Original papers
Wheat leaf rust detection at canopy scale under different LAI levels using machine learning techniques

https://doi.org/10.1016/j.compag.2018.11.016Get rights and content

Highlights

  • Addressing the effects of different LAI values on the accuracy of the selected machine learning methods in retrieval of wheat leaf rust severity level.

  • Provide separate wavelength ranges with the highest contributions in the wheat leaf rust inversion model at the three LAI levels.

Abstract

Accurate diagnosis of wheat leaf rust is of high interest for precision farming. Spectral data have been increasingly employed to detect this disease at leaf or canopy scales; however, less attention has been paid to the variations of leaf area index (LAI). Therefore, in this study, identification of wheat leaf rust was investigated at canopy scale and under different LAI levels, namely high, medium and low. Four machine learning (ML) methods including ν-support vector regression (ν-SVR), boosted regression trees (BRT), random forests regression (RFR) and Gaussian process regression (GPR) were built to estimate disease severity (DS) levels at canopy scale, where the reflectance data were measured in field situ by using a spectroradiometer, in which records spectra from 350 to 2500 nm. Results showed that ν-SVR outperformed the other ML methods at all three LAI levels with the R2 measures all being around 0.99. The results, particularly, showed that the performances of the ML methods were improved with increasing LAI value, where RFR reported the worst R2 value of 0.79 (RMSE = 8.5%) at low LAI level. The variable importance obtained using BRT showed three distinct regions of wavelengths that were appropriate across different LAI levels. The results of this research confirmed that hyperspectral signature can be reliably considered to identify wheat leaf rust disease at different LAI levels. Moreover, performances of several spectral vegetation indices (SVIs) were compared with those of the ML techniques. The results showed that the SVIs were consistently outperformed by the ML methods, particularly at low LAI level in which the SVIs were adversely affected. Nevertheless, all the SVIs, except for the RVSI, performed moderately well at high and medium LAI levels.

Introduction

Wheat leaf rust caused by P. recondita f. sp. tritici (Prt) is one of the most common and destructive fungal diseases of wheat which threats global food security. Plants infected by this disease exhibit various symptoms at different stages of development which can be simultaneously observed in various parts of the infected leaves and leaves can be in a variety of colors such as yellow, orange, gray or brown (Bolton et al., 2008).

Wheat leaf rust has been reported frequently in many wheat growing areas worldwide and it reduces both wheat yield quantity and quality. Hence, detection and monitoring of this disease is of paramount importance for precision farming (PF). Traditional disease detection methods often observe crops in situ which is costly and time consuming in large-scale studies. Therefore, an efficient method for rapid identification of plant diseases in large-scale wheat growing regions is of utmost necessity. Spectral data acquired from in situ, airborne and space borne sensors provide reliable information for early plant disease detection, being utilized in a plethora of studies (West et al., 2003, Mewes et al., 2011) at different scales including leaf-level (Yuan et al., 2013, Zhang et al., 2012), canopy level (Wang et al., 2015) and aerial/space remote sensing (Franke and Menz, 2007). Leaf-scale studies only consider optical properties of an infected leaf at different disease severity levels (Devadas et al., 2009). Disease detection at canopy scale, however, reflects not only the leaf properties but also the background soil, leaf area index and plant geometry (Franke and Menz, 2007). Finally, atmospheric effects on path radiance are investigated in aerial and space remote sensing (Kaufman and Tanre, 1992). In plant disease detection studies, canopy scale studies are the basis for airborne and space borne disease detection methods (Zhang et al., 2012).

Hyperspectral signature are exceptionally prevalent in plant disease detection applications because they provide valuable information about disease development at early stages. Recorded reflectance data at the visible and near infrared (VNIR) wavelengths have shown high potential for plant disease detection compared to the short wave infrared (SWIR) and thermal infrared (TIR) wavelengths. This is because most diseases influence biochemical material content and structure of the leaves, and thus infected areas are more sensitive to the VNIR wavelengths. Hence, narrow spectral bands in the VNIR regions have been investigated in many studies to detect plant diseases such as wheat leaf rust (Ashourloo et al., 2014, Mahlein et al., 2013).

Leaf rust, yellow rust and stem rust are the three types of wheat rust diseases. Among them, leaf rust has the highest frequency of occurrence in the world, making it an important disease for future studies. However, it has been studied in few studies (Franke and Menz, 2007, Devadas et al., 2009, Ashourloo et al., 2014, Pryzant et al., 2017, Wang et al., 2015). A variety of techniques including classification (Franke and Menz, 2007), spectral vegetation indices (SVIs) (Thenkabail et al., 2000), spectral disease indices (SDIs) (Mahlein et al., 2013) and machine learning (ML) methods (Ashourloo et al., 2016) have been successfully adopted for plant disease detection. SVIs and SDIs are popular due to their simplicity and interpretability in disease detection. Most studies have utilized prevalent SVIs for disease detection, and few have developed SDIs for identification of a specific disease. However, SVIs and SDIs usually use few spectral bands in disease detection and not the full spectrum. On the other hand, ML methods have shown promising results due to their capabilities in handling complex associations between dependent and independent variables (Friedman and Meulman, 2003, Leathwick et al., 2006, Elith et al., 2008). Moreover, ML methods are known to handle data with high dimensionality. Plant disease detection is usually conducted in narrow regions of the spectrum. These methods are appearing superior to vegetation indices and they have shown satisfactory outcomes in detecting plant diseases (Ashourloo et al., 2014). In a study, conducted by Ashourloo et al. (2016), three ML techniques were investigated for wheat leaf rust detection at canopy level. Among the ML methods, GPR was superior, where all of the ML methods showed higher accuracy than the SVIs. In addition, various symptoms of the disease did not affect the ML methods as much as they did the SVIs (Ashourloo et al., 2016).

