Most farmers who cultivate green soybeans use manual sorting, with work efficiency of 12 kg/h. To improve this low work efficiency, a sorting machine must be developed for rapid accurate detection and classification of their appearance quality. An object detection artificial intelligence (AI) was developed for sorting green soybeans. Methods to achieve high performance were discussed. An AI developed with a dataset including one variety had higher precision, recall, and a higher F-value than an AI created with a dataset including data from multiple varieties. The highest Newton efficiency was 0.57, equivalent to manual sorting, was obtained by inclusion of good and defective product images in the dataset. Results confirmed the effectiveness of AI-based object detection for sorting green soybeans.

For this study, after developing an object detection AI for appearance quality sorting of green soybeans, we investigated the effects on AI accuracy of different green soybean varieties in datasets and different object detection algorithms. After setting seven datasets by combining three varieties, we used these datasets to develop AI for YOLOv3 and Faster R-CNN. Results show that the Precision, Recall, and F-value of the AI, including images of the green soybean varieties to be sorted, were significantly higher. The Newton efficiency (η) of the AI with highest accuracy was 0.79. The contents and percentages of the green soybeans which were misclassified and undetected for appearance quality by AI were different for each object detection algorithm.

Gas chromatography-mass spectrometry-based metabolomics approach was applied to reveal volatile compounds that reflect the freshness of whole and fresh-cut cabbage. Forty and thirty volatiles were annotated in whole and fresh-cut cabbage, respectively, and twenty of which were in common. A developed partial least squares regression (PLSR) model using the detected volatiles allowed estimating the cumulative storage temperature with sufficient accuracy. A hierarchical cluster analysis with selected volatiles based on variable importance in projection of each PLSR model showed that the profiles of the volatile compounds differed with the cumulative storage temperature. These results suggest the feasibility of the freshness assessment for whole and fresh-cut cabbage by profiling the selected volatile compounds as candidates for freshness markers.

In this study, a high-speed ridge forming seeder with a disc-type ridging and a double plate-type seeding mechanism was developed to reduce moisture damage during the early growth of soybean. Comparison tests were conducted in Miyagi, Niigata, Toyama, and Saitama prefectures to compare the prototype with an existing rotary tilling and ridge-making seeder. The prototype was capable of working at 1.5 m/s, which was more than twice as fast as the existing seeder, and the seeding accuracy was comparable or better. The germination rates of both seeders were more than 95 %. There were no significant differences in yield. In areas where pre-tillage is possible, it is possible to secure yields by seeding at the right time and to expand the seeded area by increasing the speed.

Green soybean, farmers have introduced color sorting machines to improve green soybean sorting work. This investigation assessed effects of cleaning in a sorting machine based on analysis of exhaust solenoid valve. Results show the V value, indicating smoke concentration, as 20.0-27.3 at 40-80 s. That value diminished with the passage of time, but showed little difference at each measuring point. Exhaust solenoid valves were cleaned inside of the sorting machine during 10 s at 0.6 MPa (2-6 times/s) and during 20 s at 0.4 MPa (5 times/s). The smoke concentration in exhaust decreased gradually for 120 s at 0.4 MPa (2 times/s). Results show that exhaust solenoid valves inside of sorting machines have sufficient performance.

To automate the grading of external defects of potato, two classification models (a convolutional neural network (CNN) and a support vector machine (SVM) model) were built based on either color or short-wave infrared (SWIR) images captured by different optical systems positioned along a grading line. The images were manually labeled into six external defect categories. Consequently, the highest classification accuracy (96.8 %) was obtained by the CNN model based on SWIR images. Furthermore, by visualizing influential regions for validation of the classification models, surface color and edge information features were apparent in the color images, while white highlighted areas with high reflectance were observed in the SWIR images.