Spectral analysis for the early detection of anthracnose in fruits of Sugar Mango (Mangifera indica)

Highlights

• Spectral detection of anthracnose in fruits of sugar mango.

• Construction of a low-cost lighting camera to get spectral signatures without noise.

• Methodology for early identification of anthracnose using wavelengths discriminants.

• The wavelengths identified by LDA had an F-score of 91% on average in the classification.

Abstract

The use of spectroscopy in fruits provides spectral information that can be used to estimate chemical-physiological variables or to determine the phytopathological state of the fruit. Mango is a fruit prone to develop the anthracnose pathogen during its harvest, affecting its commercialization. There are different studies of mango that evaluate the development of anthracnose, however, no work in the previous literature has presented a method to estimate early the state of development of anthracnose. In this work, a spectroradiometer was used to evaluate the evolution of anthracnose in mango fruits. Three stages of development in the mango were analyzed (healthy, asymptomatic and diseased) and the performance was evaluated with random forest (RF) and support vector machines (SVM). The principal component analysis (PCA) and linear discriminant analysis (LDA) were used to reduce the dimensionality and identify the most significant bands of the spectrum used, with the help of a Gaussian filter. A total of 61 significant bands with PCA and 29 significant bands with LDA were found. The best evaluation performance was obtained with LDA reaching an accuracy of 91–100% in the three classes. The bands 399, 514, 726, 822, 912 and 1061 nm of the set of 29 bands of LDA are highlighted to identify asymptomatic fruits. This non-destructive method to identify the development of anthracnose at an early stage could benefit the farmer by helping to improve the commercialization of mango. In general, early detection of anthracnose, which is not visible, reached an average accuracy in the 29 bands identified with 91% LDA.

Introduction

Anthracnose is a disease caused by the presence of the fungus Colletotrichum sp., which affects several types of plants, mainly damaging tissues, stems, leaves, flowers and fruits (Gañán et al., 2015). This disease generates a phytosanitary problem that impairs the yield of various crops and causes economic losses in tropical and subtropical regions of the world (Huerta-Palacios et al., 2009). Mango crops are one of the most affected by anthracnose, which harms the pre-harvest and post-harvest stages of the fruit (de los Santos-Villalobos et al., 2011). This disease is characterized by colonizing initially the shell of the fruit, until it causes necrosis or death of the plant tissue inside the fruit in its final stage of development (Hu et al., 2014).

Mango (Mangifera indica) is characterized by its particular sweet taste due to the nutrients it provides to the consumer and it is one of the most commercialized fruits worldwide (Hu et al., 2014, Kamle et al., 2013). However, in its immature stage, it is prone to be infected by anthracnose without developing symptoms that appear in its maturity stage and affect its commercialization and market price (Hu et al., 2014). To combat anthracnose, non-destructive techniques have been developed to identify the progress of the disease in some types of fruits or plants and control the spread of different pathogens in crops (Yeh et al., 2016, Li et al., 2011, Sankaran et al., 2010, Abdulridha et al., 2019).

There are several remote sensing techniques that identify anomalies in the structure of fruits and leaves, using tools such as spectroscopy and thermography sensors, RGB and multispectral and hyperspectral cameras. These tools allow developing new techniques to improve productivity in the field, studying the reflectance or transmittance on the surface of fruits and leaves (Sankaran et al., 2010, Xie et al., 2017, Shuaibu et al., 2018). In Sharif et al. (2018) study 6 diseases in citrus fruits, detecting them with a digital camera and implementing segmentation techniques and feature selection with Principal Component Analysis (PCA). Although they achieve a classification accuracy close to 96%, the work focuses on classifying the disease when it is visible in the fruit. Another study by Abdulridha et al. (2019) uses a digital camera together with a multispectral camera to detect the wilting disease of the laurel in avocado trees. They identify the asymptomatic development of the disease, by means of two classifiers: K-nearest neighbors and multilayer perceptron neural network, with a yield of 79% and 85%, respectively. However, they do not specify the days on which they detect the disease in an asymptomatic state. Additionally (Abdulridha et al., 2018) analyze the same variables and use the same classifiers to identify the same disease by means of spectral bands, identifying the asymptomatic stage in the trees 14 days before it develops. However, the evaluation process of the asymptomatic state is not detailed, because the authors focus on assessing the classification performance of the healthy and diseased state.

