The rapid, stable, and undamaged picking of small-sized spherical fruits are one of the key technologies to improve the level of intelligent picking robots and reduce grading operations. Cherry tomatoes were selected as the research object in this work. Picking strategies of two-stage “Holding-Rotating” and finger-end grasping were determined. The end-effector was designed to separate the fruit from the stalk based on the linear motion of the constraint part and the rotating gripper. This work first studied the human hand-grasping of cherry tomatoes and designed the fingers with sinusoidal characteristics. The mathematical model of a single finger of the gripper was established. The structural parameters of the gripper were determined to meet the requirements of the grabbing range from 0 to 61.6 mm. Based on the simulation model, the constraint part was set to 6 speeds, and the fruit sizes were set to 20 mm, 30 mm, and 40 mm, respectively. When the speed was 0.08m/s, the results showed that the grabbing time was 0.5381 s, 0.387 s, and 0.2761 s, respectively, and the maximum grabbing force was 0.9717 N, 3.5077 N, and 4.0003 N now of clamping, respectively. It met the picking requirements of high speed and low loss. The criterions of two-index stability and undamaged were proposed, including the grasping index of the fixed value and the slip detection of variance to mean ratio. Therefore, the control strategy and algorithm based on two-stage and two-index for rapid, stable, and non-destructive harvesting of small fruit were proposed. The results of the picking experiment for seventy-two cherry tomatoes showed that the picking success rate was 95.82%, the average picking time was 4.86 s, the picking damage rate was 2.90%, the browning rate was 2.90% in 72 h, and the wrinkling rate was 1.49% in 72 h, which can meet the actual small spherical fruit picking requirements. The research will provide an idea for the flexible end-effectors with humanoid grasp function and provides a theoretical reference for small spherical fruit picking.

Multi-class detection of cherry tomatoes using improved Yolov4-tiny model

The detection and control of fungal spores in greenhouse crops are important for stabilizing and increasing crop yield. At present, the detection of fungal spores mainly adopts the method of combining portable volumetric spore traps and microscope image processing. This method is problematic as it is limited by the small field of view of the microscope and has low efficiency. This study proposes a rapid detection method for fungal spores from greenhouse crops based on CMOS image sensors and diffraction fingerprint feature processing. We built a diffraction fingerprint image acquisition system for fungal spores of greenhouse crops and collected diffraction fingerprint images of three kinds of fungal spores. A total of 13 diffraction fingerprint features were selected for the classification of fungal spores. These 13 characteristic values were divided into 3 categories, main bright fringe, main dark fringe, and center fringe. Then, these three features were calculated to obtain the Peak to Center ratio (PCR), Valley to Center ratio, and Peak to Valley ratio (PVR). Based on these features, logistics regression (LR), K nearest neighbor (KNN), random forest (RF), and support vector machine (SVM) classification models were built. The test results show that the SVM model has a better overall classification performance than the LR, KNN, and RF models. The average accuracy rate of the recognition of three kinds of fungal spores from greenhouse crops under the SVM model was 92.72%, while the accuracy rates of the LR, KNN, and RF models were 84.97%, 87.44%, and 88.72%, respectively. The F1-Score value of the SVM model was higher, and the overall average value reached 89.41%, which was 11.12%, 7.18%, and 5.57% higher than the LR, KNN, and RF models, respectively. Therefore, the method proposed in this study can be used for the remote identification of three fungal spores which can provide a reference for the identification of fungal spores in greenhouse crops and has the advantages of low cost and portability.