Spectral reflectance of vegetation is affected by several factors, including the background soil, leaf area index (LAI), leaf angle, leaf thickness, and chlorophyll content (Verhoef, 1984). Of these parameters, LAI plays an important role in affecting vegetation spectral signature (Clevers and Verhoef, 1993) and it changes across different wheat cultivars and under rainfed or irrigated wheat farms. So far, most of the disease detection studies have been conducted in field situ or in the lab without specifying variations of LAI. Therefore, it is necessary to evaluate the potential of different ML methods in disease detection at various levels of LAI. To the best of the authors’ knowledge, identification of leaf rust disease using hyperspectral remote sensing data at canopy level has not been investigated under different LAI levels. Therefore, this paper is aimed at determination of leaf rust disease severity levels at canopy scale and under three different LAI levels including low, medium and high levels through adoption of boosted regression trees (BRT), GPR, SVR and random forest regression (RFR) as state-of-the-art ML techniques which have successfully been applied in several applications (Aghighi et al., 2018, Azadbakht et al., 2018, Mutanga et al., 2012, Elith et al., 2008).

The remainder of this manuscript is organized as follows. In Section 2, experimental setup, the ML methods and the selected SVIs are presented. Section 3 provides the results of our experiments in detail and under different conditions, and the concluding remarks are given in Section 4.

Section snippets

Experimental setup

Field experiments were conducted at a site located in the Moghan fertilized plain, in the North West of Iran, in which approximately 7000 hectares of this plain is usually cultivated by wheat annually. This area is expanded from latitudes 39.465° N to 39.615° N and longitudes 47.548° E to 48.009° E.

To examine the potential of the ML methods in disease detection, hyperspectral reflectance data of healthy and diseased leaves were collected at canopy scale under natural environmental conditions.

Performance evaluation of the ML methods

In order to reduce the bias of random splitting, the ML methods were implemented 10 times on three categories of observations, representing different LAI levels. Boxplots in Fig. 3 represent the calculated R2 values of the actual disease severity levels against those predicted from each model under 10 runs, indicating how well each model approximates the real disease levels. As seen, larger R2 values are reported for all the regression models at high LAI level, followed by marginally deficient

Conclusions

In this study, wheat leaf rust disease inversion models were built using ML methods and SVIs at canopy scale and under three distinct LAI levels. Hyperspectral signature data collected in field situ by using a spectroradiometer were served as inputs of the models, where all the ML inversion models performed acceptably at high and medium LAI levels. Of the implemented methods, however, ν-SVR was less affected by variations of LAI, followed by GPR. More research is required to determine the

References (56)

  • H. Aghighi et al.

    Machine learning regression techniques for the silage maize yield prediction using time-series images of landsat 8 oli

    IEEE J. Sel. Top. Appl. Earth Observ. Rem. Sens.

    (2018)
  • H. Al-Hiary et al.

    Fast and accurate detection and classification of plant diseases

    Mach. Learn.

    (2011)
  • D. Ashourloo et al.

    An investigation into machine learning regression techniques for the leaf rust disease detection using hyperspectral measurement

    IEEE J. Sel. Top. Appl. Earth Observ. Rem. Sens.

    (2016)
  • D. Ashourloo et al.

    Developing two spectral disease indices for detection of wheat leaf rust (pucciniatriticina)

    Rem. Sens.

    (2014)
  • M. Berenson et al.

    Intermediate Statistical Methods and Applications: A Computer Package Approach

    (1983)
  • G. Biau et al.

    A random forest guided tour

    Test

    (2016)
  • M.D. Bolton et al.

    Wheat leaf rust caused by puccinia triticina

    Mol. Plant Pathol.

    (2008)
  • A. Bowley

    Elements of Statistics. Studies in Economics and Political Science: Monographs

    (1917)
  • L. Breiman

    Random forests

    Mach. Learn.

    (2001)
  • L. Breiman et al.

    Classification and Regression Trees. The Wadsworth and Brooks-Cole Statistics-Probability Series

    (1984)
  • Y. Bu et al.

    Stellar atmospheric parameter estimation using gaussian process regression

    Monthly Not. Roy. Astron. Soc.

    (2014)
  • T. Chai et al.

    Root mean square error (rmse) or mean absolute error (mae)?

    Geosci. Model Dev. Discuss.

    (2014)
  • C. Chang et al.

    Training nu-support vector regression: theory and algorithms

    Neural Comput.

    (2002)
  • V. Cherkassky et al.

    Learning from Data: Concepts, Theory, and Methods

    (2007)
  • J. Clevers et al.

    Lai estimation by means of the wdvi: a sensitivity analysis with a combined prospect-sail model

    Rem. Sens. Rev.

    (1993)
  • R. Devadas et al.

    Evaluating ten spectral vegetation indices for identifying rust infection in individual wheat leaves

    Precis. Agric.

    (2009)
  • J. Devore et al.

    Modern Mathematical Statistics with Applications. Springer Texts in Statistics

    (2011)
  • J. Elith et al.

    A working guide to boosted regression trees

    J. Anim. Ecol.

    (2008)
  • Cited by (67)

    View all citing articles on Scopus
    View full text