4 fungus has led to losses of $121 million in Indonesia, $253.3 million in Taiwan, and $14.1 million in Malaysia (Aquino, Bandoles and Lim, 2013). In northern Mozambique, where the fungus was first reported in 2013, the number of symptomatic plants rose to more than 570,000 in September 2015.
With the "Tumaini" app, a word that means "hope" in Swahili, a farmer can find the symptomatic plants by taking photos with the phone. The app provides confirmation of disease, along with recommendations for next steps and control measures.
Existing detection models focus primarily on leaf symptoms and can only accurately function when pictures contain detached leaves on a plain background. Selvaraj's system, however, finds symptoms on any part of the crop, and is trained to read lower-quality images, inclusive of background noise, like other plants or leaves.
To build Tumaini, researchers uploaded 20,000 images depicting various visible banana disease and pest symptoms: plant wilt, leaf discoloration, or an ooze, for example. With this information, the app scans photos of parts of the fruit, bunch, or plant to determine the nature of the infection.
The tool is designed to help smallholder banana growers quickly detect a disease or pest. Tumaini's creators aim to link farmers to agricultural advisors who can quickly stem the outbreak.
"We are planning a chatbot in the future so that this platform will link farmers directly to the government extension people," Selvaraj told Tech Briefs.
The app, currently in the test phase, can also upload data to a global system for large-scale monitoring.
Selvaraj spoke with Tech Briefs about the importance to farmers of such an accessible AI. Read his edited responses below.
Tech Briefs: Tumaini has a 90% detection rate. What explains the 10% missed detection? What is still challenging for your app to detect?
Michael Selvaraj: In artificial intelligence, accuracy is based on how much the machine is learning from your data sets. The more data you have, the more accurate the app will be. So, this 10% can be improved by new data sets and training on [plant] features. We are also inventing new ways to differentiate more closely related diseases.
Tech Briefs: You tested this in in Colombia, the Democratic Republic of the Congo, India, Benin, China, and Uganda. Can you walk us through one test?
Michael Selvaraj: Currently this app is in the testing stage. Once this app is released, we plan to do campaigning through our national partners to explain how to use it. If farmers are seeing some symptoms, they take a photograph and click the Scan button on the app. Probability of disease will be shown in real time. If probability is very high, they click a Recommendations button to see the control measures of particular pest and diseases. Farmers also can choose the plant part where they are seeing the symptoms. We have 6 options, including whole plant, cut fruits, fruit bunches, leaf, and corm roots.
Tech Briefs: How do you imagine this app being used exactly? Does a farmer go around regularly and take pictures, or does a farmer only use the app when a questionable crop is seen?
Michael Selvaraj: This app will also detect healthy plants. If farmers see symptoms, they can take pictures and confirm the diseases early; these photos are GPS-tagged and will come to our server, so we can confirm the diseases of a particular area, prevent outbreak, and monitor the status by satellites and drones.
Tech Briefs: How does your application demonstrate a more “accessible” AI?
Michael Selvaraj: Right now global smartphone penetration is increasing. The Internet has become very cheap. We developed the API, which is accessible through cheap Android phones. Also, the app is free.
Tech Briefs: Can you give us a sense of the pest problem that today's farmers have to deal with?
Michael Selvaraj: Bananas are affected by major fungal, bacterial, and viral diseases, causing huge economic problems. Rapid identification is a key to preventing outbreaks. Right now farmers are identifying the diseases using empirical knowledge.
Early identification is also often not possible because of a lack of communication. Our app can be a decision-support system, to help farmers to identify field problems. Also it will be very useful to the scientific community to track the diseases on a global scale.
Selvaraj and his team's finding were published this month in the journal Plant Methods.
What do you think? What other applications are possible with a more accessible AI? Share your comments and questions below.