Mango Fruit Diseases Severity Estimation based on Image Segmentation and Deep Learning

Abstract

Abstract Plant disease severity is the ratio between the surface area of disease symptoms and the total surface area of the plant unit (e.g. fruit, leaf). It is related to plant disease diagnosis and has several advantages for farmers. It is therefore a key element in the protection and management of plant diseases. In the literature, there are three proposed categories of plant disease severity determination solutions: those based on segmentation algorithms, those based on classical ML algorithms and those based on DL algorgorithms. Despite their many advantages, these solutions have a number of limitations, including i) subjectivity in data labeling, ii) loss of information on disease lesion contours during (manual) data labeling, and iii) the proposed solutions have focused on estimating plant disease severity from leaves, although diseases can also affect other parts of the plant, such as fruits. In this paper, we present a solution for estimating the severity of four mango fruit diseases, namely alternaria, anthracnose, aspergillus rot and stem rot. This solution is based on ResNet50 CNN and uses a dataset automatically labeled by a proposed algorithm based on two segmentation algorithms such as image color space segmentation and image thresholding. The solution has achieved an accuracy and a F1_score of 97.82% and 97.79%, respectively, on test data. It is then deployed in a mobile application with a diagnostic solution we previously proposed. This mobile application will help mango growers, particularly those in Sahelian countries like Senegal, to manage their mango diseases earlier.

Plant phenotyping relevance

マンゴー果実の病斑面積に基づく病害重症度を、画像セグメンテーションと深層学習で推定する方法が中心であり、植物状態の定量的フェノタイピングに該当する。

Code and data availability

The paper uses the authors' own public SenMangoFruitDDS dataset of 862 mango fruit images, explicitly stated to be downloadable from Mendeley Data, as the image input for their severity estimation and automatic labeling pipeline. No code or trained model deposit is mentioned.