Deep learning for the differentiation of downy mildew and spider mite in grapevine under field conditions

Highlights

• Spider mite and downy mildew can be differentiated automatically.

• Computer vision and deep learning allow for fast modelling of disease in grapevine.

• The results show that accurate models can be trained without the need of large numbers of images.

Abstract

Diseases and pests cause serious damage in crop production, reducing yield and fruit quality. Their identification is often time-consuming and requires trained personnel. New sensing technologies and artificial intelligence could be used for automatic identification of disease and pest symptoms on grapevine in precision viticulture. The aim of this work was to apply deep learning modelling and computer vision for the detection and differentiation of downy mildew and spider mite symptoms in grapevine leaves under field conditions. RGB images of grapevine canopy leaves with downy mildew symptoms, with spider mite symptoms and without symptoms were taken under field conditions in a commercial vineyard. The images were prepared using computer vision techniques to increase disease visual features. Finally, deep learning was used to train a model capable of differentiating leaf images of the three classes. An accuracy up to 0.94 (F1-score of 0.94) was obtained by classifying leaves with downy mildew, spider mite and without symptoms at the same time, using a hold-out validation. Additionally, accuracies between 0.89 and 0.91 (F1-scores between 0.89 and 0.91) were obtained in the binary classification of the disease and pest, obtaining the best results in differentiating downy mildew from spider mite symptoms. This high accuracy demonstrates the effectiveness of deep learning and computer vision techniques for the classification of grapevine leaf images taken under field conditions, automatically finding complex features capable of differentiating leaves with spider mite symptoms, with downy mildew symptoms and without any. These results prove the potential of these non-invasive techniques in the detection and differentiation of pests and diseases in commercial crop production.

Introduction

Diseases and pests are one of the major problems in crop production. Grapevine can be attacked worldwide by numerous pathogens and pests, causing diseases that induce serious damage (Galet, 1996, Wilcox et al., 2015) reducing yield and grape quality. Traditionally, pests and diseases detection in vineyards has been performed by visual evaluation of the symptoms in leaves or/and fruits. This evaluation can be confirmed by serological (Maree et al., 2013), microscopical or molecular biology techniques (Gindro et al., 2014). Early detection has also been developed by air fungal spore trapping by volumetric or passive traps, and posterior identification and quantification of these structures (Carisse et al., 2009, Martínez-Bracero et al., 2019). Indirect methodologies have been used too, to detect molecules or particles produced by the pathogen of interest (Andolfi et al., 2009). Pest monitoring can be also carried out by visual searches of insects causing damage, or by installing capture traps. These traps can be from different nature: attraction of the insects with specific pheromones (Alcudia-León et al., 2017), attraction to colored plates or to different broths (Lasa et al., 2020), or vacuuming (Shapira et al., 2018). All cases suppose the subsequent identification of the insect by visual assessment of an expert or by using molecular techniques (Ali et al., 2019, Giblot-Ducray et al., 2016). However, the described techniques to identify crop pests and diseases are time-consuming, require trained personnel and/or are expensive, as molecular techniques (Ali et al., 2019, Sankaran et al., 2010). They always require taking the samples in the field and process samples in the lab to obtain a conclusive result.

The pest and disease symptoms should be identified accurately in-field commercial vineyards in order to settle the appropriate decision to manage the vine crop protection. However, the symptoms can be quite similar and very difficult to distinguish and identify correctly. Regarding to common diseases in grapevine, initial stages of the infection process of downy mildew caused by Plasmopara viticola shows yellowish oil spots in grapevine leaves (Unger et al., 2007, Wilcox et al., 2015). These symptoms can be confused with caused by pests such as some spider mites—Eotetranychus carpini Oud and Tetranychus urticae Koch (Wilcox et al., 2015). The presence of spider mites in grapevine induced yellow or reddish symptoms, depending on the grapevine variety, spots in leaves, mainly along the main veins. The differentiation between all the yellowings is highly important to identify accurately the causal agent. In many times both symptoms, those caused by downy mildew and by spider mites appeared in the same vineyard and the same plant. To manage the pest (spider mite) or disease (downy mildew) in commercial vineyards it is essential to get a precise diagnosis, as it helps to the correct identification and quantification of the causal agent.

New sensing technologies and artificial intelligence have been recently applied in phytopathology and crop protection. Emerging technologies can be used for detecting plant diseases in agriculture with several advantages versus conventional methods (Ali et al., 2019, Sankaran et al., 2010). Sensing technologies have been applied for both pests and diseases identification in grapevine, like the combination of image texture and spectral signals (Al-Saddik et al., 2018) or the analysis of thermal information (Mastrodimos et al., 2019). Additionally, authors have taken advantage of the massive data collection that sensing technologies have to offer to apply several machine learning techniques to them for disease detection, like neural networks (Zhu et al., 2020) or statistical methods applied to data from IoT devices (Patil and Thorat, 2016). Finally, applications of lastest trends of deep learning have been reported in several studies: assessing spider mites’ damage level in cotton (Yang et al., 2019), detection of nutrition deficiencies in plants (Barbedo, 2019a), identification of yellow symptoms in grapevine (Cruz et al., 2019) or differentiation of diseases in commercial crop production (Mahlein et al., 2019). These solutions must be robust and reliable systems capable of discerning the symptoms produced by different pathogens that often appear simultaneously in the same plot.

The goal of this work was to examine the potential of deep learning for the detection and differentiation of key grapevine disease (downy mildew) and pest (spider mite) using in-field RGB images in a commercial vineyard.