Early detection of plant leaf diseases using stacking hybrid learning

Abstract

The early identification of pests and diseases in crops now presents a significant challenge. Different methods have been used to resolve this problem. Sticky traps and black light traps, used to identify diseases and for field monitoring, are examples of a manual procedure for analysing the diseases. A lot of time is required, and it is less effective to manually inspect larger crop fields manually. To serve requires a professional, so it is, therefore, costly. The use of sticky traps, where by bugs stick to the material upon contact, is one method of disease monitoring. A camera is used to take a picture of the sticky trap. From the picture using the average disease count, this image is then processed to ascertain the pet density for a specific time period. Such manual methods, as well as providing an effective outcome also pose a danger to the environment. This is because farmers spray pesticides in large quantities as a preventative measure. Various approaches have been used to identify diseases, including image processing and sophisticated algorithms. The most effective method of disease identification from crops is automatic detection using methods of image processing and classification algorithms for the diseases to be categorised based on different picture attributes. With a stacking stacking hybrid learning with scratch and transfer learning strategies, which is utilised in this work, a model that has already been trained is used to learn on images of diverse fruit plant leaves from the Plant Village dataset, spanning both safe samples and various illnesses. This reasearch paper used ensemble CNN and we achieved accuracy between 99.75% to 100%.

Plant phenotyping relevance

植物葉の画像から病害状態を自動分類する画像解析・機械学習手法が研究の中心であり、植物の病害表現型を直接推定しているため。

Code and data availability

The paper's plant-phenotyping input is the public New Plant Diseases Dataset (Kaggle), explicitly cited in the Data Availability statement as the data supporting the findings. No author analysis code or trained model checkpoints are disclosed.