RDA’s ‘AI Isak 2.0’ diagnoses pest and disease photos with 92% accuracy…ahead of GPT in its own comparison

If you photograph a diseased leaf and ask a general-purpose generative AI what disease it is, quite a few wrong answers come back. In the Rural Development Administration’s own comparison, in which pest and disease photos were shown and the systems were asked to identify the disease name, the accuracy of GPT-5.6 was 46.5%, wrong about once every two tries, and GPT-6 Astra reached only 77%. In the same test, the RDA’s artificial intelligence (AI) farming consultation service ‘AI Isak’ got 92% correct.

On the 30th of last month, the RDA released ‘AI Isak 2.0’ and changed the nature of the service from a ‘farming consultation chatbot’ to a ‘farming secretary AI agent’ that performs diagnosis and prescriptions. Shin Jae-hoon, head of the RDA Agricultural Intelligence Data Team, said, “AI like ChatGPT can also search public data and answer well, but it cannot perform the pest and disease diagnoses, fertilizer prescriptions, or weather forecasts that RDA experts do,” adding, “The biggest feature of Isak 2.0 is that it lets Isak call up and answer with these functions that existed as separate systems.”

Farmers’ response is showing up in numbers. From January to September this year, the number of questions submitted to Isak was 104,702, more than five times higher than in December last year. Except for February (6,174), it exceeded 10,000 each month, and in September, ahead of the launch of 2.0, it was the highest at 14,163. Cumulative downloads stood at 20,096 as of the 10th of this month, and 6,839 users have updated to the new version since the release of 2.0.

 

17 information systems, through one chat window

The RDA has separately operated services such as soil information, early warnings for agricultural weather disasters, and pesticide registration information by system. Isak 2.0 is a project to connect them so they can be called up and used within a conversation with Isak. Pest and disease diagnosis and fertilizer prescriptions were connected first in September, and early warnings for agricultural weather disasters, pest and disease forecasting information, legal information, soil characteristics and test information, and fruit-tree growth quality information will be connected in stages by year-end, expanding to 36 services across 17 systems.

Team leader Shin explained, “To conduct a soil diagnosis, farmers have to enter Heuktoram (the RDA soil environment information system) and select various items, and to receive early warnings, they have to install and access a separate service, which is difficult for farmers,” adding, “We gathered these things so they can be solved with Isak alone.”

Fertilizer prescriptions recommend appropriate fertilizer amounts for 253 crops based on field-by-field soil test results. If there are no test results, it first provides standard usage amounts by crop, and guides farmers to receive an accurate prescription after soil testing at an agricultural technology center. 

300,000 pest and disease photos, each identified by experts

Pest and disease diagnosis is a function that provides the disease name and control information when a crop photo is taken and uploaded. It currently identifies 190 types of pests and diseases (94 diseases and 96 pests) in 25 crops. These 25 crops account for 74.1% of the domestic cultivated area. The RDA plans to expand the target by 2029 to 60 crops, covering 93.0% by cultivated area.
What determined accuracy was the training data. Unlike general-purpose AI, which broadly learns from images and text publicly available on the internet, Isak was trained on 300,000 photos for which RDA pest and disease experts identified the disease name one by one. Team leader Shin said, “Isak was trained after our pest and disease experts identified what disease each of all 300,000 photos showed.”

 

Consideration for elderly farmers with dialect voice input and large text

Features for elderly farmers were also added. It is a voice secretary function that lets users ask questions and hear answers by speaking without touching the screen. Voice recognition uses Naver Clova Speech, and the dialect recognition accuracy announced by the RDA is 99.2%. For example, if a farmer says in a Gyeongsang dialect, “The weeds are everywhere,” it understands it as “There are too many weeds.”

The text, which had been criticized as small in Isak 1.0, was enlarged, and a new farming diary function for recording and checking farming activities was added. Team leader Shin said, “Older farmers almost never type on mobile phones and mostly use dictation (voice input),” adding, “Existing information systems are very difficult, but the advantage is that if they speak, it understands and answers.”

