Integrating spectral and image data to detect Fusarium head blight of wheat

Highlights

• Our decision tree method selected optimal spectral and image features of wheat Fusarium head blight.

• Spectral and image data were integrated into fusion feature.

• A deep learning model performed best for detecting wheat Fusarium head blight.

Abstract

Fusarium head blight (FHB), caused by the fungus Gibberella zeae, infects spikelets on wheat heads and can cause significant yield and quality losses in wheat. Application of hyperspectral imaging on the detection of FHB was evaluated in the current study. Hyperspectral images were acquired from a total of 1,680 Fusarium-infected wheat head samples over a wavelength range of 400–1000 nm. The principal component analysis was used to reduce dimension of the hyperspectral image. The central wavelengths at 660, 560 and 480 nm were combined into the RGB image and then transferred to YDbDr space. The texture features of the first six principal components were extracted based on gray level co-occurrence matrix and dual-tree complex wavelet transform and the color features extracted in the color space of RGB and YDbDr. Gradient boosting decision tree and sequential backward elimination were applied to select the optimal features, and 50 spectral features and 40 image features were screened. The random forest model was built based on spectral, image, and fusion features of both spectral and image features of wheat heads to determine the optimal features dataset. Then, the deep convolutional neural network (DCNN) was established based on the optimal features dataset. This process resulted in the development of the DCNN model that predicted disease severity most accurately (R² = 0.97 and RMSE = 3.78). The DCNN model developed from this study can be used as a new tool to detect and predict the FHB disease in wheat.

Introduction

Wheat (Triticum aestivum L.) is one of the world’s leading food crops, making up a high percentage of calories consumed by the world’s population. Fusarium head blight (FHB), caused by the fungus Gibberella zeae, is a major disease in wheat worldwide. The disease can cause significant yield and quality reductions in the form of atrophy, weight reduction, and discoloration (Bauriegel et al., 2011). The fungus also produces mycotoxins including deoxynivalenol, nivalenol and zearalenones, which can adversely affect livestock and human health (Barbedo et al., 2015, Desjardin, 2006). Early and quick detection and monitoring of the development of the FHB disease are key to the control of this disease and the mycotoxins it produces in wheat.

Spectroscopy can potentially be a practical method for rapid detection of FHB in wheat and other plant diseases since it has many advantages in operation, including simple pretreatment and fast measurement. Hyperspectral imaging (HSI), an emerging spectra-based technology, is different from traditional computer vision technology. HSI combines spectroscopic and imaging techniques into one that can obtain a group of monochromatic images in hundreds of nearly continuous wavebands, in addition to providing both spectral and spatial information (Zhang et al., 2014). At present, the HSI technology has been widely used in crop growth detection and monitoring, especially in the areas of disease detection and crop damage in cereal crops (Singh et al., 2009, Zhang et al., 20122; Barbedo et al., 2017). Jiang et al. (2010) used hyperspectral images to identify healthy and yellow rust-infected wheat plants based on red edge and yellow edge positions. Alisaac et al. (2018) applied hyperspectral sensors to phenotype wheat varietal response to the infection of FHB and also investigated the correlations among disease severity, spike weight and spectral wavelengths. Bauriegel et al. (2011) used hyperspectral imaging technology to classify Fusarium infected and healthy ear tissues of wheat in the three spectral wavelength ranges with their results showing that application of spectral sub-ranges has a promising prospect. Furthermore, the data obtained from HSI also contains image data that can be used to identify wheat heads damaged by the Fusarium infection. Jin et al. (2018) applied a deep neural network (DNN) to the pixels of hyperspectral images to accurately discern the infected areas of wheat heads by FHB, illustrating that DNN is an excellent classification algorithm and that image information of hyperspectral data has the potential to classify healthy and diseased wheat ears. It can be seen that hyperspectral imaging system can not only provide spatial information but also provide spectral information of each pixel in the image, which enables hyperspectral technology to detect the intrinsic chemical and molecular information of the object and obtain the external features of the object (Wu and Sun, 2013). However, most of the previous studies are use only the spectral information without using the image information or vice versa. At present, there are some reports showing that integrating spectral and image information can have advantages over them being used alone to detect wheat FHB. Yang et al. (2015) classified corn seed varieties by using spectral and image features extracted from hyperspectral images and obtained an accuracy of over 96% based on support vector machines model, which was better in accuracy than spectral or image information used alone. Ropelewska and Zapotoczny (2018) studied the effect of the ventral and dorsal sides of wheat kernels on classifying Fusarium-infected samples; their results showed that due to both sides containing different information, there was a difference in the model performance between the ventral and dorsal sides. Little research has been conducted on the use of the front side (side A) and reverse side (side B) information to detect FHB disease in wheat. Thus, in the current study, we evaluated infected wheat heads using fusion features of spectral and image information from both the side A and side B of wheat heads to develop models that can detect wheat FHB. To investigate the optimal features dataset required to detect wheat FHB, the random forest model was built based on spectral, image and fusion features of both spectral and image features of wheat heads. A deep convolutional neural network (DCNN) model was finally constructed based on double-sided comprehensive information to develop a method that can accurately detect and predict the FHB disease in wheat.

Deep learning method is derived from the artificial neural network Multi-Layer Perceptron. DCNN has been developed rapidly and used widely in processing images, face detection, and so on (Delgado et al., 2013, Li et al., 2015). At present, DCNN has been introduced into hyperspectral image analysis with excellent performance. Hu et al. (2015) demonstrated that using DCNN to classify hyperspectral images in the spectral domain can achieve better performance than the traditional methods evaluated. Similarly, Li et al. (2017) developed a hyperspectral imaging reconstruction model based on DCNN to improve spatial features and concluded that DCNN-based framework can achieve desirable performance. Jin et al. (2018) applied a deep neural network to the pixels of hyperspectral images to classify healthy and diseased wheat heads in the field with excellent results on the detection of healthy and Fusarium-infected wheat heads.

The use of hyperspectral imaging on wheat FHB analysis depends on professional knowledge and skills. However, there is a potential to avoid such need by using machine learning or deep learning methods. Random forest, partial least squares regression, support vector regression and convolutional neural network (CNN) were often used to develop an analysis model for hyperspectral imaging (Adam et al., 2017, Tekle et al., 2015, Chu et al., 2017, Lee and Kwon, 2017). The first three models could be regarded as shallow models, which may not be as effective as CNN.

In this study, we compared those four models with the objective to develop a robust and generalized model based on hyperspectral imaging to predict the FHB disease in wheat (Fig. 1).