Hyperspectral imaging technique is known as a promising non-destructive way for detecting plants diseases and pests. In most previous studies, the utilization of the whole spectrum or a large number of bands as well as the complexity of model structure severely hampers the application of the technique in practice. If a detection system can be established with a few bands and a relatively simple logic, it would be of great significance for application. This study established a method for identifying and discriminating three commonly occurring diseases and pests of wheat, i.e., powdery mildew, yellow rust and aphid with a few specific bands. Through a comprehensive spectral analysis, only three bands at 570, 680 and 750 nm were selected. A novel vegetation index namely Ratio Triangular Vegetation Index (RTVI) was developed for detecting anomalous areas on leaves. Then, the Support Vector Machine (SVM) method was applied to construct the discrimination model based on the spectral ratio analysis. The validating results suggested that the proposed method with only three spectral bands achieved a promising accuracy with the Overall Accuracy (OA) of 83%. With three bands from the hyperspectral imaging data, the three wheat diseases and pests were successfully detected and discriminated. A stepwise strategy including background removal, damage lesions recognition and stresses discrimination was proposed. The present work can provide a basis for the design of low cost and smart instruments for disease and pest detection.
Crop diseases and pests are among the major agricultural concerns and they are characterized with multiple types, large impacts, and frequent outbreaks, which have a significant impact on the global agricultural production and food security [
To differentiate the powdery mildew and take-all diseases in wheat, Graeff et al. [
In summary, the authors suggest the development of portable instruments for identification and distinguishing of diseases and pests that is based on hyperspectral imaging data. Fewer bands should be selected when considering the cost of instrument development and the feasibility of algorithm implementation. Considering that the hyperspectral imaging technique can obtain both the detailed spectral information on the disease and pest lesion area on plant leaves and the image information of each wavelength, this data is ideal for feature selection. Therefore, instead of directly apply the expensive hyperspectral imaging technique in practice, this study takes the hyperspectral imaging data as an experimental setting for understanding spectral response, development of features and models.
This study aims to decrease the spectral confusion of different diseases and pests during their detection. Based on the hyperspectral imaging system, the methods and models are studied for detecting and discriminating three typical wheat diseases and pests, including powdery mildew (
Leaf samples used in this study were taken from the experimental field of the Beijing Academy of Agricultural and Forestry Sciences (39°56’ N, 116°16’ E). In this paper, three types of winter wheat leaf’s diseases and pests were selected: yellow rust (YR), powdery mildew (PM), and aphid (AH). Hyperspectral imaging data were collected from a hyperspectral imaging system (ImSpector V10E-QE, Spectral Imaging, Ltd., Finland) in a dark box, with the halogen lamps as a light source (
In this study, the spectral signals for model training were extracted from hyperspectral images corresponding to six infected leaves with high representativeness (two per each with PM, YR, and AH), while the spectral signals for model validation were extracted from hyperspectral images corresponding to 23 leaves (12 with PM, seven with YR, and four with AH).
In this study, typical spectra of diseases, pest and corresponding healthy samples were investigated for band optimization. For selection of ROI, considering that some natural variations of the physiological and biochemical states among leaves will result in spectral heterogeneity, selecting both normal and lesion areas within one leaf can mitigate this background difference, and facilitate the spectral comparison between the spots. The spectral curves are extracted from the ROIs based on the six modeling samples. The center pixels of stress lesions and the healthy pixels that were parallel to the selected stress lesions along the scanning direction were selected in order to avoid the difference in illumination (i.e., the radiation intensity was equal in the vertical scanning direction). Each ROI was composed of 3 × 3 pixels, while the spectral curve corresponding to the ROI was obtained after averaging all ROI pixels. The standard deviations of the spectral reflectance within the ROI (3 × 3 pixel) of each disease/pest were derived. The coefficients of variation (CV = Standard deviation/mean) among each type of samples were lower than 0.05, indicating that the spectral signals have a certain degree of variation within the ROI, but the reflectance curves were relatively stable. On this basis, 18 typical spectral curves were obtained from different leaf samples (
When crops are affected by diseases and pests, chlorophyll and cell structure will change and affect the position and area of red valley in the spectral curve of stressed crops [
The aim of this study was to propose a novel vegetation index, based on the phenomenon observed from the TVI, which will respond to different types of diseases and pests.
