Detecting powdery mildew of winter wheat using leaf level hyperspectral measurements

https://doi.org/10.1016/j.compag.2012.03.006Get rights and content

Abstract

Powdery mildew (Blumeria graminis) is one of the most destructive diseases, which has a significant impact on the production of winter wheat. Detecting powdery mildew via spectral measurement and analysis is a possible alternative to traditional methods in obtaining the spatial distribution information of the disease. In this study, hyperspectral reflectances of normal and powdery mildew infected leaves were measured with a spectroradiometer in a laboratory. A total of 32 spectral features (SFs) were extracted from the lab spectra and examined through a correlation analysis and an independent t-test associated with the disease severity. Two regression models: multivariate linear regression (MLR) and partial least square regression (PLSR) were developed for estimating the disease severity of powdery mildew. In addition, the fisher linear discriminant analysis (FLDA) was also adopted for discriminating the three healthy levels (normal, slightly-damaged and heavily-damaged) of powdery mildew with the extracted SFs. The experimental results indicated that (1) most SFs showed a clear response to powdery mildew; (2) for estimating the disease severity with SFs, the PLSR model outperformed the MLR model, with a relative root mean square error (RMSE) of 0.23 and a coefficient of determination (R2) of 0.80 when using seven components; (3) for discrimination analysis, a higher accuracy was produced for the heavily-damaged leaves by FLDA with both producer’s and user’s accuracies over 90%; (4) the selected broad-band SFs revealed a great potential in estimating the disease severity and discriminating severity levels. The results imply that multispectral remote sensing is a cost effective method in the detection and mapping of powdery mildew.

Highlights

► We detected powdery mildew at leaf level by hyperspectral measurements. ► PLSR outperforms MLR in estimating disease severity. ► FLDA can reliably discriminate the three healthy levels of winter wheat. ► Broad-band spectral features have great potential in disease detection.

Introduction

At least 10% of global food production is lost due to plant disease (Christou and Twyman, 2004, Strange and Scott, 2005). Powdery mildew, caused by Blumeria graminis, is one of the most widely destructive plant diseases in the world (Reuveni and Reuveni, 1998, Olsen et al., 2003, Nofal and Haggag, 2006). The disease affects a wide range of commercial crops, and can result in a significant yield loss (Sharma et al., 2004, Strange and Scott, 2005). Therefore people have paid a great attention to the impact of powdery mildew on food security (Hardwick et al., 1994). Several studies have addressed the influencing mechanism of powdery mildew from physiological or genetic perspectives, attempting to breed varieties with strong resistance to powdery mildew or to develop effective fungicides (Gooding et al., 1994, Wright et al., 1995, Hu et al., 2008). Meanwhile, great progresses were achieved in preventing and controlling powdery mildew in wheat planted areas. For example, Hardwick et al. (1994) found that a fungicide with fenpropidin and fenpropimorph appeared to be effective in controlling the powdery mildew. Jørgensen and Olesenb (2002) discovered that the infection of powdery mildew can be successfully prevented with fungicides containing ergosterol biosynthesis inhibitors. However, although the application of fungicides is effective in controlling the powdery mildew, it is impossible to eradicate the disease at a regional scale since many species of plants can host this pathogen (Eichmann and Hückelhoven, 2008). Therefore, it is important in practice for crop managers to obtain information about the spatial distribution of powdery mildew in time to guide the spray of fungicide. In addition, an inaccurate application of fungicide can lead to missing infected areas or overuse, especially when using automatic spray systems such as tractors or aircraft.

To obtain the information of disease infected boundaries in the field, the most common and conventional way is conducting a field survey. The traditionally ground-based survey method is very expensive and inefficient and, therefore, is problematic over large areas. However, remote sensing technology may be a possible alternative for obtaining the spatial distribution information of powdery mildew over a large area with a relatively low cost.