Other studies related to spectroscopy focus on looking for discriminating bands that identify different states of some disease in fruits and leaves, with the purpose of providing solutions. The reduction of dimensionality of spectral bands in the area of spectroscopy allows identifying the most discriminating bands that explain the behavior or development of certain diseases in fruits and leaves. Sinha et al. (2019) determine a group of spectral bands using multilinear stepwise regression techniques and partial least squares regression, managing to identify the grapevine leafroll disease (GLD) in grape leaves, differentiating between the states of healthy and diseased leaf. The authors deduce that the bands identified can be used to determine the asymptomatic state on the leaf, but without performing an evaluation of the bands in the asymptomatic state. Similarly, (Mahlein et al., 2013) find a group of discriminating spectral bands when analyzing hyperspectral images obtained from sugar beet leaves using the RELIEF-F algorithm. This made it possible to identify the healthy state and different diseases of the leaf (Xie et al., 2017) use multispectral images to classify healthy and diseased tomato leaves with a gray mold into five time-intervals, reaching a classification accuracy of about 62% with the technique of K-nearest neighbors. Although the studies described above identify different states of certain diseases such as GLD or gray mold, they do not focus on identifying an asymptomatic state of the disease or do not adequately describe the processes for determining this state. However, Yeh et al. (2016) is studying the development of strawberry leaf anthracnose by analyzing hyperspectral images and determining a group of spectral bands that serve to differentiate between a healthy and an asymptomatic leaf, obtaining classification results of 88% through a progressive discriminant analysis.

Another development in the area of spectroscopy is the analysis of mechanical damage and defects in leaves and fruits. Hu et al. (2016) detected mechanical damages that are present in the pulp of blueberries and not visible in the shell. They obtained a number of spectral bands that allow to differentiate blueberries with mechanical damages from healthy ones, implementing the algorithm Competitive Adapted from Repeat Sampling (CARS). The classification is not time continuous, since performance is evaluated at different time intervals, varying the accuracy of the implemented classifiers. In Li et al. (2011), the PCA technique is used to determine the most significant bands in oranges and differentiate healthy fruits from defective ones affected by insect damage, wind scars, thrips scars, scale infestation, canker spot, copper burn, heterochromatic fringe, phytotoxicity and stem end. The mentioned works can be applied to different sectors of the industry, improving the quality of the product that reaches the consumption (Hu et al., 2016, Li et al., 2011). However, these studies only focus on damage caused by external agents and not on phytopathological conditions of the fruit.

More specific studies on the handling focus on identifying damage, defects and maturity states, using different sensors. The proposal by Nagle et al. (2012) implements a computer vision system (CVS) with UV-A illumination, for the acquisition of images of mango Nam Dokmai and Maha Chanok fruits, which are inoculated with anthracnose, and other bruised fruits, evaluating their affected areas. The study indicates that the disease can be detected early with UV-A illumination. Similarly, Patel et al. (2019) implement a CVS, but with a UV camera and UV-A illumination. They contrast the detection of defects on mango fruits with images in the RGB space and highlight the potential of working with wavelengths of 360–400 nm. Rivera et al. (2014) use a hyperspectral camera in Manila mangoes for the timely identification of the mechanical damage by segmenting the healthy and damaged areas of the fruit with the technique of K-nearest neighbors. They contrast the classification with all the bands (650–1100 nm) and with the most discriminating of this group, obtaining results of 97.7% and 91.4%, respectively.

Other studies focus on establishing technological strategies to define the state of ripeness in mango fruits, with the aim of reducing harvest times. In Naik and Patel (2017) develop a classification method to estimate the state of maturity of the Langdo mango, from thermal images. However, the process has several limitations to reach an accuracy of about 89% with the proposed algorithm, such as the storage of the mango at the same temperature or using a thermal reference mango, among others. Wendel et al. (2018) and Gutiérrez et al. (2019) use terrestrial robots with hyperspectral cameras attached to estimate the degree of ripening of mango fruits in the field. The first work uses the full spectrum offered by the camera (411.3–867 nm), to assess the ripeness of the fruits with Convolutional Neural Networks (CNN). In the second work, they extract the most discriminating bands from the 400 to 1100 nm range using parametric filters by evaluating them with Support Vector Machines (SVM). Although these studies focus on analyzing the state of maturity of the fruit to improve the harvest process, they solve none of the phytosanitary problems that arise in fruits, such as anthracnose. This paper presents a methodology focused on the early detection of anthracnose in fruits of sugar mango. However, the methodology implemented can be used for the analysis of any vegetative (leaves, stem) or reproductive (flowers and fruits) structure of interest, under controlled laboratory conditions.

Identifying the anthracnose disease in mango fruits before it develops notoriously would allow to apply fungicides in a localized manner and eliminate the spread of the disease. For this reason, in this study sugar mangoes were inoculated with the fungus Colletotrichum sp., in order to analyze the spectral signatures obtained with a spectroradiometer under controlled lighting conditions. The progress of the disease in the fruits was evaluated by analyzing the bands in the range of 350–1900 nm. The most discriminating spectral bands in the different stages of anthracnose development are determined by dimensionality reduction techniques, such as PCA and LDA. Finally, different classifiers were implemented, evaluating their performance in the classification into the three fruit states: healthy, asymptomatic and diseased.

This document has the following structure. Section 2 describes the process for capturing spectral signatures, the sample preparation, the definition of classes and the manipulation of those spectral signatures. Section 3 describes the manual evaluation of the classes. Section 4 describes the methods of band dimensionality reduction, component filtering and transformation vectors. Section 5 evaluates the performance of the bands identified. Finally, the conclusions and future work are presented in Section 6.