where, SABC refers to the area of triangle ABC and SABED is the area of trapezoidal ABED. Furthermore, the area of triangle ABC and trapezoidal ABED is calculated as follows:
Therefore, the calculation formula of RTVI is:
To evaluate the sensitivity of bands, two statistical analyses, including an independent
The majority of the existing research related to the recognition and diagnosis of plant diseases and pests by hyperspectral imaging technology uses original spectral reflectance or vegetation index as input variables for modeling and analysis. The model based on the previously described approach will ignore the subtle spectral differences among various stress types due to differences in the test sample, time, testing conditions, etc. In order to overcome the abovementioned problems, this study proposes a new method that applies the spectral ratio of stressed to healthy samples as the model input, which can strengthen the differences obtained between different sample spectrums. Instead of original spectral reflectance, the spectral ratio is able to overcome some possible spectral baseline differences among leaves. Specifically, three selected bands (570, 680, and 750 nm) were calculated according to
where, i represents band i (e.g., 570, 680, and 750 nm), D represents pixels in leaf diseased area, H represents pixels in normal leaf area, Ref(i,D) represents reflectance intensity of each pixel in leaf diseased area of band i, Ref(i,H) represents the average value of all pixels in the normal leaf area of band i. Here the diseased area and normal area refer to the identified abnormal area and normal area in the step of leaf lesion detection. The Ratio(i,D) represents the ratio of Ref(i,D) and Ref(i,H).
The specific research methods applied in this study are: (1) Mask the original hyperspectral images to remove the background. In this regard, the influence of background noise on the subsequent leaf extraction and the area extraction of disease and pest lesions is eliminated; (2) The area extraction of disease and pest lesions based on RTVI; (3) Combining the spectral ratio characteristics and SVM method to construct the disease discrimination model; and (4) Model accuracy evaluation. The main steps are presented in
In the first processing step, the plant pixels are separated from the background pixels according to the simple, intuitive, and effective threshold segmentation method. Plant clusters are selected by using a 750 nm threshold of 0.08. At the same time, 750 nm has been selected to create the RTVI index. Accordingly, the information can be reused and the cost of instrument development can be reduced. Only the pixels from plant clusters are regarded in further analysis.
Based on the masked leaf images obtained in
Based on the calculated spectral ratio image, the SVM method was applied to construct the discriminant model. This paper selects the RBF kernel function for model training [
The ratio characteristics of the three bands of the selected ROIs were considered as the model training samples, and the classification model was trained based on the SVM method. To determine the type of leaf stress, it was necessary to input the image data into the model after it was processed for leaf background separation (see
The accuracy evaluation of the discrimination model was conducted at the pixel and leaf levels. Six statistical parameters were calculated from the confusion matrix to reflect the accuracy of the discriminant model. They include the overall accuracy (OA), producers accuracy (P.’s a. (%)), user accuracy (U.’s a.(%)), kappa coefficient, commission error (%), and omission error (%).
The spectral ratio curve can reflect the change (increase or decrease) in a certain stress spectral reflection at each band relative to the normal spectrum (
The new index, RTVI, introduced the trapezoidal area (ABED in
Leaf abnormal area identification as the basis of leaf diseases and pests identification included leaf and background separation and leaf lesion area extraction.
Based on the RTVI index image, a threshold segmentation method was used to identify the lesion areas. According to a stepwise threshold optimization within a range of 0.25–0.50, the threshold of RTVI was determined as 0.4 in subsequent analysis.