During the last two decades, several studies were successfully conducted to detect crop diseases by means of remote sensing techniques (e.g. West et al., 2003, Sankaran et al., 2010). As stated by Sankaran et al. (2010), optical remote sensing, particularly using spectral features (SFs) extracted from visible and near-infrared (NIR) regions, has great potential in plant disease diagnosis and detection. For example, by using multispectral data, Franke and Menz (2007) successfully detected powdery mildew and leaf rust in a winter wheat field by using normalized difference vegetation index (NDVI). Qin and Zhang (2005) obtained the infected area information of rice sheath blight with broadband high spatial-resolution data. In addition, some researchers have applied hyperspectral remote sensing technique to detection and mapping crop disease. Bravo et al. (2003) and Moshou et al. (2004) developed a ground-based real-time remote sensing system for disease detection in winter wheat field, which achieved a classification accuracy of over 90%. Huang et al. (2007) found that the Photochemical Reflectance Index (PRI) had a strong estimating power for yellow rust disease in winter wheat at canopy level. In their study, a relationship between PRI and disease severity of yellow rust in winter wheat was further confirmed with airborne hyperspectral data. Liu et al. (2010) also used hyperspectral reflectance measurements to make an accurate discrimination of rice fungal diseases at different severity levels. Based on the literature review, it is apparent that the hyperspectral remote sensing has shown an even greater potential in identifying and detecting crop diseases. Hyperspectral remote sensing refers to a special type of imaging technology that collects image data in many narrow contiguous spectral bands (<10 nm band width) throughout the visible and solar-reflected infrared portions of the spectrum (Goetz et al., 1985). Given the fact that various symptoms and the corresponding spectral responses may vary with the diseases, it is thereby necessary to conduct an independent examination on the performance of several commonly used SFs in detecting powdery mildew.

The infection caused by powdery mildew usually leads to a contiguous stretched distribution pattern in the field, which thus provides a good chance for remote sensing applications (Lorenzen and Jensen, 1989). Moreover, the most distinct symptom of powdery mildew of winter wheat is that pustules in light white (sometimes light yellow) color appear on leaves (Rémus-Borel et al., 2005). The portion of pustules on leaves will increase with the severity level, which leads to a significant spectral difference between normal leaves and infected ones, allowing the disease to be detected based on spectral signatures (Jones et al., 2010).

To date, there are a few studies addressing powdery mildew detection using spectral discrimination. The knowledge about the spectral responses to powdery mildew in winter wheat is still lacking. Lorenzen and Jensen (1989) reported the spectral characteristics of powdery mildew in barley. Rumpf et al. (2010) differentiated between diseases Cercospora leaf spot, leaf rust and powdery mildew for sugar beet at leaf level by using hyperspectral data. However, none of them has systematically explored the spectral responses that are induced by powdery mildew. In their studies, instead of using extracted SFs, the entire reflectance spectral bands were utilized with some statistical analysis methods to improve estimated accuracy, which would inevitably increase the computational load. The pivotal question, at what severity level can powdery mildew be detected, has not been answered yet. Therefore, the objectives of this study are: (1) To examine responses of a set of possible SFs to powdery mildew in winter wheat at a leaf level, and identify the most suitable SFs for disease detection; (2) to develop multivariate models in estimating the disease severity at a leaf level; and (3) to determine the severity level of powdery mildew that could be identified with an acceptable accuracy by means of a spectral discrimination analysis.

Section snippets

Study site and materials

The winter wheat (Triticum aestivum L.) plants were grown in an experimental field in Beijing Academy of Agriculture and Forestry Sciences, China, which was located at 39°56′N, 116°16′E at an altitude of 56 m (Fig. 1). Cultivar ‘Jingdong 8’ was chosen, as it was widely grown in Beijing and Hebei province and is moderately susceptible to powdery mildew. During the months of May and June, 2010, the powdery mildew (B. graminis) occurred naturally in approximately a half of the experimental field.

Spectral curves of powdery mildew

Fig. 2 illustrates curves of raw reflectances, first-derivative spectra, and reflectance ratios of slightly-damaged (3% < lesion portion < 30%) and heavily-damaged (lesion portion > 30%) leaf spectra to normal spectrum (by averaging all the measurements from normal leaves and leaves with a lesion portion <3%). From Fig. 2a and c, it is easy to see that the spectral difference between normal and slightly-damaged leaves is much smaller than that between normal and heavily-damaged leaves especially in

Discussion

The spectral response of powdery mildew was first observed by Lorenzen and Jensen (1989). In their study, they found that the disease severity of leaves become explicit with the passage of time after they were inoculated. The longer the time after the plants were inoculated, the more serious symptoms would become. However, to measure different severity degrees of powdery mildew on plant leaves from different developing stages of the disease would inevitably bring about a mixture phenomenon of

Conclusions

The remote detection of powdery mildew infection would be of value for monitoring the disease and offering a direction of fungicide spray tasks. In this study, it was found that the powdery mildew could induce a significant spectral change in both visible and NIR regions, which enables the detection of the disease by remote sensing means. PLSR and FLDA were demonstrated to be efficient in estimating and discriminating the disease severity levels using the selected spectral features. It is

Acknowledgements

This work was subsidized by the natural science foundation of Beijing city (4122032), the National Natural Science Foundation of China (41071276, 41101395), the National Basic Research Program of China (2011CB311806). The authors are grateful to Mr. Weiguo Li and Mrs. Hong Chang for help of data collection. Thanks also to Ms. Amor Elder, University of South Florida, Tampa, FL for her valuable comments on this paper.