Aboveground biomass (AGB) is an important indicator used to predict crop yield. Traditional spectral features or image textures have been proposed to estimate the AGB of crops, but they perform poorly at high biomass levels. This study thus evaluated the ability of spectral features, image textures, and their combinations to estimate winter wheat AGB. Spectral features were obtained from the wheat canopy reflectance spectra at 400–1000 nm, including original wavelengths and seven vegetation indices. Effective wavelengths (EWs) were screened through use of the successive projection algorithm, and the optimal vegetation index was selected by correlation analysis. Image texture features, including texture features and the normalized difference texture index, were extracted using gray level co-occurrence matrices. Effective variables, including the optimal texture subset (OTEXS) and optimal normalized difference texture index subset (ONDTIS), were selected by the ranking of feature importance using the random forest (RF) algorithm. Linear regression (LR), partial least squares regression (PLS), and RF were established to evaluate the relationship between each calculated feature and AGB. Results demonstrate that the ONDTIS with PLS based on the validation datasets exhibited better performance in estimating AGB for the post-seedling stage (R² = 0.75, RMSE = 0.04). Moreover, the combinations of the OTEXS and EWs exhibited the highest prediction accuracy for the seeding stage when based on the PLS model (R² = 0.94, RMSE = 0.01), the post-seedling stage when based on the LR model (R² = 0.78, RMSE = 0.05), and for all stages when based on the RF model (R² = 0.87, RMSE = 0.05). Hence, the combined use of spectral and image textures can effectively improve the accuracy of AGB estimation, especially at the post-seedling stage.

In this paper, hyperspectral imaging technology, combined with chemometrics methods, was used to detect the nitrogen content of soybean leaves, and to achieve the rapid, non-destructive and in situ detection of the nitrogen content in soybean leaves. Soybean leaves under different fertilization treatments were used as the research object, and the hyperspectral imaging data and the corresponding nitrogen content data of soybean leaves at different growth stages were obtained. Seven spectral preprocessing methods, such as Savitzky–Golay smoothing (SG), first derivative (1-Der), and direct orthogonal signal correction (DOSC), were used to establish the quantitative prediction models for soybean leaf nitrogen content, and the quantitative prediction models of different spectral preprocessing methods for soybean leaf nitrogen content were analyzed and compared. On this basis, successive projections algorithm (SPA), genetic algorithm (GA) and random frog (RF) were employed to select the characteristic wavelengths and compress the spectral data. The results showed the following: (1) The full-spectrum prediction model of soybean leaf nitrogen content based on DOSC pretreatment was the best. (2) The PLS model of soybean leaf nitrogen content based on the five characteristic wavelengths had the best prediction performance. (3) The spatial distribution map of soybean leaf nitrogen content was generated in a pixel manner using the extracted five characteristic wavelengths and the DOSC-RF-PLS model. The nitrogen content level of soybean leaves can be quantified in a simple way; this provides a foundation for rapid in situ non-destructive detection and the spatial distribution difference detection of soybean leaf nitrogen. (4) The overall results illustrated that hyperspectral imaging technology was a powerful tool for the spatial prediction of the nitrogen content in soybean leaves, which provided a new method for the spatial distribution of the soybean nutrient status and the dynamic monitoring of the growth status.

Soybean is one of the world’s major oil-bearing crops and occupies an important role in the daily diet of human beings. However, the frequent occurrence of soybean leaf diseases caused serious threats to its yield and quality during soybean cultivation. Rapid identification of soybean leaf diseases could provide a better solution for efficient control and subsequent precision application. In this study, a lightweight deep convolutional neural network (CNN) based on multiscale feature extraction fusion (MFEF) and combined with a dense connectivity (DC) network (MFEF-DCNet) was proposed for soybean leaf disease identification. In MFEF-DCNet, a multiscale feature extraction fusion (MFEF) module for soybean leaves was constructed by utilizing a convolutional attention module and depth-separable convolution to improve the model feature extraction capability. Multiscale features are fused by using dense connections (DC) in the backbone network to improve the model generalization capability. Experiments were implemented on eight distinct disease and deficiency classes of soybean images (including bacterial blight, cercospora leaf blight, downy mildew, frogeye leaf spot, healthy, potassium deficiency, soybean rust, and target spot) using the proposed network. The results showed that the MFEF-DCNet had an accuracy of 0.9470, an average precision of 0.9510, an average recall of 0.9480, and an F1-score of 0.9490 for soybean leaf disease identification. And MFEF-DCNet had certain performance advantages in terms of classification accuracy, convergence speed and other effects compared with VGG16, ResNet50, DenseNet201, EfficientNetB0, Xception and MobileNetV3_small models. In addition, the accuracy of the MFEF-DCNet model in recognizing soybean diseases in local data was 0.9024, which indicated that the MFEF-DCNet model had favorable application in practical applications. The proposed model and experience in this study could provide useful inspiration for automated disease identification in soybean and other crops.