The spectral ratio feature images show the spectral response of affected lesions. The direction (increase or decrease) and amplitude of reflectance changes at representative bands of vegetation can be understood as a spectral fingerprint of infection specific for different diseases and pests. In this study, extracted spectral ratio feature images (according to
Class | Reference | U.’s a. (%) | Commission error (%) | OA | Kappa | |||
---|---|---|---|---|---|---|---|---|
PM | YR | AH | Sum | |||||
PM | 913 | 1 | 0 | 914 | 99.89 | 0.11 | 0.97 | 0.95 |
YR | 30 | 893 | 42 | 965 | 92.54 | 7.46 | ||
AH | 0 | 0 | 376 | 376 | 100.00 | 0.00 | ||
Sum | 943 | 894 | 418 | 2255 | ||||
P.’s a. (%) | 96.82 | 99.89 | 89.95 | |||||
Omission error (%) | 3.18 | 0.11 | 10.05 |
Note: P.’s a. = Producer’s accuracy; U.’s a. = User’s accuracy; OA = Overall accuracy.
The accuracy of the SVM discriminant model was verified with 23 independent leaf samples. The confusion matrix of classification accuracy is presented in
Class | Reference | U.’s a. (%) | Commission error (%) | OA | Kappa | |||
---|---|---|---|---|---|---|---|---|
PM | YR | AH | Sum | |||||
PM | 10 | 2 | 0 | 12 | 83.33 | 16.67 | 0.83 | 0.72 |
YR | 1 | 5 | 0 | 6 | 83.33 | 16.67 | ||
AH | 1 | 0 | 4 | 5 | 80.00 | 20.00 | ||
Sum | 12 | 7 | 4 | 23 | ||||
P.’s a. (%) | 83.33 | 71.43 | 100.00 | |||||
Omission error (%) | 16.67 | 28.57 | 0.00 |
Note: P.’s a. = Producer’s accuracy; U.’s a. = User’s accuracy; OA = Overall accuracy.
The discriminant results of several representative samples are given in
In this study, the hyperspectral imaging technique was used as an experimental system, which may provide insights in understanding the spectral characteristic of the stresses, selection of optimal spectral bands or features, establishment of models for detection and discrimination. However, from a practical perspective, the application of the corresponding techniques in the real world requires the development of portable instruments. Instead of using the entire hyperspectral bands, instruments with some specific bands are able to achieve the trade-off between costs and functionality [
Given the complexity of the type of the abnormality under field conditions, some abiotic stresses (e.g., nutrient stress, water stress) may co-occur with the biotic stresses, which thereby also need to be taken into account. It is encouraging that the identified three bands were also included in some relevant vegetation indices. For example, Photochemical/Physiological Reflectance Index (PRI; including 570 nm bands) and MERIS Terrestrial Chlorophyll Index (MTCI; including 680 and 750 nm bands) could be potentially used in N-stress estimation [
Identification and differentiation of different plant diseases and pests is a practically important task in plant monitoring via remote sensing technology. In this study, based on hyperspectral imaging technique, a detection and discrimination method was proposed and successfully applied in differentiating three typical diseases and pests of winter wheat (powdery mildew, yellow rust, and aphid) in Northern China. The main conclusions are as follows: The wheat diseases and pests can be effectively detected and discriminated through a proposed stepwise procedure that includes background elimination, extraction of spectral features, identification of damaged areas, and discrimination of stresses. A relatively satisfactory accuracy is achieved of the discriminant model, with OA of 83% and kappa coefficient of 0.72. It is encouraging that the detecting procedure was successfully established on only three specific bands at 570, 680, and 750 nm, which significantly reduced the computational load. Based on these bands, an improved index (RTVI) was proposed for distinguishing damaged areas on leaves, that was able to facilitate further discrimination of diseases and pests. Based on the constructed spectral ratio fingerprint and the SVM algorithm, the wheat diseases and pests can be effectively differentiated on a leaf level. While the feasibility of the discrimination model on canopy level is still unknown, it needs to be tested in future research. Its adaptability to some more complex scenarios is also worth studied.
The proposed method for detecting and differentiating wheat diseases and pests is established on only a few spectral bands and some simple algorithms. It has great potential to serve as a core method in designing a customized optical sensor for detecting plants’ stresses. Such non-contact and low-cost techniques may bring new insights in crop protection and phenotypic studies of crop resistance to diseases and pests.