Visual inspection of rice leaf blast resistance is time-consuming and labor-intensive with low accuracy. Therefore, this study aims to identify and detect rice leaf blast spots based on RGB imaging of rice canopy combined with mask regions with convolutional neural network (Mask-RCNN), and develop multiple classification models to quantify the number of disease spots and evaluate the association between the number of disease spots and the resistance level by analyzing the quantitative information of different categories of disease spots in RGB images of rice. First, we collected RGB images from different rice breeding lines at the seedling stage, including japonica lines, early indica lines and indica recovery lines. Preprocessing and labeling of the input images were then performed. A Mask-RCNN model for the recognition of rice leaf blast spots was developed to perform the rectangular frame detection, mask segmentation and classification. The classification result of rice leaf blast spots with the mean intersection over union (mIoU) of 0.603 was achieved. The mean average precision (mAP) of the test dataset was 0.716, when the intersection over union (IoU) threshold of 0.5 was used. Among all the classification models, Gaussian process support vector machine obtained the highest prediction accuracy of 94.30% (proportion of disease spots in each category corresponding to different resistances) on the test dataset. The above results demonstrate that RGB images of rice canopy combined with Mask-RCNN have the great potential for the accurate identification of rice leaf blast spots, and the number of detected disease spots is highly correlated with the rice leaf blast resistance level. The proposed method is promising for efficient selection of disease-resistant rice varieties in breeding.

In order to recognize citrus leaf canker disease accurately and quickly, a diagnosis method of citrus leaf canker disease based on naive Bayesian classification was proposed. The digital images of leaves with different severities of citrus leaf canker disease were used as the data source. According to the characteristics of color space, a disease spot recognition model based on naive Bayesian classification was established for rapid diagnosis of citrus leaf canker disease, and the diagnostic abilities of naive Bayesian classification, fixed threshold, adaptive threshold and support vector machine for citrus leaf canker disease were compared. The results showed that the method based on naive Bayesian classification was effective in the segmentation of citrus leaf canker disease, and the incorrect segmentation rate was only 3.58%, which was far better than the threshold methods and support vector machine. In terms of performance efficiency, the time order of the four algorithms was fixed threshold method＜adaptive threshold＜naive Bayesian＜support vector machine, all of which were within a reasonable range. Combined with the preparation time, naive Bayesian method had the best performance efficiency. Therefore, the naive Bayesian classification algorithm has a rapid and accurate application ability in the diagnosis of citrus leaf canker disease, and can provide a new way for the accurate diagnosis of fruit tree disease severities.

In order to more accurately monitor the vertical distribution of the leaf area index (LAI) of maize, we propose a conditionally constrained LAI vertical distribution inversion method based on the simulation dataset constructed by the discrete anisotropic radiative transfer (DART) model. First, the simulation effects of DART model on canopy reflectance and photosynthetically active radiation (PAR) were evaluated based on the three-layer vertical distribution scenario, and constructed the corresponding simulation dataset with PAR. Second, a single parameter inversion model for LAI and PAR was built based on the simulated dataset. Finally, using the single parameter inversion model as a priori knowledge, the inversion of the vertical distribution of maize canopy LAI based on the hyperspectral vegetation index was realized by solving the constraint problem. The results showed that the accuracy of the constraint optimization inversion model was higher than that of the single parameter inversion model. The coefficient of determination (R²) of LAI inversion results for top layer of maize increased by 0.022, root-mean-square error (RMSE) decreased by 0.016 m²/m², and normalized root-mean-square error (NRMSE) decreased by 1.3%. The R² of LAI inversion results for middle layer of maize increased by 0.08, RMSE decreased by 0.219 m²/m², and NRMSE decreased by 10.1%. The R² of LAI inversion results for bottom layer of maize increased by 0.069, RMSE decreased by 0.041 m²/m², and NRMSE decreased by 4.6%. Therefore, it can be concluded that the inversion of vertical distribution of LAI in maize canopy using the conditional constraint optimization method can effectively improve the inversion accuracy.

This study focuses on the accuracy and stability of deep learning method for maize tassel segmentation at different tasseling stages and varieties. The RGB images were collected in the experimental base of Chinese Academy of Agricultural Sciences in Xinxiang City of Henan Province from July to September in 2019, and four models, PspNet, DeepLab V3＋, SegNet and U-Net, based on the lightweight network as the feature extraction layer, were applied to compare the accuracy of different models for maize tassel segmentation. Then, the U-Net model with the best segmentation accuracy (mIoU=0.780) was selected to segment the maize tassel in different varieties at different tasseling stages. The results showed that the accuracy of U-Net model at different tasseling stages was generally better (mIoU=0.703 to 0.798), and the segmentation accuracy of maize tassel in fully emerged tassel stage was the highest (mIoU=0.798); the segmentation accuracy of maize tassel in different varieties was significantly different, but the average segmentation accuracy of maize tassel for all varieties was higher (mIoU=0.749), and the segmentation accuracy of Zhengdan 958 (ZD958) was the highest (mIoU=0.814). In summary, the U-Net model has good universality and robustness for maize tassel segmentation, which provides an effective method for tassel monitoring in maize phenotypic test in the future.

This research was based on the combined planting model of corn-soybean strip intercropping and the corns under different nitrogen levels were used as the test materials. The reflectance spectrum and chlorophyll content of leaves and canopies of corns were measured at the jointing stage, tasseling stage and filling stage. Red edge position (REP) was extracted by continuous wavelet transform (CWT) and other algorithms [maximum first derivative method (FD), four-point interpolation method (FPI) and linear extrapolation method (LEM)]. The quantitative relationships between REP and chlorophyll contents were systematically analyzed to compare the accuracy and stability of the REP extracted by each red edge algorithm on the two scales of leaf and canopy. The results showed that, based on the REP-CWT, the estimation accuracy of chlorophyll content was higher on leaf and canopy scales, and the stability was the strongest, which indicated that REP-CWT was feasible in extracting the REP of corn reflectance spectrum. The quantitative estimation models of corn leaf chlorophyll content and canopy chlorophyll content base on REP-LEM and REP-FPI, respectively, were the best. This study provides a new method for extracting the REP of corn reflectance spectrum, and then constructs the best quantitative estimation model of corn chlorophyll content on different observation scales (leaf and canopy), and offers an effective way to monitor the nitrogen nutrition status of corn.