Reference (59)

  • D. Haboudane et al.

    Hyperspectral vegetation indices and novel algorithms for predicting green LAI of crop canopies: Modeling and validation in the context of precision agriculture

    Remote Sensing Environment

    (2004)
  • N.V. Hardwick et al.

    Powdery mildew (Erysiphe graminis) on winter wheat: control with fungicides and the effects on yield

    Crop Protection

    (1994)
  • T.Z. Hu et al.

    Molecular mapping and chromosomal location of powdery mildew resistance gene in wheat cultivar tangmai

    Acta Agronomica Sinica

    (2008)
  • C.D. Jones et al.

    Diagnosis of bacterial spot of tomato using spectral signatures

    Computers and Electronics in Agriculture

    (2010)
  • L.N. Jørgensen et al.

    Fungicide treatments affect yield and moisture content of grain and straw in winter wheat

    Crop Protection

    (2002)
  • B. Li et al.

    Model selection for partial least squares regression

    Chemometrics and Intelligent Laboratory Systems

    (2002)
  • H.K. Lichtenthaler

    Chlorophylls and carotenoids: pigments of photosynthetic biomembranes

    Methods in Enzymology

    (1987)
  • Z.Y. Liu et al.

    Application of neural networks to discriminate fungal infection levels in rice panicles using hyperspectral reflectance and principal components analysis

    Computers and Electronics in Agriculture

    (2010)
  • B. Lorenzen et al.

    Changes in leaf spectral properties induced in barley by cereal powdery mildew

    Remote Sensing of Environment

    (1989)
  • E. Luedeling et al.

    Remote sensing of spider mite damage in California peach orchards

    International Journal of Applied Earth Observation and Geoinformation

    (2009)
  • D. Moshou et al.

    Automatic detection of ‘yellow rust’ in wheat using reflectance measurements and neural networks

    Computers and Electronics in Agriculture

    (2004)
  • M.A. Nofal et al.

    Integrated management of powdery mildew of mango in Egypt

    Crop Protection

    (2006)
  • M. Olsen et al.

    Effect of powdery mildew of pecan shucks on nut weight and quality and relevance to fungicide application

    Crop Protection

    (2003)
  • J. Penuelas et al.

    Reflectance indices associated with physiological changes in nitrogen- and water-limited sunflower leaves

    Remote sensing of Environment

    (1994)
  • Z.H. Qin et al.

    Detection of rice sheath blight for in-season disease management using multispectral remote sensing

    International Journal of Applied Earth Observation and Geoinformation

    (2005)
  • R. Reuveni et al.

    Foliar-fertilizer therapy—a concept in integrated pest management

    Crop Protection

    (1998)
  • W. Rémus-Borel et al.

    Silicon induces antifungal compounds in powdery mildew-infected wheat

    Physiological and Molecular Plant Pathology

    (2005)
  • J. Roujean et al.

    Estimating PAR absorbed by vegetation from bidirectional reflectance measurements

    Remote Sensing of Environment

    (1995)
  • T. Rumpf et al.

    Early detection and classification of plant diseases with Support Vector Machines based on hyperspectral reflectance

    Computers and Electronics in Agriculture

    (2010)
  • Cited by (150)

    • Remote sensing detection algorithm for apple fire blight based on UAV multispectral image

      2022, Computers and Electronics in Agriculture
      Citation Excerpt :

      The emergence of remote sensing technology has provided an important technical tool to achieve large area, rapid as well as accurate detection of crop pests and diseases. Many researchers have demonstrated the ability of remote sensing in the field of crop pests and diseases detection (Zhang et al., 2012), and Dlamini et al. (2019) described numerous remote sensing indices that are relevant for disease mapping. Zhang et al. (2014) used multi-temporal medium-resolution HJ-CCD images to monitor the occurrence and development of wheat powdery mildew on a regional scale.

    • Deep learning assisted continuous wavelet transform-based spectrogram for the detection of chlorophyll content in potato leaves

      2022, Computers and Electronics in Agriculture
      Citation Excerpt :

      Although this method can obtain accurate results, it involves disadvantages of destructive sampling and time-consuming operation (Song et al., 2021a). Visible and near-infrared (Vis-NIR) spectroscopy has proved to be a nondestructive and fast method to estimate the chlorophyll content (Li et al., 2020b; Liu et al., 2018; Zhang et al., 2012) due to the large quantities of features that spectral data contain (Liu et al., 2020a). In spectral analysis, extracting effective features from spectral data is one of the significant issues in chlorophyll content detection.

    View all citing articles on Scopus
    View full text