 

Answer sources only from RDA-verified materials

The chronic problem of generative AI, ‘hallucinations (plausible wrong answers),’ was addressed by narrowing the grounds for answers. Isak finds documents related to questions from within 560,000 pages of agricultural technology materials published by the RDA, summarizes them, and answers, without using external materials. Under each answer, it displays the source document and page number, such as ‘Agricultural Technology Guide: Apple, page 335,’ so users can check which material the answer came from.

Team leader Shin said, “Because the structure searches registered documents, extracts related documents, and summarizes them to answer, in principle there is little room for hallucinations to intervene,” adding, “We minimized hallucinations by using already verified information such as Heuktoram or weather information.”

However, even if the basis is narrowed, it cannot eliminate errors such as selecting documents that do not match the question or content deviating during the summarization process. On the 1st of last month, the RDA formed a quality assurance team to continuously check answer quality. A total of 30 people, including 5 dedicated staff and 25 cooperating personnel, compare 300 answer sheets prepared by agricultural experts with Isak’s answers. It also analyzed farmer Q&As and additionally registered about 1,000 reference documents.

 

Gochang, North Jeolla maca farm “It speaks only based on available information…so I trust it”

Reactions in the field are also changing. Kim Il-soo, head of the AI Education Project Group at the Korea Federation of Informatized Farmers, who grows maca and Chinese artichoke in Gochang, North Jeolla Province, has used Isak since the release of version 1.0.

He had also used general-purpose AI such as ChatGPT, Gemini, and Claude from early on, but said, “They answered broadly, but lacked expertise in specialized areas.” Isak 1.0 was also disappointing at first. Because there was little material on maca, other information sometimes came back instead of the answers he needed. Kim said, “Pest and disease information is meaningless if it is not about my crop, and at first it felt like receiving newsletters from the Ministry of Agriculture, Food and Rural Affairs or the RDA.”

His assessment is that things changed with Isak 2.0. When he asks about maca, a winter crop planted in September and harvested the following March to April, information on pests and diseases such as powdery mildew and information suited to the cultivation period appears. He emphasized, “Because Isak shows answers based on the materials it has, it is reliable,” adding, “Since it speaks only with the information it has, unlike other AI it has no hallucination symptoms, so I trust it.” He said it has improved enough to understand a considerable portion even when he uses a strong dialect.

He also pointed out areas for improvement. Kim said, “In the field, it is not a situation where you can search, but the voice-input microphone button is small, so it is inconvenient to press while wearing gloves,” adding, “It should be possible to ask and answer much more by voice.” He continued, “Right now it is at the stage of teaching information, but in the future it should develop into a one-on-one service that helps with each farmer’s business plan or even subsidy applications,” advising, “Farmers also need to take more interest and try using it so it can develop further.”

 

Training about 550 instructors…linking with private agricultural AI as well

The RDA plans to expand users through education. On the 26th, it will open a training course for extension instructors and train 433 field instructors at city and county agricultural technology centers. They will teach how to use Isak at Farmers’ Colleges (10,000 people), agricultural machinery education (30,000 people), and New Year practical education for farmers (250,000 people).

Farmer instructors will also be trained. The target is 120 members of the Korea Federation of Informatized Farmers, with which an MOU was signed in August. The first course was completed on the 6th and 7th of this month, and the second course will open in January next year. By year-end, it will also connect with public and private services such as Daedong’s ‘AI Daedongi,’ NongHyup’s ‘NH Today Farming,’ the Korea Agency of Education, Promotion and Information Service in Food, Agriculture, Forestry and Fisheries’ ‘Agriculture ON,’ and FarmHannong’s ‘Exaone Farm.’

In November, a legal information search function will be added, and next year version 3.0, including advanced management diagnosis and web search functions, will be released. AI Isak was selected last December as one of the Ministry of the Interior and Safety’s ‘30 Key Tasks for Realizing an AI Democratic Government.’ Team leader Shin promised, “We will continue updating it until users are satisfied.”