Image-based field monitoring of Cercospora leaf spot in sugar beet by robust template matching and pattern recognition

Highlights

• An image-based foliar disease monitoring on a single leaf scale is proposed.

• OCM based robust foliage tracking against various changes in open field.

• Novel feature of L^∗, a^∗, Entropy × Density for classifying CLS from soil background.

• Promising for automatic CLS progress observation in real field.

Abstract

This paper presents a novel image algorithm using template matching and pattern recognition frameworks for monitoring Cercospora leaf spot (CLS) development on sugar beets on a single leaf scale under real field conditions. Due to the variety and complexity of the open field, it is a great challenge to achieve continuous and robust foliar disease observation in real field conditions. We propose a novel and compact algorithm, composed of two frameworks and a post-processing. The algorithm has continuous and highly discriminative capabilities for observing the process of disease in a single leaf from plant-level time sequence images. The first framework is based on robust template matching by orientation code matching (OCM), which implements successive tracking of a single leaf from a beet plant against severe illumination changes and non-rigid plant movements. The second framework uses a pattern recognition method of support vector machine (SVM) for achieving further disease classification from clutter field background. Prior to SVM, we propose a three feature combination of L^∗, a^∗, Entropy × Density, which has strong discrimination power to classify CLS disease from the clutter scene containing sandy soil, leaves, leaf stalks, and specular reflection. Additionally, post-processing is introduced to filter false positive noise to enhance the precision of the classification. Field experiment results demonstrate the feasibility and applicability of the proposed algorithm for disease monitoring under real field conditions. Meanwhile, comparative results with other conventional matching methods and feature combinations show the effectiveness of our proposed algorithm in both foliage tracking and disease classification.

Introduction

Sugar beets (cv. Beta vulgaris), which account for 20% of the world’s sugar production (FAO, 2009), are large-rooted plants with high content of sucrose for sugar extraction. However, they are vulnerable to many pathogens which negatively affect sugar production. Among those, Cercospora leaf spot (CLS) is one of the most destructive foliar diseases in sugar beets with high incidences affecting over one-third of sugar beet cultivation area worldwide (Holtschulte et al., 2000) and approaching 40% sugar losses (Robert and Extension plant pathologist, 2013). Therefore, fungicide spraying is necessary and the most important tool for disease control and plant protection (Ioannidis and Karaoglanidis, 2010). Usually, profitable and optimal decisions for fungicide spraying are made in terms of disease severity and local weather conditions (Jones and Windels, 1991), by which both economic losses and chemical usage can be reduced. Therefore, disease severity assessment is a crucial premise to facilitate optimal disease management.

The traditional way of CLS damage assessment is manual field monitoring, which is a process in which specialists first need to walk into the fields at certain intervals and assess the disease severity by naked eye observation based on a certain rating system, such as a single leaf scale based (Jones and Windels, 1991) or whole plant scale based (Vereijssen et al., 2003). However, this hand-operated field monitoring has limitations of being laborious, discontinuous, somewhat subjective with large-scale fields, and imprecise with subtle disease variations.

In recent years, image sensing techniques have captured many researches’ attention and been broadly introduced into plant disease study for their benefits in automatic, low-cost, noninvasive disease analysis capabilities. And a number of prior researches have demonstrated their inspiring ideas and results in foliar disease detection and measurements using both color images (visible spectrum) and spectral imaging. Sankaran et al. (2010) provides a comprehensive overview for currently used technologies that can assist in monitoring health and diseases in plants under field conditions, in which both visible and spectroscopic imaging techniques can be accessible. Polder et al. (2014) designs a mobile platform consisting of two RGB–NIR multispectral cameras, a color camera, a black and white camera and two infrared LED floodlights to automatically detect diseased tulip plants infected by tulip breaking virus (TBV) in open fields. Diseased plants were detected using Fisher’s linear discriminant classification with four features derived from four bands (R, G, B and NIR) of the multispectral images. Camargo and Smith (2009) develop a colored image based method that can identify the visual symptoms from plant disease regions with a large range of intensity distributions by analyzing the gradient change between the neighborhoods of each pixel. Huang (2007) employs an artificial neural network along with color and texture features for detecting and classifying three foliar diseases in Phalaenopsis seedlings based on color images. Mahlein et al., 2012, Rumpf et al., 2010 introduce hyperspectral imaging to detect and differentiate foliar diseases of sugar beets in the pixel level by analyzing the reflectance properties of the plant and diseases. Oberti et al. (2014) investigates the sensitivity of disease detection from different view angles of the canopy by multispectral image analysis and suggests a range of view angles from 40° to 60° to improve early detection of disease symptoms. However, most of the researches in this field have previously focused on detecting diseases under controlled environments, static images with zoom-in or small patches of leaves, and without considering soil background, which limit their further applications for long-term, consistent, and comprehensive disease observation under real field conditions.

Alternatively, this study focuses on monitoring CLS development from plant-level RGB image sequences under real field conditions. In our previous work (Zhou et al., 2014, Zhou et al., 2013a, Zhou et al., 2013b), we proposed a robust template matching based image algorithm, which achieved continuous small-scale CLS observation under various illumination changes. However, the challenging issue of sandy soil in real field was not addressed. In this study, we aim to solve the soil problem and also extend the small-scale based CLS observation into single leaf scale based, which, in addition to achieving motivations for complying with one of the manual CLS assessment criteria by specialists, such as a rating system used in Minnesota and North Dakota (Jones and Windels, 1991, Windels et al., 1998), also avoids potentially inconsistent disease observations due to mutual overlap of cluster leafage. For achieving field monitoring for a single leaf based CLS development, the algorithm should meet two fundamental capabilities. First is continuity, which can align identical leaves through neighboring frames for continuous leaf tracking. Second is discriminability, which can further classify CLS against the field background. However, challenges in the complex and unstructured field conditions will disturb the robustness of image algorithms. Firstly, changeable illumination can lead the intensities of plant images captured under natural sunlight conditions to vary over time. Secondly, constant changes in movements with translation, slight rotation, and occlusion of non-rigid plant, caused by both internal circadian growth and external environment interactions, are difficult for identical leaf alignment. Thirdly, cluttered field scenes with the key problem of sandy soil resembling the reddish brown color of CLS will increase the difficulties for CLS classification and detection. Meanwhile, highlighted leaf stalks and specular reflection which share the light-brown color with the inner of spots are also parts of the problem.

In this study, we propose an algorithm that consists mainly of a two-stage framework and post-processing to realize the research goal of the continuity and discriminability for a single leaf based CLS monitoring as well as attack the challenges in real field conditions. The first stage employs a robust template matching called orientation code matching (OCM) (Ullah et al., 2001) for identical leaf alignment from adjacent time-series frames against various illuminations and plant movements. Then, an extended adaptive OCM is used to achieve the continuity for the leaf tracking. The second stage uses a pattern recognition method of support vector machine (SVM) for further classifying CLS against the field background. We propose a novel three feature combination of L^∗, a^∗, Entropy × Density as inputs for SVM, which has strong CLS discrimination power from the clutter field background with sandy soil, leaves, leaf stalks and specular reflection. Finally, an edge detection based post-processing is conducted to filter false positive noise aroused from the proposed feature to improve the precision of classification results. A schematic diagram of the proposed algorithm is shown in Fig. 1. In summary, this study provides a visible image sensing technique for automatic and effective CLS progress monitoring in a single leaf manner under challenging real field conditions. Additionally, to the best of our knowledge, this is the first study to achieve continuous foliar disease monitoring on a single leaf scale from plant-level image sequences.