Development of robust plant diagnosis system with deep learning and dimensional reduction

2019 Fiscal Year Final Research Report

• Project/Area Number: 17K08033
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Research Field: Agricultural environmental engineering/Agricultural information engineering
• Research Institution: Hosei University
• Principal Investigator: Iyatomi Hitoshi 法政大学, 理工学部, 教授 (10386336)
• Project Period (FY): 2017-04-01 – 2020-03-31
• Keywords: 植物病害 / 自動診断 / 深層学習 / 機械学習 / 早期発見

Outline of Final Research Achievements

We have developed automated image-based diagnosis and various key technologies for the early detection of plant diseases. We developed disease discriminators using more than 200,000 images of four crops (cucumber, tomato, eggplant, and strawberry) taken at 24 prefectural agricultural research centers. Most of the obtained discriminators have achieved more than 90% disease discrimination by the cross-validation method, and we have made them available on the web to our collaborators for use in the field.
We have also achieved a wide range of results, including various methods to reduce over-fitting in real-world environments (extraction of regions of interest such as leaves and stems, various new training methods, and methods for generating training data), diagnostic techniques that allow simultaneous diagnosis from wide-angle images, and super-resolution techniques that reduce the loss of accuracy in such situations.

Free Research Field

機械学習

Academic Significance and Societal Importance of the Research Achievements

植物病害は、農業生産の量と質に大きな損失となる。早期発見が重要であるが、診断は専門家の目視に頼っており時間的、金銭的なコストが課題となっている。スマートフォンや遠隔カメラなどによる、高速かつ安価で正確な自動診断技術が実現し、普及できれば、世界中において農業経済的な損失を減らすのみならず、今後の世界的な人口増大に伴う食糧事情の改善にも大きく貢献できる。我々は、本研究により得られた様々な成果を元に構築した植物病害識別器を、画像を提供してくださった24府県に実検証用のためwebで公開した。今後も逐次更新を予定しており、研究としての成果のみならず実践的な成果として今後の活用が期待される。