Distinguishing rice varieties using plant image analysis by deep learning methods

Abstract

Among the rice varieties developed for different purposes, Basmati varieties are unique for their morphological characters and quality. The origin, evolution and development of Basmati varieties has thrown challenges in terms of varietal classification and correct identification. Besides the classical method used in DUS testing for variety identification, new method consisting of whole plant images using deep learning algorithms was studied to identify basmati rice varieties. Classification of varieties by images of whole plant at different growth stages using deep learning algorithms was carried out to find the best algorithm and the best stage for effective discrimination of varieties. The ripening stage (terminal panicles ripened) was identified as the most suitable stage for effective classification of the varieties among the four stages namely, booting stage, 50% flowering, milk stage and ripening stage. The testing accuracy of all algorithms ranged between 60 to 73%. The testing accuracy at the ripening stage was found to be 73% using VGG 16, a deep learning model. Pusa Basmati 1609 and Pusa Basmati 1637 were identified with 100% accuracy. High testing accuracy was observed in identifying some other varieties namely, Pusa Basmati 1121, Pusa Basmati 1401, Pusa Basmati 1609, Pusa Sugandh 3. There was a high chance of misclassification among the genetically close varieties. Genetically close varieties that could not be differentiated using leaf and panicle characteristics, could be classified up to 90% accuracy using plant images and VGG 16. From this study it is concluded that plant image analysis by deep learning methods can be a viable alternative approach for identification of rice varieties.

Plant phenotyping relevance

水稲の全草画像と深層学習による品種識別手法を開発・比較しており、植物画像から品種状態を抽出する方法が研究の中心である。

Code and data availability

The paper describes field photographs of 15 rice varieties and CNN classification (VGG16, InceptionV3, ResNet152V2, VGG19), but no supplied block contains any data availability statement, public dataset deposit, or author code/model release. The only URLs in the text are the journal DOI and cited references; no paper-­

No evidence-backed public reproduction asset is currently recorded.