Fusarium head blight (FHB) is a devastating disease of wheat worldwide. In addition to reducing the yield of the crop, the causal pathogens also produce mycotoxins that can contaminate the grain. The development of resistant wheat varieties is one of the best ways to reduce the impact of FHB. To develop such varieties, breeders must expose germplasm lines to the pathogen in the field and assess the disease reaction. Phenotyping breeding materials for resistance to FHB is time-consuming, labor-intensive, and expensive when using conventional protocols. To develop a reliable and cost-effective high throughput phenotyping system for assessing FHB in the field, we focused on developing a method for processing color images of wheat spikes to accurately detect diseased areas using deep learning and image processing techniques. Color images of wheat spikes at the milk stage were collected in a shadow condition and processed to construct datasets, which were used to retrain a deep convolutional neural network model using transfer learning. Testing results showed that the model detected spikes very accurately in the images since the coefficient of determination for the number of spikes tallied by manual count and the model was 0.80. The model was assessed, and the mean average precision for the testing dataset was 0.9201. On the basis of the results for spike detection, a new color feature was applied to obtain the gray image of each spike and a modified region-growing algorithm was implemented to segment and detect the diseased areas of each spike. Results showed that the region growing algorithm performed better than the K-means and Otsu’s method in segmenting diseased areas. We demonstrated that deep learning techniques enable accurate detection of FHB in wheat based on color image analysis, and the proposed method can effectively detect spikes and diseased areas, which improves the efficiency of the FHB assessment in the field.

Hyperspectral imaging is becoming an increasingly popular tool for high-throughput plant phenotyping, because it provides remarkable insights about the health status of plants. Feature selection is a key component in a hyperspectral image analysis, largely because a significant portion of spectral features are redundant and/or irrelevant, depending on the desired application. This paper presents an ensemble feature selection method to identify the most informative spectral features for practical applications in plant phenotyping. The hyperspectral data set contained the images of four wheat lines, each with a control and a salt (NaCl) treatment. To rank spectral features, six feature selection methods were used as the base for the ensemble: correlation-based feature selection, ReliefF, sequential feature selection, support vector machine-recursive feature elimination (SVM-RFE), LASSO logistic regression, and random forest. The best results were achieved by the ensemble of ReliefF, SVM-RFE, and random forest, which drastically reduced the dimension of the hyperspectral data set from 215 to 15 features, while improving the accuracy in classifying the salt-treated vegetation pixels from the control pixels by 8.5%. To transform the hyperspectral data set into a multispectral data set, six wavelengths as the center of broad multispectral bands around the most prominent features were determined by a clustering algorithm. The result of salt tolerance assessment of the four wheat lines using the derived multispectral data set was similar to that of the hyperspectral data set. This demonstrates that the proposed feature selection pipeline can be utilized for determining the most informative features and can be a valuable tool in the development of tailored multispectral cameras.

Salinity stress has significant adverse effects on crop productivity and yield. The primary goal of this study was to quantitatively rank salt tolerance in wheat using hyperspectral imaging. Four wheat lines were assayed in a hydroponic system with control and salt treatments (0 and 200 mM NaCl). Hyperspectral images were captured one day after salt application when there were no visual symptoms. Subsequent to necessary preprocessing tasks, two endmembers, each representing one of the treatment, were identified in each image using successive volume maximization. To simplify image analysis and interpretation, similarity of all pixels to the salt endmember was calculated by a technique proposed in this study, referred to as vector-wise similarity measurement. Using this approach allowed high-dimensional hyperspectral images to be reduced to one-dimensional gray-scale images while retaining all relevant information. Two methods were then utilized to analyze the gray-scale images: minimum difference of pair assignments and Bayesian method. The rankings of both methods were similar and consistent with the expected ranking obtained by conventional phenotyping experiments and historical evidence of salt tolerance. This research highlights the application of machine learning in hyperspectral image analysis for phenotyping of plants in a quantitative, interpretable, and non-invasive manner.

Assessment of the nitrogen status of grapevines with high spatial, temporal resolution offers benefits in fertilizer use efficiency, crop yield and quality, and vineyard uniformity. The primary objective of this study was to develop a robust predictive model for grapevine nitrogen estimation at bloom stage using high-resolution multispectral images captured by an unmanned aerial vehicle (UAV). Aerial imagery and leaf tissue sampling were conducted from 150 grapevines subjected to five rates of nitrogen applications. Subsequent to appropriate pre-processing steps, pixels representing the canopy were segmented from the background per each vine. First, we defined a binary classification problem using pixels of three vines with the minimum (low-N class) and two vines with the maximum (high-N class) nitrogen concentration. Following optimized hyperparameters configuration, we trained five machine learning classifiers, including support vector machine (SVM), random forest, XGBoost, quadratic discriminant analysis (QDA), and deep neural network (DNN) with fully-connected layers. Among the classifiers, SVM offered the highest F1-score (82.24%) on the test dataset at the cost of a very long training time compared to the other classifiers. Alternatively, QDA and XGBoost required the minimum training time with promising F1-score of 80.85% and 80.27%, respectively. Second, we transformed the classification into a regression problem by averaging the posterior probability of high-N class for all pixels within each of 150 vines. XGBoost exhibited a slightly larger coefficient of determination (R² = 0.56) and lower root mean square error (RMSE) (0.23%) compared to other learning methods in the prediction of nitrogen concentration of all vines. The proposed approach provides values in (i) leveraging high-resolution imagery, (ii) investigating spatial distribution of nitrogen across a vine’s canopy, and (iii) defining spatial zones for nitrogen application and smart sampling.